GoodVibes

Plug in. Receive good vibes.

One Claude Code plugin, three MCP servers, 25 tools: structure-aware code intelligence, session cost analytics, and an API workbench — opt-in capabilities the native harness doesn't have, and honest about the many things it already does well.

What it is

Server	Tools	What it does
intel	15	Structure-first code reading (code_read outline/ranges), multi-query search (code_grep with count/locations formats, AST structural queries, expand_to, preview_replace), code_glob with per-file stats, TypeScript API-surface and safe-delete analysis on a single compiler host, API route/spec/contract analyzers, DB schema intelligence with Prisma usage analysis, React component/hook/boundary analyzers, a merged layout analyzer, project scaffolding — and structural_edit, the one write tool, preview-gated
analytics	7	Per-session and cross-project token/cost analytics from transcript actuals with per-model cache-aware pricing, budgets, tags, export, and a self-contained HTML report (dashboard {"action":"report"} writes it to .goodvibes/reports/) — plus live modes: query {"mode":"live_cost"} prices the running session, "doctor" surfaces host load and orphaned processes with ready-to-run cleanup commands (never auto-runs them), "agents" classifies background agents as thinking / executing / wedged
connect	3	api_request (batched HTTP with per-entry error isolation), service (a credential registry: 0600 secrets file, env-var indirection, bounded OAuth2 refresh), db_query (live databases under the same trust model)

Measured, not promised

Every performance claim in this README was measured against the native tool baseline at default settings, with ground-truth verification — the benchmark harness ships in packages/intel/bench/ so you can rerun it:

• code_read outline: −40% to −72% of the token cost of a full native read, sufficient to answer structure questions with zero follow-up reads
• code_grep files mode: −63% vs native search with identical match counts (76 = 76)
• Responses self-report their cost (token_estimate computed from the actual rendered payload), enforce output.max_tokens, and set truncated/effective_caps only when something was actually trimmed

If a claim stops being true, it comes out of the README. That's the policy.

Design principles

• Native tools are never blocked, redirected, or deprecated. These are additions, not replacements — use whichever tool wins for the job.
• Hooks observe, inform, and assist. Zero hooks block, rewrite, or steer. All are fail-open: a broken hook can never brick a tool call.
• Servers live for the life of your session. No idle self-exit, ever (regression-tested). Orphan defense is a parent-liveness watchdog that fires only when the session itself dies.
• Every tool runs under a per-request time budget and returns an honest partial rather than hanging a client forever.
• connect is restricted by default: credentials are pinned to their registered origins (never toggleable), destinations are allowlisted, writes are per-service opt-in, and "open" mode is a human-only, ephemeral toggle that announces itself every session it persists.
• State is namespaced under .goodvibes/ in your project; outside it the plugin writes only its own dependency home under ~/.claude/.goodvibes/ (/goodvibes:plugin install-prompts remains opt-in and has a real uninstall).

Install

claude plugin marketplace add mgd34msu/goodvibes-plugin
claude plugin install goodvibes@goodvibes-market

Restart Claude Code to activate the servers. Tool schemas load on demand when your client has Tool Search active (the default in current Claude Code); the fixed context cost is the tool names only.

Native dependencies install automatically in the background: the first session after install spawns a detached installer and the tools are ready next session. Installs live in ~/.claude/.goodvibes/deps/ and survive plugin updates — the SessionStart hook relinks them automatically. If the background install fails, one line points at the log and /goodvibes:setup, the manual foreground repair path. Until an install lands, the servers still boot and every non-native capability works; anything that needs a native dependency returns an honest "run /goodvibes:setup" message instead of failing.

Commands

• /goodvibes:plugin — status, native-dependency repair, prompt-pointer install/uninstall
• /goodvibes:setup — re-run the native dependency install in the foreground (repair path)
• /goodvibes:analytics — session cost, HTML report, budgets, doctor, agent liveness
• /goodvibes:services — connect's registry and the restricted/open toggle
• /goodvibes:codebase-review — an in-band review→fix workflow with refutation-based verdicts and grounded checks

History

GoodVibes v1 was a 77-tool replacement harness built in an earlier era of Claude Code. A measured deep review found its best ideas had been absorbed by the platform and its policy layer cost more than it saved — so v2 kept the verified winners, fixed every field defect (all nine are regression-tested), and retired the rest. The complete v1 lives forever on the v1 branch; the full investigative record — the deep review, the field-issue reports, the plan — lives in docs/history/. The living architecture record is docs/goodvibes-plan.md and docs/carveout-architecture.md.

Versioning

Semantic versioning from 2.0.0: patches for fixes, minors for features, majors for breaking changes.

License

MIT