原始内容
The secure AI agent runtime for builders who can't afford a security incident.
Run autonomous, self-hosted AI agent fleets that are isolated, auditable, and production-ready. Every agent runs in its own Docker container, API keys never leave the credential vault, and per-agent budgets cap spend. A source-available, security-first OpenClaw alternative for teams. Chat via Telegram, Discord, Slack, or WhatsApp. 100+ LLM providers via LiteLLM.
What is OpenLegion? · Quick Start · OpenLegion vs OpenClaw · Security Model · Docs
Demo
https://github.com/user-attachments/assets/8bd3fe95-5734-474d-92f0-40616daf91ad
openlegion start→ inline setup → multiple agents running. Live cost tracking. No configuration files edited by hand. Connect Telegram, WhatsApp, Slack, and Discord.
Table of Contents
- What is OpenLegion?
- Who is OpenLegion for?
- Quick Start
- OpenLegion vs OpenClaw
- What It Does
- Architecture
- Mesh Host
- Agent Architecture
- Operator & Teams
- Memory System
- Triggering & Automation
- Cost Tracking & Budgets
- Security Model
- CLI Reference
- Configuration
- MCP Tool Support
- Testing
- Dependencies
- Project Structure
- Design Principles
- FAQ
What is OpenLegion?
OpenLegion is a secure, self-hosted AI agent runtime for running fleets of autonomous AI agents in production. Each agent runs in its own hardened Docker container (or microVM), with its own memory, tools, schedule, and budget. Agents never hold API keys - every LLM and API call routes through a central credential vault that also enforces per-agent spend limits. A trusted mesh host coordinates the fleet through shared state and pub/sub events, with permission ACLs checked on every cross-agent action.
It is source-available under the PolyForm Perimeter License 1.0.1: you can self-host it for free — including for your own commercial, internal operations — read the entire ~77,000-line codebase, and audit it in a day. The one thing you can't do is offer it to others as a product or service that competes with OpenLegion (e.g. reselling it as your own hosted/managed service). Managed hosting is available from OpenLegion LLC for teams that prefer not to run their own infrastructure. OpenLegion is built as a production- and team-focused OpenClaw alternative - it keeps the autonomy of single-user assistant frameworks and adds container isolation, credential vaulting, per-agent budgets, and auditable workflows.
In one line: a multi-agent framework where security, isolation, and cost control are part of the architecture, not an afterthought.
Who is OpenLegion for?
- Developers and AI builders who want a programmable, self-hosted runtime for multi-agent systems instead of a hosted black box.
- Self-hosters and technical founders who need AI agents that run on their own infrastructure, with data and credentials never leaving their control.
- Teams running agents in production that need per-agent budgets, permission ACLs, Docker/microVM isolation, and a codebase small enough to audit.
- OpenClaw, CrewAI, LangGraph, and AutoGen users who have outgrown single-user or library-only setups and now need security and cost controls around their agents.
- Managed-hosting customers who want the same runtime without operating the infrastructure themselves.
If you just want a personal assistant on one machine, a single-user tool is simpler. OpenLegion is for when agents become shared, always-on, or handle anything you cannot afford to leak or overspend.
Quick Start
Requirements: Python 3.10+, Docker (running), at least one LLM provider key (Anthropic, OpenAI, Gemini, Moonshot, Deepseek, xAI, Groq, Minimax, Zai, or Ollama — the setup wizard walks you through it; existing Anthropic Claude CLI or OpenAI Codex CLI logins can be imported).
macOS / Linux:
git clone https://github.com/openlegion-ai/openlegion.git && cd openlegion
./install.sh # checks deps, creates venv, makes CLI global
openlegion start # inline setup on first run, then launch agents
Windows (PowerShell):
git clone https://github.com/openlegion-ai/openlegion.git
cd openlegion
powershell -ExecutionPolicy Bypass -File install.ps1
openlegion start
Windows note: Docker Desktop (not Docker Engine) is required on Windows. WSL2 must be enabled. See Docker's WSL2 backend guide if containers fail to start.
First install downloads dependencies into a venv; this may take several minutes the first time. Subsequent installs are fast.
First run: On the very first
openlegion start, Docker builds theopenlegion-agent:latestandopenlegion-browser:latestimages from theDockerfile.agentandDockerfile.browserin the repo root. The browser image is significantly larger (Camoufox + KasmVNC + Openbox + Xvnc) and can take several minutes with no progress output — this is normal. Subsequent starts are fast.Background mode:
openlegion start -dpolls for startup for up to 90 seconds. If a Docker image build is needed on first run, this timeout may be exceeded — wait for the build to finish and re-runopenlegion start -d.First run also creates:
config/agents.yaml,config/permissions.json,config/mesh.yaml, agent volumes, and anoperatoragent that you didn't define — that's a built-in fleet-management agent (lighter resource caps, excluded from cost/quota math). See CLI Reference foropenlegion resetif you want to wipe state and start over.Need help? See the full setup guide for platform-specific instructions and troubleshooting.
Common commands
# Start (interactive REPL); use /add inside the REPL to register more agents
openlegion start
# Run in background
openlegion start -d
openlegion chat <agent_name> # connect from another terminal to an agent you created
openlegion stop # clean shutdown
openlegion reset # destructive: wipe config/, data/, agent_tools/* (keeps .env)
OpenLegion vs OpenClaw
OpenLegion is an OpenClaw alternative built for production and team use. OpenClaw is the most popular personal AI assistant framework (200K+ GitHub stars) and is genuinely great for single-user setups. The trade-off shows up once agents become shared, always-on, or handle untrusted input - areas where it has documented security and cost gaps:
- 42,000+ exposed instances with no authentication (Bitsight, Feb 2026)
- 341 malicious skills found stealing user data (Koi Security via The Hacker News)
- CVE-2026-25253: one-click remote code execution
- No per-agent cost controls — runaway spend is a real risk
- No deterministic routing — a CEO agent (LLM) decides what runs next
- API keys stored directly in agent config
OpenLegion was designed from day one assuming agents will be compromised.
| OpenClaw | OpenLegion | |
|---|---|---|
| API key storage | Agent config files | Vault proxy — agents never see keys |
| Agent isolation | Process-level | Docker container per agent + microVM option |
| Cost controls | None | Per-agent daily + monthly budget caps |
| Multi-agent routing | LLM CEO agent | Fleet model — blackboard + pub/sub coordination |
| LLM providers | Broad | 100+ via LiteLLM with health-tracked failover |
| Test coverage | Minimal | 5800+ tests across 193 test files including full Docker E2E |
| Codebase size | 430,000+ lines | ~77,000 lines in src/ — still auditable in a day |
What It Does
OpenLegion is an autonomous AI agent framework for running multi-agent fleets in isolated Docker containers. Each agent gets its own memory, tools, schedule, and budget — coordinated through blackboard shared state and pub/sub events with no LLM routing layer.
Chat with your agent fleet via Telegram, Discord, Slack, WhatsApp, or CLI. Agents act autonomously via cron schedules, webhooks, and heartbeat monitoring — without being prompted.
5800+ tests passing across 193 test files. Fully auditable in a day. No LangChain. No Redis. No Kubernetes. No CEO agent. Source-available (PolyForm Perimeter).
Security by architecture — every agent runs in an isolated Docker container (microVM when available). API keys live in the credential vault — agents call through a proxy and never handle credentials directly. Defense-in-depth with 6 security layers.
Production-grade cost control — per-agent LLM token tracking with enforced daily and monthly budget caps at the vault layer. Agents physically cannot spend what you haven't authorized. View live spend with
/costsin the REPL.Acts autonomously — cron schedules, heartbeat probes, and webhook triggers let agents work without being prompted.
Self-aware and self-improving — agents understand their own permissions, budget, fleet topology, and system architecture via auto-generated
SYSTEM.mdand live runtime context. They learn from tool failures and user corrections, injecting past learnings into future sessions.Self-extends — agents write their own Python tools at runtime and hot-reload them. Agents can also spawn sub-agents for specialized work.
Multi-channel — connect agents to Telegram, Discord, Slack, and WhatsApp. Also accessible via CLI and API.
Real-time dashboard — web-based fleet observability with consolidated navigation, slide-over chat panels, keyboard command palette, grouped request traces, live event streaming, streaming broadcast with real-time per-agent responses, LLM prompt/response previews, agent management, agent settings editor (personality, instructions, preferences, heartbeat rules, memory, activity logs, learnings), cost charts, cron management, and embedded KasmVNC viewer for persistent browser agents.
Tracks and caps spend — per-agent LLM cost tracking with daily and monthly budget enforcement.
Fails over across providers — configurable model failover chains cascade across LLM providers with per-model health tracking and exponential cooldown.
Token-level streaming — real-time token-by-token LLM responses across CLI, dashboard, Telegram, Discord, and Slack with progressive message editing and graceful non-streaming fallback.
Architecture
OpenLegion's architecture separates concerns across three trust zones: untrusted external input, sandboxed agent containers, and a trusted mesh host that holds credentials and coordinates the fleet. All inter-agent communication flows through the mesh. Agents do not contact each other directly — no direct peer-to-peer connections.
┌──────────────────────────────────────────────────────────────────────────┐
│ User Interface │
│ │
│ CLI (click) Webhooks Cron Scheduler │
│ - setup - POST /webhook/ - "0 9 * * 1-5" │
│ - start (REPL) hook/{id} - "every 30m" │
│ - stop / status - Trigger agents - Heartbeat pattern │
│ - chat / status │
└──────────────┬──────────────────┬──────────────────┬────────────────────┘
│ │ │
▼ ▼ ▼
┌──────────────────────────────────────────────────────────────────────────┐
│ Mesh Host (FastAPI) │
│ Port 8420 (default) │
│ │
│ ┌────────────┐ ┌─────────┐ ┌───────────┐ ┌────────────────────────┐ │
│ │ Blackboard │ │ PubSub │ │ Message │ │ Credential Vault │ │
│ │ (SQLite) │ │ │ │ Router │ │ (API Proxy) │ │
│ │ │ │ Topics, │ │ │ │ │ │
│ │ Key-value, │ │ subs, │ │ Permission │ │ LLM, image_gen, │ │
│ │ versioned, │ │ notify │ │ enforced │ │ Apollo, Hunter, │ │
│ │ TTL, GC │ │ │ │ routing │ │ Brave Search │ │
│ └────────────┘ └─────────┘ └───────────┘ └────────────────────────┘ │
│ │
│ ┌──────────────┐ ┌──────────────┐ ┌──────────────┐ │
│ │ Permission │ │ Container │ │ Cost │ │
│ │ Matrix │ │ Manager │ │ Tracker │ │
│ │ │ │ │ │ │ │
│ │ Per-agent │ │ Docker life- │ │ Per-agent │ │
│ │ ACLs, globs, │ │ cycle, nets, │ │ token/cost, │ │
│ │ default deny │ │ volumes │ │ budgets │ │
│ └──────────────┘ └──────────────┘ └──────────────┘ │
└──────────────────────────────────────────────────────────────────────────┘
│
│ Docker Network (bridge by default; host opt-in via
│ OPENLEGION_HOST_NETWORK=1 / _BROWSER_ALLOW_HOST_NETWORK=1)
│
┌─────────┼──────────┬──────────────────────┐
▼ ▼ ▼ ▼
┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐
│ Agent A │ │ Agent B │ │ Agent C │ ... │ Agent N │
│ :8400 │ │ :8400 │ │ :8400 │ │ :8400 │
└─────────┘ └─────────┘ └─────────┘ └─────────┘
Each agent: isolated Docker container, own /data volume,
own memory DB, own workspace, 384MB RAM, 0.15 CPU. FastAPI
listens on :8400 *inside* the container; host port is allocated
dynamically by the runtime. The built-in `operator` agent runs
with lighter caps (128MB / 0.05 CPU).
To reach an agent from the host, use the mesh proxy at :8420, not the agent's internal :8400.
Trust Zones
| Level | Zone | Description |
|---|---|---|
| 0 | Untrusted | External input (webhooks, user prompts). Sanitized before reaching agents. |
| 1 | Sandboxed | Agent containers. Isolated filesystem, no credentials. External network access gated through SSRF-protected mesh proxy — restricted Docker bridge with NAT egress; private/CGNAT/IPv4-mapped/6to4/Teredo ranges blocked by http_tool.py. The shared browser container has its own iptables egress filter (set up by entrypoint with NET_ADMIN, then dropped) — that is the authoritative SSRF control for browser-initiated traffic. |
| 2 | Trusted | Mesh host. Holds credentials, manages containers, routes messages. |
Mesh Host
The mesh host is the central coordination layer. It runs on the host machine as a single FastAPI process.
Blackboard (Shared State Store)
SQLite-backed key-value store with versioning, TTL, and garbage collection.
Team agents' blackboard access is automatically scoped to teams/{name}/* —
agents use natural keys (e.g. tasks/research_abc123) while the MeshClient
transparently namespaces them under the team. Solo agents have no automatic
team scoping; they can access only keys explicitly granted via ACL.
The on-disk prefix is teams/{name}/*.
| Namespace | Purpose | Example |
|---|---|---|
tasks/* |
Task assignments | tasks/research_abc123 |
context/* |
Shared agent context | context/prospect_acme |
signals/* |
Inter-agent signals | signals/research_complete |
history/* |
Append-only audit log | history/action_xyz |
These prefixes are conventions, not enforced schemas — agents can write any key that matches their blackboard_write glob.
Credential Vault (API Proxy)
Agents never hold API keys. All external API calls route through the mesh.
The vault uses a two-tier prefix system: OPENLEGION_SYSTEM_* for LLM
provider keys (never agent-accessible) and OPENLEGION_CRED_* for agent-tier
tool/service keys. Budget limits are enforced before dispatching LLM calls
and token usage is recorded after each response. OPENLEGION_SYSTEM_* credentials are never resolvable by agents (mesh-proxy-only). OPENLEGION_CRED_* are gated by per-agent allowed_credentials. Misclassifying one as the other yields silent invisibility.
Model Failover
Configurable failover chains cascade across LLM providers transparently.
ModelHealthTracker applies exponential cooldown per model (transient errors:
60s → 300s → 1500s, billing/auth errors: 1h). Streaming failover is supported — if streaming fails mid-response (including empty/zero-length responses that indicate upstream provider failure),
the next model in the chain retries the full request from the start.
Permission Matrix
Every inter-agent operation is checked against per-agent ACLs. The shape — agents call the blackboard with natural keys (e.g. read_blackboard("tasks/foo")) and MeshClient transparently namespaces them under the active team, so the patterns below are matched against the resolved key (teams/myteam/tasks/foo):
{
"researcher": {
"can_message": ["*"],
"can_publish": ["research_complete"],
"can_subscribe": ["new_lead"],
"blackboard_read": ["teams/myteam/*"],
"blackboard_write": ["teams/myteam/*"],
"allowed_apis": ["llm", "brave_search"],
"allowed_credentials": ["brightdata_*"],
"browser_actions": null
}
}
Matching is exact match (or *) for can_message, can_publish, and can_subscribe, and glob (fnmatch) for blackboard_read, blackboard_write, and allowed_credentials.
browser_actions semantics: null (default) = all known actions allowed; ["*"] = explicit allow-all; specific list (e.g. ["browser_navigate", "browser_screenshot"]) = narrow allowlist; [] = deny all browser use even when can_use_browser is true.
Blackboard patterns use the teams/{name}/* namespace. When an agent joins a
team, it receives read/write access to that namespace. Solo agents have no
automatic team scoping; they can access only keys explicitly granted via ACL.
Team scope is enforced by default (env: OPENLEGION_TEAM_SCOPE_MODE=enforce). Agents in different teams cannot read each other's blackboard without explicit permission.
Container Manager
Agent containers are slim — no browser. Browsing is handled by a shared browser service container (Camoufox + KasmVNC).
Agent container:
- Image:
openlegion-agent:latest(Python 3.12, system tools — no browser) - Network: Bridge with port mapping (macOS/Windows) or host network (Linux)
- Volume:
openlegion_data_{agent_id}mounted at/data(agent names with spaces/special chars are sanitized) - Resources: 384MB RAM, 0.15 CPU (agents are I/O-bound — waiting on LLM APIs). The built-in
operatoragent runs at 128MB / 0.05 CPU. - Security:
no-new-privileges,cap_drop=[ALL],read_only=True,tmpfs=/tmp, non-root UID 1000 - Port: 8400 (FastAPI, inside the container; host port allocated dynamically)
Browser service container (shared across all agents):
- Image:
openlegion-browser:latest(Camoufox stealth browser + KasmVNC) - Resources: 2–8GB RAM, 1–2 CPU, 512MB–2GB shared memory — scaled by
OPENLEGION_MAX_AGENTSplan tier - Ports: 8500 (browser API) is the only exposed port. Per-agent KasmVNC instances run internally on 6100..6163 and are reverse-proxied by the mesh at
/agent-vnc/{agent_id}/...(no direct port exposed to the host). - Capacity: autoscales by
OPENLEGION_MAX_AGENTS— ≤1 agent → 1 session; ≤5 → 5; ≤15 → min(N, 10); >15 → min(N, 30). Absolute cap 64 viaOPENLEGION_BROWSER_MAX_CONCURRENT(legacy aliasMAX_BROWSERS). Managed deployments may layer their own plan tiers on top of these autoscale rules. Restart the browser service to apply a change.
Browser Capabilities
Beyond the basic navigation/screenshot/click tools, the browser service ships with:
- CAPTCHA solving. Optional 2captcha or capsolver provider configured per-fleet via
CAPTCHA_SOLVER_KEY+CAPTCHA_SOLVER_PROVIDER. Solver credentials (CAPTCHA_SOLVER_KEY,CAPTCHA_SOLVER_KEY_SECONDARY,CAPTCHA_SOLVER_PROXY_LOGIN,CAPTCHA_SOLVER_PROXY_PASSWORD) are env-only by design — they bypass theOPENLEGION_CRED_*vault and are stripped fromconfig/settings.jsonat load (_ENV_ONLY_FLAGSinsrc/browser/flags.py). Auto-solve runs afterbrowser_navigate; behavioral / persistent challenges escalate torequest_captcha_helpwhich posts a card to the dashboard for the user to clear via the live VNC viewer. Disabled fleet-wide withCAPTCHA_DISABLED=1. Behavioral / vendor JS challenges that 2captcha and capsolver cannot solve escalate viarequest_captcha_help, surfacing a VNC card to the user. - Per-agent + per-tenant solver cost caps.
CAPTCHA_COST_LIMIT_USD_PER_AGENT_MONTHandCAPTCHA_COST_LIMIT_USD_PER_TENANT_MONTHenforce monthly spend with 50/80/100% threshold alerts. Per-tenant rollups available at/dashboard/api/billing/captcha-rollup(requires a dashboard session cookie and theX-Requested-WithCSRF header). - Fingerprint health monitoring. A rolling per-agent rejection window detects when a fingerprint is "burned" (>50% rejection over the last 10 events across Cloudflare / DataDome / PerimeterX / Imperva / Akamai BMP signals); subsequent CAPTCHA envelopes carry
fingerprint_burn=Trueand aretry_with_fresh_profilehint. Operator clears state manually after profile rotation. - JS-challenge detection. Vendor-specific selectors detect Cloudflare 1xxx / Under Attack / Press & Hold and similar interstitials before the agent attempts to extract content.
- Mobile emulation profiles.
BROWSER_DEVICE_PROFILEenv var (per-agent or fleet-wide) selects a mobile UA + viewport + touch profile when sites gate on desktop fingerprints. Configured via env, not the dashboard. - Session continuity (opt-in).
BROWSER_SESSION_PERSISTENCE_ENABLED=1enables a per-agent storage-state sidecar so cookies and localStorage survive container restarts. Default-off; operator/curl-only management via/dashboard/api/agents/{id}/session. - Two-stage workspace upload.
browser_upload_filereads from the agent's/dataand uploads via a stage-then-apply protocol with idempotency keys and a tmpfs partial reaper, so a half-completed upload can never end up attached to a form. Per-file cap 50 MB (OPENLEGION_UPLOAD_STAGE_MAX_MB), max 5 files per call.
Agent Architecture
Each agent container runs a FastAPI server with endpoints for task assignment, chat, status, capabilities, and results.
┌─────────────────────────────────────────────────────────────┐
│ Agent Container │
│ │
│ FastAPI Server (:8400) │
│ POST /task POST /chat POST /chat/reset │
│ GET /status GET /result GET /capabilities │
│ GET /workspace GET|PUT /workspace/{file} │
│ GET /heartbeat-context │
│ │
│ ┌──────────────────────────────────────────────────────┐ │
│ │ AgentLoop │ │
│ │ │ │
│ │ Task Mode: bounded 20-iteration loop │ │
│ │ Chat Mode: conversational with tool use │ │
│ │ │ │
│ │ Both: LLM call → tool execution → context mgmt │ │
│ └──┬──────────┬──────────┬──────────┬──────────┬───────┘ │
│ │ │ │ │ │ │
│ ┌──▼───┐ ┌──▼───┐ ┌──▼──────┐ ┌─▼──────┐ ┌─▼─────────┐ │
│ │ LLM │ │ Mesh │ │ Tool │ │Work- │ │ Context │ │
│ │Client│ │Client│ │Registry │ │space │ │ Manager │ │
│ │(mesh │ │(HTTP)│ │(builtins│ │Manager │ │(token │ │
│ │proxy)│ │ │ │+custom) │ │(/data/ │ │tracking, │ │
│ └──────┘ └──────┘ └─────────┘ │workspace│ │compact) │ │
│ └─────────┘ └───────────┘ │
└─────────────────────────────────────────────────────────────┘
Task Mode (POST /task)
Accepts a TaskAssignment for task execution. Runs a bounded loop
(max 20 iterations) of decide → act → learn. Returns a TaskResult with
structured output and optional blackboard promotions.
Chat Mode (POST /chat)
Accepts a user message. On the first message, loads bootstrap workspace files
into the system prompt — TEAM.md (team members only), SYSTEM.md, INSTRUCTIONS.md,
SOUL.md, USER.md, MEMORY.md — injects
a live Runtime Context
block (permissions, budget, fleet, cron), and searches memory for relevant facts.
Executes tool calls in a bounded loop with three caps from loop.py:
CHAT_MAX_TOOL_ROUNDS=30 per turn, CHAT_MAX_TOTAL_ROUNDS=200 total before
auto-compaction kicks in, and _MAX_SESSION_CONTINUES=5 auto-continuations
(hardcoded, not env-overridable — after which the session halts with an error rather than continuing forever).
Tool Loop Detection
Both modes include automatic detection of stuck tool-call loops. A sliding
window tracks recent (tool_name, params_hash, result_hash) tuples and
escalates through three levels:
| Level | Trigger | Action |
|---|---|---|
| Warn | 2nd identical call | System message: "Try a different approach" |
| Block | 4th identical call | Tool skipped, error returned to agent |
| Terminate | 9th call with same params | Loop terminated with failure status |
memory_search is exempt since repeated searches are legitimate. Detection uses SHA-256 hashes of
canonicalized parameters and results over a 15-call sliding window.
Built-in Tools
| Tool | Purpose |
|---|---|
run_command |
Shell command execution with timeout |
read_file |
Read file contents from /data |
write_file |
Write/append file in /data |
list_files |
List/glob files in /data |
http_request |
HTTP GET/POST/PUT/DELETE/PATCH |
browser_navigate |
Open URL, extract page text via shared browser service. Auto-detects CAPTCHAs and may auto-solve or surface a help envelope. |
browser_get_elements |
Accessibility tree snapshot with element refs (e1, e2, ...) |
browser_find_text |
Locate elements by visible/accessible name (Unicode case-fold match) |
browser_screenshot |
Capture page screenshot |
browser_click |
Click element by ref or CSS selector |
browser_click_xy |
Click at viewport-relative pixel coordinates (canvas / non-accessible widgets) |
browser_type |
Fill input by ref or CSS selector |
browser_fill_form |
Fill multiple labeled form fields in one call |
browser_hover |
Hover over element to trigger dropdowns/tooltips |
browser_scroll |
Scroll page up/down or scroll element into view |
browser_wait_for |
Wait for CSS selector to appear/disappear |
browser_press_key |
Press keyboard key or shortcut (Escape, Enter, Control+a) |
browser_go_back / browser_go_forward |
Navigate browser history |
browser_open_tab |
Open a URL in a new tab (becomes the active page) |
browser_switch_tab |
List open tabs or switch to a specific tab |
browser_upload_file |
Upload workspace files to a file-input element (1-5 files) |
browser_download |
Click a ref to trigger a download and save it as an artifact (≤50MB) |
browser_inspect_requests |
List recent network request URLs (redacted; no bodies or headers) |
browser_reset |
Reset browser session (profile preserved) |
browser_detect_captcha |
CAPTCHA detection (usually not needed — browser_navigate auto-detects) |
browser_solve_captcha |
Explicitly request a CAPTCHA solve on the current page |
request_captcha_help |
Hand a behavioral / persistent CAPTCHA to the user via the dashboard viewer |
request_browser_login |
Navigate browser to a URL and send a VNC login card to the user for manual credential entry |
generate_image |
Generate an image via Gemini or DALL-E 3 and save as an artifact |
memory_search |
Hybrid search across workspace files and structured DB |
memory_save |
Save fact to workspace and structured memory DB |
web_search |
Search the web via DuckDuckGo (HTML scrape — no API key, but subject to occasional rate limits / CAPTCHAs) |
notify_user |
Send notification to user across all connected channels |
list_agents |
Discover agents on your team (solo agents see only themselves) |
read_blackboard |
Read from the shared blackboard |
write_blackboard |
Write to the shared blackboard |
list_blackboard |
Browse blackboard entries by prefix |
publish_event |
Publish event to mesh pub/sub |
subscribe_event |
Subscribe to a pub/sub topic at runtime |
hand_off |
Hand a work item to another agent via structured coordination protocol |
check_inbox |
Check for pending work items handed off by other agents |
update_status |
Update the status of an in-progress work item visible to coordinators |
complete_task |
Mark a coordination work item as complete with a result |
watch_blackboard |
Watch blackboard keys matching a glob pattern |
claim_task |
Atomically claim a task from the shared blackboard |
save_artifact |
Save deliverable file and register on blackboard |
update_workspace |
Update identity files (SOUL.md, INSTRUCTIONS.md, USER.md, HEARTBEAT.md) |
set_cron |
Schedule a recurring job (set heartbeat=true for autonomous wakeups) |
list_cron / remove_cron |
Manage scheduled jobs |
create_tool |
Write a new Python tool at runtime |
reload_tools |
Hot-reload all tools |
spawn_fleet_agent |
Spawn an ephemeral sub-agent in a new container |
spawn_subagent |
Spawn a lightweight in-container subagent for parallel subtasks |
list_subagents |
List active subagents and their status |
wait_for_subagent |
Wait for a subagent to complete and return its result |
vault_generate_secret |
Generate and store a random secret (returns opaque handle) |
vault_list |
List credential names (names only, never values) |
wallet_get_address |
Get Ethereum/Solana wallet address for an agent (requires [wallet] extras) |
wallet_get_balance |
Get wallet balance (ETH or SOL) (requires [wallet] extras) |
wallet_read_contract |
Read data from an Ethereum smart contract (requires [wallet] extras) |
wallet_transfer |
Transfer ETH or SOL to an address (requires [wallet] extras) |
wallet_execute |
Execute an Ethereum smart contract function (requires [wallet] extras) |
get_system_status |
Query own runtime state: permissions, budget, fleet, cron, health |
read_agent_history |
Read another agent's conversation logs |
Custom tools are Python functions decorated with @tool, auto-discovered
from the agent's tools_dir at startup. Agents can also create new tools
at runtime and hot-reload them.
Agents also support MCP (Model Context Protocol) — any MCP-compatible tool server can be plugged in via config, giving agents access to databases, filesystems, APIs, and more without writing custom tools.
Operator & Teams
OpenLegion is built around one autonomous loop: you talk to the operator, the operator builds teams, and the fleet carries the work toward a goal — pulling you in for credentials, approvals, or genuine forks, not for routine coordination.
Operator — a built-in agent (
operator; lighter caps, excluded from cost/quota math) that is your entry point. It creates agents and teams, sets goals, watches progress, and reallocates work. It is the only agent that can apply fleet templates, install a team lead, or confirm hard config edits. It is not a message router — work still flows peer-to-peer.Team lead — every non-solo team has one accountability owner. The mesh auto-appoints the first non-operator member the moment a team reaches two or more real members (on create and on add), and re-appoints another member automatically whenever the current lead leaves by any path (removed, offboarded, evicted by a move, or deleted). The lead is an ordinary member under the ordinary security model — it plans and stewards, gaining no routing privilege. Team stewardship (goal-coverage probes, standups, blocked-task escalation, delivery review) is gated on there being a lead, so a leaderless team is never silently left un-stewarded.
Goals that reach the workers — the operator sets a team's north star and up to ten success criteria with
set_team_goal. The goal is persisted, mirrored into a## Goalsection of the team'sTEAM.md, and pushed to members — so a running worker picks it up on its next turn, not only at onboarding. Redirecting a busy team is no longer inert.update_team_context(writes## Context) and team briefs are section-scoped, so updating one never clobbers the goal.A spend envelope per team — teams carry daily/monthly USD budgets the operator reallocates within. Enforcement can be set fail-closed (
OPENLEGION_TEAM_ENVELOPE_FAIL_CLOSED, default fail-open) so a budget-store outage stops spend instead of waving it through.Observe and correct, without leaving chat —
inspect_teams(goal, criteria, budget),inspect_team_spend(per-team spend by period),inspect_agentsatprofiledepth (budget headroom + track record), andassess_team_progress(success criteria bundled with the team's actual completed outputs, for an evidence-grounded verdict). The dashboard Team Room shows the goal side at a glance: north star, success criteria, probe names, and each member's current and blocked work. (Per-team spend is aninspect_team_spend/ mesh read, not shown in the Team Room.)
Honest failure is part of the loop: when a worker asks for a credential that cannot be delivered,
request_credential returns an explicit failure envelope instead of reporting success — so
autonomous work parks visibly rather than silently forever.
Today's guardrails are spend caps + Docker/microVM isolation + SSRF egress filtering. By design the current posture does not yet gate individual outbound actions (publish / email / purchase) behind a broker; that action-broker layer and an immutable human spend ceiling are planned follow-ups. See
docs/plans/2026-07-16-autonomous-team-delivery.mdfor the roadmap and current implementation status.
Memory System
Five layers give agents persistent, self-improving memory:
Layer 5: Context Manager ← Manages the LLM's context window
│ Monitors token usage
│ Proactive flush at 60% capacity
│ Auto-compacts at 70% capacity
│ Extracts facts before discarding messages
│
Layer 4: Learnings ← Self-improvement through failure tracking
│ learnings/errors.md (tool failures with context)
│ learnings/corrections.md (user corrections and preferences)
│ Auto-injected into system prompt each session
│
Layer 3: Workspace Files ← Durable, human-readable storage
│ Bootstrap files loaded into the first-message system prompt:
│ TEAM.md (team members only), SYSTEM.md, INSTRUCTIONS.md,
│ SOUL.md, USER.md, MEMORY.md
│ Other workspace files:
│ HEARTBEAT.md (autonomous monitoring rules)
│ INTERFACE.md, AGENTS.md (channel + roster context)
│ memory/YYYY-MM-DD.md (daily session logs)
│ FTS5 keyword search across files
│
Layer 2: Structured Memory DB ← Hierarchical vector database
│ SQLite + sqlite-vec + FTS5
│ Hybrid search: 0.7 vector similarity + 0.3 FTS5 keyword
│ Auto-categorization with category-scoped search
│ 3-tier retrieval: categories → scoped facts → flat fallback
│ Reinforcement scoring with access-count boost + recency decay
│
Layer 1: Salience Tracking ← Prioritizes important facts
Access count, decay score, last accessed timestamp
High-salience facts auto-surface in initial context
Write-Then-Compact Pattern
Before the context manager discards messages, it:
- Asks the LLM to extract important facts from the conversation
- Stores facts in both
MEMORY.mdand the structured memory DB - Summarizes the conversation
- Replaces message history with: summary + last 3–4 messages (role-aware, preserving message alternation invariant)
Nothing is permanently lost during compaction.
Cross-Session Memory
Facts saved with memory_save are stored in both the workspace (daily log)
and the structured SQLite database. After a reset or restart, memory_search
retrieves them via hybrid search:
Session 1: User says "My cat's name is Whiskerino"
Agent saves to daily log + structured DB
═══ Chat Reset ═══
Session 2: User asks "What is my cat's name?"
Agent recalls "Whiskerino" via memory_search
Triggering & Automation
Agents act autonomously through trigger mechanisms running in the mesh host (not inside containers, so they survive container restarts).
Cron Scheduler
Persistent cron jobs that dispatch messages to agents on a schedule. Agents
can schedule their own jobs using the set_cron tool.
Supports 5-field cron expressions (minute hour dom month dow), interval
shorthand (every 30m, every 2h), and state persisted to config/cron.json.
Cron jobs can also dispatch in tool-mode (tool_name + tool_params), invoking a built-in tool directly without an LLM round — useful for cheap deterministic monitoring. For example, set_cron with tool_name="http_request" and tool_params={"url": "...", "method": "GET"} polls an endpoint on a schedule without spending tokens.
Heartbeat System
Cost-efficient autonomous monitoring. Heartbeat jobs run cheap deterministic probes first — disk usage, pending signals, pending tasks — and only dispatch to the agent (costing LLM tokens) when probes detect something actionable.
When a heartbeat fires, the agent receives enriched context: its HEARTBEAT.md rules, recent daily logs, probe alerts, and actual pending signal/task content — all in a single message. If HEARTBEAT.md is the default scaffold, no recent activity exists, and no probes triggered, the dispatch is skipped entirely (zero LLM cost).
This 5-stage architecture (scheduler → probes → context → policy → action) makes always-on agents economically viable.
Task Hand-off & Auto-close
Handed-off tasks auto-transition to terminal status (done/failed) only when the wake chain carries x-task-id (via hand_off's task_id propagation). Legacy callers, heartbeats, and manual chats won't auto-close — intentional.
Webhook Endpoints
Named webhook URLs that dispatch payloads to agents. Create one from the dashboard (System → Automation) or via the mesh API; the URL it returns is what you POST to. Payloads are sanitized and capped at 1MB.
# The full URL is returned in the `url` field of the `POST /api/webhooks`
# response (and listed on System → Automation in the dashboard). Pattern:
# {base}/webhook/hook/{hook_id} e.g. http://localhost:8420/webhook/hook/hook_3f9a1c8b2d4e6f70
curl -X POST "$WEBHOOK_URL" \
-H "Content-Type: application/json" \
-d '{"event": "push", "repo": "myproject"}'
Cost Tracking & Budgets
Every LLM call is tracked at the Credential Vault layer. Per-agent budgets
prevent runaway spend. View costs from the interactive REPL (/costs) or
configure budgets in config/agents.yaml:
agents:
researcher:
budget:
daily_usd: 5.00
monthly_usd: 100.00
When an agent exceeds its budget, the vault rejects LLM calls with an error instead of forwarding them to the provider.
CAPTCHA solver spend is tracked separately from LLM spend (per-agent and per-tenant USD caps with 50/80/100% threshold alerts) — see Browser Capabilities. The two budgets do not share a pool.
Security Model
Defense-in-depth with six layers:
| Layer | Mechanism | What It Prevents |
|---|---|---|
| Runtime isolation | Docker containers (default); Docker Sandbox microVMs with --sandbox (Docker Desktop 4.58+ required) |
Agent escape, kernel exploits |
| Container hardening | Non-root user, no-new-privileges, memory/CPU limits | Privilege escalation, resource abuse |
| Credential separation | Vault holds keys, agents call via proxy | Key leakage, unauthorized API use |
| Permission enforcement | Per-agent ACLs for messaging, blackboard, pub/sub, APIs | Unauthorized data access |
| Input validation | Path traversal prevention, SSRF blocking, safe condition eval (no eval()), token budgets, iteration limits, rate limiting |
Injection, runaway loops, network abuse |
| Unicode sanitization | Invisible character stripping at ~110 call sites across 17 source files, covering all external input boundaries | Prompt injection via hidden Unicode |
Per-agent rate limits on 17 mesh endpoints (token bucket; HTTP 429 + audit log on overflow). See docs/security.md for the full table.
Containers run with no-new-privileges, cap_drop=ALL, read_only=True, /tmp tmpfs (100MB noexec/nosuid), UID 1000. Tools self-authored by agents pass AST validation (23 forbidden imports, 16 forbidden calls, 11 forbidden attrs).
Dual Runtime Backend
OpenLegion supports two isolation levels:
| Docker Containers (default) | Docker Sandbox microVMs | |
|---|---|---|
| Isolation | Shared kernel, namespace separation | Own kernel per agent (hypervisor) |
| Escape risk | Kernel exploit could escape | Hypervisor boundary — much harder |
| Performance | Native speed | Near-native (Apple Virtualization.framework on macOS / Hyper-V on Windows) |
| Requirements | Any Docker install | Docker Desktop 4.58+ |
| Enable | openlegion start |
openlegion start --sandbox |
Docker containers (default) run agents as non-root with no-new-privileges, 384MB memory limit, 0.15 CPU cap, and no host filesystem access. Browser operations are handled by a shared browser service container (2–8GB RAM, 0.15–4.0 CPU — scaled by fleet size). This is secure for most use cases.
Docker Sandbox microVMs give each agent its own Linux kernel via Apple Virtualization.framework (macOS) or Hyper-V (Windows). Even if an agent achieves code execution, it's trapped inside a lightweight VM with no visibility into other agents or the host. Use this when running untrusted code or when compliance requires hypervisor isolation.
# Default: container isolation (works everywhere)
openlegion start
# Maximum security: microVM isolation (Docker Desktop 4.58+ required)
openlegion start --sandbox
Check compatibility: Run
docker sandbox version— if it returns a version number, your Docker Desktop supports sandboxes. If not, update Docker Desktop to 4.58+.
CLI Reference
openlegion [--verbose/-v] [--quiet/-q] [--json]
├── start [--config PATH] [-d] [--sandbox] [--port PORT] # Start runtime + interactive REPL (inline setup on first run)
├── stop # Stop the runtime + agent containers
├── chat [name] [--port PORT] # Connect to a running agent
├── status [--port PORT] [--wide/-w] [--watch N] [--json] # Show agent status
├── teams [--port PORT] [--json] # List active teams
├── team <team_id> [--port PORT] [--json] # Show one team (members, blockers, task counts)
├── tasks [--agent X] [--team Y] [--status S] [--port PORT] [--json] # List durable task records
├── pending [--port PORT] [--json] # List pending actions awaiting confirmation
├── confirm <nonce> [--port PORT] # Confirm a pending action
├── cancel <nonce> [--port PORT] # Cancel a pending action
├── reset [-y] # DESTRUCTIVE: stop everything and wipe config/, data/, agent_tools/* (keeps .env)
├── version [--verbose/-v] # Show version and environment info
└── wallet # Manage agent wallets (derives EVM + Solana from one master seed)
├── init # Generate the master wallet seed (shown once; refuses to regenerate while ``OPENLEGION_SYSTEM_WALLET_MASTER_SEED`` is set)
└── show [agent_id] # Show wallet addresses
Agent management, credentials, blackboard, cron, and channels are managed via REPL commands (below) inside a running session, or via the web dashboard at
http://localhost:8420(default port; change with--portflag ormesh.portinconfig/mesh.yaml).
Interactive REPL Commands
@agent <message> Send message to a specific agent
/use <agent> Switch active agent
/agents List all running agents
/add Add a new agent (hot-adds to running system)
/agent [edit|view] Agent overview, config editing, workspace files
/edit [name] Edit agent settings (model, browser, budget)
/remove [name] Remove an agent
/restart [name] Restart an agent container
/status Show agent health
/broadcast <msg> Send message to all agents
/steer <msg> Inject message into busy agent's context
/history [agent] Show recent conversation messages
/costs Show today's LLM spend + context usage + model health
/blackboard [list|get|set|del] View/edit shared blackboard entries
/queue Show agent task queue status
/cron [list|del|pause|resume|run] Manage cron jobs
/team [list|use|info] Manage multi-team namespaces
/credential [add|list|remove] Manage API credentials
/traces [id] Show recent request traces
/logs [--level LEVEL] Show recent runtime logs
/addkey <svc> [key] Add an API credential to the vault
/removekey [name] Remove a credential from the vault
/reset Clear conversation with active agent
/quit Exit and stop runtime
Aliases: /exit = /quit, /agents = /status, /debug = /traces
All /edit and /agent edit changes apply immediately. Soft-field edits (instructions, soul, heartbeat, heartbeat_schedule, interface, role) are undoable for 5 minutes; hard-field edits (model, permissions, budget, thinking) are undoable for 30 minutes.
Team Templates
Templates are offered during first-run setup (via openlegion start):
| Template | Agents | Description |
|---|---|---|
starter |
assistant | Single general-purpose agent |
sales |
researcher, qualifier, outreach | Sales pipeline team |
devteam |
pm, engineer, reviewer | Software development team |
content |
researcher, writer, editor | Content creation team |
deep-research |
scout, analyst, writer | Deep research and analysis team |
monitor |
watcher, analyst | Autonomous monitoring agent |
competitive-intel |
scout | Market and competitor analysis |
lead-enrichment |
enricher, formatter | Lead data enrichment |
price-intelligence |
crawler, analyst | Price monitoring and analysis |
review-ops |
monitor, responder | Review and feedback management |
social-listening |
monitor, writer | Social media monitoring |
opportunity-finder |
gap-scout, evaluator, modeler | Market opportunity discovery |
research |
researcher | General-purpose research agent |
Configuration
TEAM.md — Per-Team Context
Each team has its own TEAM.md stored in
config/teams/{name}/team.md.
The file is mounted into team member agents' containers and loaded into
their system prompts. Solo agents (not on a team) do not receive any
TEAM.md.
The operator maintains it through section-scoped writers: set_team_goal owns the ## Goal
section (north star + success criteria), update_team_context owns ## Context, and team briefs
own their own named sections — so writing one never clobbers another, and a goal change reaches
running members on their next turn (see Operator & Teams).
# TEAM.md
## Goal
**North star:** Grow the newsletter to 10,000 engaged subscribers.
**Success criteria:**
- 10,000 confirmed opt-in subscribers
- ≥35% average open rate sustained for 4 weeks
## Context
Ship the email personalization pipeline this week. Audience: B2B SaaS founders.
No cold outreach to .edu or .gov domains.
Workspace files have per-file size caps (4–16 KB; HEARTBEAT.md uncapped). See docs/configuration.md for the table.
config/mesh.yaml — Framework Settings
mesh:
host: "0.0.0.0"
port: 8420
llm:
default_model: "openai/gpt-4o-mini"
embedding_model: "text-embedding-3-small" # "none" to disable vector search
collaboration: true # allow agents to message each other (default: true for new agents)
config/agents.yaml — Agent Definitions
Created automatically by openlegion start (inline setup) or the /add REPL command.
agents:
researcher:
role: "research"
model: "openai/gpt-4o-mini"
tools_dir: "./agent_tools/researcher"
initial_instructions: "You are a research specialist..."
thinking: "medium" # off (default), low, medium, or high
budget:
daily_usd: 5.00
monthly_usd: 100.00
config/permissions.json — Agent Permissions
Per-agent access control with glob patterns for blackboard paths and explicit allowlists for messaging, pub/sub, and API access.
.env — API Keys
Managed automatically by openlegion start (setup wizard) and the /addkey REPL command. You can also edit directly. Uses a two-tier prefix system:
# System tier — LLM provider keys (never accessible by agents)
OPENLEGION_SYSTEM_ANTHROPIC_API_KEY=sk-ant-...
OPENLEGION_SYSTEM_OPENAI_API_KEY=sk-...
OPENLEGION_SYSTEM_MOONSHOT_API_KEY=sk-...
# Agent tier — tool/service keys (access controlled per-agent)
OPENLEGION_CRED_BRAVE_SEARCH_API_KEY=BSA...
# Channel tokens
OPENLEGION_CRED_TELEGRAM_BOT_TOKEN=123456:ABC...
OPENLEGION_CRED_DISCORD_BOT_TOKEN=MTIz...
OPENLEGION_CRED_SLACK_BOT_TOKEN=xoxb-...
OPENLEGION_CRED_SLACK_APP_TOKEN=xapp-...
OPENLEGION_CRED_WHATSAPP_ACCESS_TOKEN=EAAx...
OPENLEGION_CRED_WHATSAPP_PHONE_NUMBER_ID=1234...
OPENLEGION_CRED_WHATSAPP_APP_SECRET=... # X-Hub-Signature-256 verification (production)
# Log format: "json" (default — production / structured) or "text" (human-readable).
# If you see JSON logs locally and want plain text, set this to "text".
OPENLEGION_LOG_FORMAT=text
# Plan limits (0 = unlimited). HTTP 403 once exceeded.
# OPENLEGION_MAX_AGENTS=0
# OPENLEGION_MAX_TEAMS=0
Connecting Channels
Channels are configured via the setup wizard during openlegion start, or by
adding the appropriate tokens to .env directly:
# Telegram
OPENLEGION_CRED_TELEGRAM_BOT_TOKEN=123456:ABC...
# Discord
OPENLEGION_CRED_DISCORD_BOT_TOKEN=MTIz...
# Slack (both required)
OPENLEGION_CRED_SLACK_BOT_TOKEN=xoxb-...
OPENLEGION_CRED_SLACK_APP_TOKEN=xapp-...
# WhatsApp (both required; APP_SECRET required for production webhook signature verification)
OPENLEGION_CRED_WHATSAPP_ACCESS_TOKEN=EAAx...
OPENLEGION_CRED_WHATSAPP_PHONE_NUMBER_ID=1234...
OPENLEGION_CRED_WHATSAPP_APP_SECRET=... # X-Hub-Signature-256 verification
On next openlegion start, a pairing code appears — send it to your bot to link.
MCP Tool Support
OpenLegion supports the Model Context Protocol (MCP) — the emerging standard for LLM tool interoperability. Any MCP-compatible tool server can be plugged into an agent via config, with tools automatically discovered and exposed to the LLM alongside built-in tools.
Note: the
mcpSDK is a core dependency (it powers both in-container stdio servers and the mesh-side gateway for remote connectors) —./install.shinstalls it;pip install -e '.[mcp]'remains as a no-op alias. Agent containers install it viaDockerfile.agent. Wallet tools still require the optional[wallet]group (web3,eth-account,mnemonic,solders,solana).
Configuration
There are two ways to attach MCP servers to an agent:
1. Dashboard (recommended for most users). Open the agent's settings → Config
tab → MCP Servers section. Click + Add MCP server, fill in name +
command + args + env, hit Save. Env values that hold secrets go through a
credential picker (saved as $CRED{name} handles, resolved by the mesh at
agent start — no plaintext on disk or in the API). Per-server status dots
(green / red with the captured error / gray pending) tell you whether each
server actually came up. See docs/mcp.md for the full UX.
2. Fleet template (for repeatable deployments). Add mcp_servers to an
agent in src/templates/<template>.yaml:
agents:
researcher:
role: "research"
model: "openai/gpt-4o-mini"
mcp_servers:
- name: filesystem
command: mcp-server-filesystem
args: ["/data"]
- name: database
command: mcp-server-sqlite
args: ["--db", "/data/research.db"]
env:
DB_PASSWORD: "$CRED{research_db_password}"
The same MCPServerConfig model validates both paths: name matches
^[a-zA-Z0-9][a-zA-Z0-9_-]{0,63}$, command cannot contain $CRED{...}
handles (use env or args instead), case-insensitive duplicate names are
rejected. Each server is launched as a subprocess inside the agent container
using stdio transport; tools are discovered via the MCP protocol and appear
in the LLM's tool list alongside built-in tools.
How It Works
- Agent container reads
MCP_SERVERSfrom environment (set by the runtime) MCPClientlaunches each server subprocess via stdio transport- MCP protocol handshake discovers available tools and their schemas
- Tools are registered in
ToolRegistrywith OpenAI function-calling format - LLM tool calls route through
MCPClient.call_tool()to the correct server - Name conflicts with built-in tools are resolved by prefixing (
mcp_{server}_{tool})
Server Config Options
| Field | Type | Description |
|---|---|---|
name |
string | Server identifier (used for logging and conflict prefixes) |
command |
string | Command to launch the server |
args |
list | Command-line arguments (optional) |
env |
dict | Environment variables for the server process (optional) |
See the full MCP Integration Guide for advanced usage, custom server setup, and troubleshooting.
Testing
# Unit and integration tests (fast, no Docker needed)
pytest tests/ --ignore=tests/test_e2e.py --ignore=tests/test_e2e_chat.py \
--ignore=tests/test_e2e_memory.py --ignore=tests/test_e2e_triggering.py
# E2E tests (requires Docker + API key)
pytest tests/test_e2e.py tests/test_e2e_chat.py \
tests/test_e2e_memory.py tests/test_e2e_triggering.py
# Everything
pytest tests/
Test Coverage
Roughly 5800+ test cases across 193 test files (find tests -name '*.py' | xargs grep -c '^def test_'). Coverage includes every module under src/ — tests/test_FOO.py maps to src/.../FOO.py (see CLAUDE.md for the full mapping). The four tests/test_e2e*.py files require Docker and a real LLM key; everything else runs in CI in under a few minutes per shard.
Dependencies
| Package | Purpose |
|---|---|
| fastapi | HTTP servers (mesh + agent + browser service) |
| uvicorn | ASGI server |
| httpx | Async HTTP client |
| pydantic | Data validation |
| litellm | Multi-provider LLM interface (100+ providers) |
| sqlite-vec | Vector search in SQLite |
| pyyaml | YAML config parsing |
| click | CLI framework |
| docker | Docker API client |
| python-dotenv | .env file loading |
| camoufox | Stealth browser automation (in browser service container only) |
| mcp | MCP tool server client (in agent container only, optional) |
| slack-bolt | Slack channel adapter (optional) |
Dev: pytest, pytest-asyncio, pytest-cov, ruff.
No LangChain. No Redis. No Kubernetes. Real-time web dashboard. Optional channels: python-telegram-bot, discord.py, slack-bolt.
Project Structure
src/
├── cli/
│ ├── main.py # Click commands and entry point
│ ├── config.py # Config loading, Docker helpers, agent management
│ ├── runtime.py # RuntimeContext — full lifecycle management
│ ├── repl.py # REPLSession — interactive command dispatch
│ ├── channels.py # ChannelManager — messaging channel lifecycle
│ └── formatting.py # Tool display, styled output, response rendering
├── agent/
│ ├── __main__.py # Container entry
│ ├── loop.py # Execution loop (task + chat)
│ ├── loop_detector.py # Tool loop detection (warn/block/terminate)
│ ├── tools.py # Tool registry + discovery
│ ├── mcp_client.py # MCP server lifecycle + tool routing
│ ├── memory.py # Hierarchical memory (SQLite + sqlite-vec + FTS5)
│ ├── workspace.py # Persistent workspace + BM25 search
│ ├── context.py # Context manager (token tracking, compaction)
│ ├── llm.py # LLM client (routes through mesh proxy)
│ ├── mesh_client.py # Mesh HTTP client
│ ├── server.py # Agent FastAPI server
│ └── builtins/
│ ├── exec_tool.py # Shell execution
│ ├── file_tool.py # File I/O (read, write, list)
│ ├── http_tool.py # HTTP requests
│ ├── browser_tool.py # Browser automation via shared Camoufox service
│ ├── web_search_tool.py # Web search via DuckDuckGo
│ ├── memory_tool.py # Memory search and save
│ ├── mesh_tool.py # Shared state, fleet awareness, artifacts
│ ├── vault_tool.py # Credential vault operations
│ ├── tool_authoring.py # Runtime tool creation + hot-reload
│ ├── introspect_tool.py # Live runtime state queries
│ ├── subagent_tool.py # Spawn in-process subagents
│ ├── coordination_tool.py # Structured inter-agent coordination (hand_off, check_inbox, update_status, complete_task)
│ ├── image_gen_tool.py # Image generation via Gemini or DALL-E 3
│ └── wallet_tool.py # Wallet operations (Ethereum + Solana)
├── host/
│ ├── server.py # Mesh FastAPI server
│ ├── mesh.py # Blackboard, PubSub, MessageRouter
│ ├── permissions.py # Permission matrix
│ ├── credentials.py # Credential vault + API proxy
│ ├── failover.py # Model health tracking + failover chains
│ ├── runtime.py # RuntimeBackend ABC + Docker/Sandbox backends
│ ├── transport.py # Transport ABC + Http/Sandbox transports
│ ├── cron.py # Cron scheduler + heartbeats
│ ├── webhooks.py # Named webhook endpoints
│ ├── costs.py # Cost tracking + budgets (SQLite)
│ ├── health.py # Health monitor + auto-restart
│ ├── lanes.py # Per-agent FIFO task queues
│ ├── traces.py # Request tracing + grouped summaries
│ ├── transcript.py # Provider-specific transcript sanitization
│ ├── wallet.py # WalletService — Ethereum + Solana operations
│ └── api_keys.py # Named API key management (salted SHA-256 hashes)
├── shared/
│ ├── types.py # All Pydantic models (the contract)
│ ├── utils.py # ID generation, logging, sanitization
│ ├── trace.py # Trace ID generation + correlation
│ ├── models.py # Model cost / context window registry (LiteLLM-backed)
│ ├── redaction.py # Central credential / URL redactor
│ └── operator_playbooks.py # Built-in operator agent prompts
├── browser/
│ ├── __main__.py # Container entry (KasmVNC + Openbox + FastAPI)
│ ├── server.py # Browser service FastAPI server
│ ├── service.py # Camoufox session management (per-agent profiles)
│ ├── captcha.py # CAPTCHA solver core (2captcha, capsolver)
│ ├── captcha_policy.py # Per-site classifier (auto-solve vs hand-off)
│ ├── captcha_cost_counter.py # Per-agent + per-tenant solver cost rollups
│ ├── js_challenge.py # JS-challenge / vendor-fingerprint detection
│ ├── session_persistence.py # Storage-state sidecar (opt-in continuity)
│ ├── profile_schema.py # Profile schema versioning + uBO migration
│ ├── flags.py # Centralized browser flag registry
│ ├── ref_handle.py # RefHandle / ShadowHop element resolver
│ ├── canary.py # Stealth canary probe
│ ├── recorder.py # Behavior recorder
│ ├── stealth.py # Anti-detection configuration
│ ├── timing.py # Human-like timing jitter
│ └── redaction.py # Credential redaction for browser content
├── channels/
│ ├── base.py # Abstract channel with unified UX
│ ├── telegram.py # Telegram adapter
│ ├── discord.py # Discord adapter
│ ├── slack.py # Slack adapter (Socket Mode)
│ └── whatsapp.py # WhatsApp Cloud API adapter
├── dashboard/
│ ├── server.py # Dashboard FastAPI router + API
│ ├── events.py # EventBus for real-time streaming
│ ├── auth.py # Session cookie verification (CSRF + HMAC)
│ ├── notifications.py # Persistent notifications store (SQLite)
│ ├── telemetry.py # SPA telemetry event sink
│ ├── platform_success.py # Per-tenant success scoring
│ ├── conversations.py # Chat-panel conversation state
│ ├── templates/index.html # Dashboard UI (Alpine.js + Tailwind via CDN)
│ └── static/ # CSS + JS assets + avatars
├── setup_wizard.py # Guided setup wizard
├── marketplace.py # Tool marketplace (git-based install/remove)
└── templates/
├── starter.yaml # Single-agent template
├── sales.yaml # Sales pipeline team
├── devteam.yaml # Dev team template
├── content.yaml # Content creation team
├── deep-research.yaml # Deep research and analysis team
├── monitor.yaml # Autonomous monitoring agent
├── competitive-intel.yaml # Competitive intelligence team
├── lead-enrichment.yaml # Lead data enrichment
├── price-intelligence.yaml # Price monitoring and analysis
├── review-ops.yaml # Review and feedback management
├── social-listening.yaml # Social media monitoring
├── opportunity-finder.yaml # Market opportunity discovery
└── research.yaml # General-purpose researcher
config/
├── mesh.yaml # Framework settings
├── agents.yaml # Agent definitions (per-team)
├── permissions.json # Per-agent ACLs
└── teams/ # Multi-team namespaces
Design Principles
| Principle | Rationale |
|---|---|
| Messages, not method calls | Agents communicate through HTTP/JSON. Never shared memory or direct invocation. |
| The mesh is the only door | No agent has network access except through the mesh. No agent holds credentials. |
| Private by default, shared by promotion | Agents keep knowledge private. Facts are explicitly promoted to the blackboard. |
| Explicit failure handling | Domain-specific exceptions propagated with context. No silent error swallowing. |
| Small enough to audit | ~77,000 lines in src/. The entire codebase is still auditable in a day. |
| Tools over features | New capabilities are agent tools, not mesh code. |
| SQLite for all state | Single-file databases. No external services. WAL mode for concurrent reads. |
| Zero vendor lock-in | LiteLLM supports 100+ providers. Markdown workspace files. No proprietary formats. |
FAQ
What is OpenLegion? OpenLegion is a secure, self-hosted AI agent runtime for running fleets of autonomous AI agents in isolated Docker containers. A central mesh host holds all credentials, enforces per-agent budgets and permission ACLs, and coordinates agents through shared state and pub/sub - so agents stay isolated, auditable, and cost-bounded.
Is OpenLegion open source?
OpenLegion is source-available, not open source. It is licensed by OpenLegion LLC under the PolyForm Perimeter License 1.0.1: you can view, modify, and self-host the full codebase for free — including for commercial, internal use — but you cannot provide it to others as a product or service that competes with OpenLegion (for example, reselling it as a hosted or SaaS product). The entire ~77,000-line src/ tree is readable and auditable.
Can I self-host OpenLegion? Yes. Self-hosting is the default and is free under the license, including for your own commercial, internal use. You need Python 3.10+, Docker, and at least one LLM provider key. See the Quick Start and the full setup guide.
Is OpenLegion a good OpenClaw alternative? For production and team use, yes - it adds container/microVM isolation, a credential vault so agents never hold API keys, per-agent budget caps, and permission ACLs on top of autonomous agents. For a single-user assistant on one machine, OpenClaw or a lighter tool may be simpler. See OpenLegion vs OpenClaw.
How does OpenLegion compare to Hermes Agent? Hermes Agent (Nous Research) is an open-source single-user agent known for self-authored, self-improving skills and strong default memory. OpenLegion solves a different problem: running fleets of agents safely in production. It adds per-agent container isolation, a credential vault so agents never hold API keys, per-agent budget caps, and default-deny permission ACLs - controls aimed at teams that cannot afford a security incident. If you want a self-improving personal assistant, Hermes is a strong choice; if you need isolated, auditable, cost-bounded multi-agent fleets, that is what OpenLegion is built for.
How is OpenLegion different from CrewAI, LangGraph, and AutoGen? Those are primarily libraries/frameworks for orchestrating agent logic inside one process. OpenLegion is a runtime: it runs each agent in its own isolated container, vaults credentials away from agents, enforces budgets and ACLs at a mesh host, and ships browser automation, memory, and multi-channel chat. They are not mutually exclusive - you can run framework-style logic on top of OpenLegion's isolation and cost controls.
How does OpenLegion secure AI agents?
Defense-in-depth: per-agent Docker containers (or Docker Sandbox microVMs), a credential vault that proxies every API call so agents never see keys, per-agent budget caps enforced before each LLM call, default-deny permission ACLs, SSRF protection, path-traversal and prompt-injection sanitization, and an auditable codebase. See the Security Model and docs/security.md.
What LLM providers does OpenLegion support? 100+ providers via LiteLLM (Anthropic, OpenAI, Gemini, Moonshot, DeepSeek, xAI, Groq, Minimax, Zai, Ollama, and more), with health-tracked failover across providers.
Does OpenLegion offer managed hosting? Yes. Managed hosting is available from OpenLegion LLC for teams that prefer not to run their own infrastructure, while self-hosting stays free under the license.
License
OpenLegion is source-available software, copyright © 2026 OpenLegion LLC, licensed under the PolyForm Perimeter License 1.0.1.
You may view, modify, and self-host the software — including for your own commercial, internal use.
You may NOT provide it to others as a product or service that competes with OpenLegion (for example, offering it as a hosted, managed, or SaaS product).
Need a commercial or hosting license? See COMMERCIAL.md or contact admin@openlegion.ai.
"OpenLegion" and the OpenLegion logo are trademarks of OpenLegion LLC; see TRADEMARK.md. See LICENSE for the full terms and CONTRIBUTING.md / CLA.md to contribute.
Related Projects & Comparisons
Looking for alternatives? OpenLegion is often compared to:
- OpenClaw — personal AI assistant, 200K+ stars, not designed for production security
- Hermes Agent — open-source self-improving agent (Nous Research), strong memory and self-authored skills, single-user focused, no container isolation or credential vault
- nanobot — ultra-lightweight Python agent (~4K lines), limited multi-agent support
- ZeroClaw — Rust-based AI agent runtime, extreme resource efficiency, early-stage
- NanoClaw — container-isolated, Claude-only, no cost tracking
- LangChain Agents — feature-rich but complex, heavy framework overhead
- CrewAI — multi-agent framework, no built-in container isolation or cost controls
- AutoGen — Microsoft's multi-agent framework, requires Azure/OpenAI, no self-hosting
OpenLegion differs from all of these in combining fleet orchestration, Docker isolation, credential vaulting, and cost enforcement in a single ~77,000-line auditable codebase.
Keywords: secure AI agent runtime, self-hosted AI agents, AI agent platform, multi-agent framework, autonomous AI agents, OpenClaw alternative, Hermes Agent alternative, OpenClaw vs Hermes Agent, AI agent security, Docker-isolated AI agents, AI agent orchestration, sandboxed AI agents, managed AI agent hosting