Context Autopilot MCP server

by ChiragBachani·io.github.ChiragBachani/context-autopilot·v0.6.0

No reviews yet

Mines your Claude Code and Cursor sessions into evidence-based CLAUDE.md / AGENTS.md rules.

A88/100grade A

Write a review Install

What users say

No reviews yet

Be the first

Safety scan

A88/100

full report

Adoption

New

0 stars48 downloads/wk

Reviews

Write one

Nobody has reviewed Context Autopilot yet.

If you have run it, two minutes of your experience saves the next person an afternoon.

Context Autopilot tools (7, 1 write)

write = sends, deletes, buys or posts

Read from the package source without running it. The installed server may list more.

apply_context_proposals
Apply the user's explicit decisions on pending context proposals. Call ONLY after the user has said in conversation which proposals to accept and/or reject — never decide for them. Pass the exact proposal titles. Accepted rules are written into the managed block of the context files; rejected ones are remembered and never re-proposed; unmentioned proposals stay pending.
distill_context_proposalswrite action
Run the full Context Autopilot pipeline: scan session history, then distill the signals into evidence-backed CLAUDE.md/AGENTS.md proposals. Proposals are saved for review; NOTHING is written to context files until the user approves via apply_context_proposals. Can take a minute. Good moments to proactively suggest this: after the user corrects you repeatedly, or at the natural end of a substantial
distill_global_context
Mine ALL of the user's projects for cross-project signals about how they like agents to work (verification effort, planning style, workflow preferences) and distill them into proposals for their personal global context file (~/.claude/CLAUDE.md). Proposals are saved to ~/.claude/.ctxlayer/proposals.json for review; nothing is applied automatically. Can take a minute.
find_stale_context
Check a project's CLAUDE.md / AGENTS.md for references the repo has outgrown: files that no longer exist and npm scripts that were removed. Returns findings as JSON.
list_observable_projects
List projects that have agent session history (Claude Code or Cursor) available for context mining.
promote_to_global
Scan the memory the user's agents have already saved per project (Claude Code auto-memory directories plus each project's CLAUDE.md/AGENTS.md) and find rules that belong in the user's GLOBAL ~/.claude/CLAUDE.md: how-the-user-works rules, and rules duplicated across 2+ projects. Additive only — project files are never edited. Proposals are saved to ~/.claude/.ctxlayer/proposals.json; NOTHING is app
scan_context_signals
Mine a project's agent-session history for durable context signals: instructions the user repeated across sessions, corrections they made, and tool calls they rejected. Returns the signals as JSON. Proactively call this when the user has corrected you more than once in a session, or when you notice them repeating an instruction they have given before — if strong signals come back, offer to distill

Public scan report

scanner v0.1.9 · 2026-09-26 · same rubric, same numbers if you re-run it

no findings

Code scan29 source files scanned25/25
–Live reliabilityno gateway calls yet and no remote to proben/a
–Tool poisoningtools not inspected (local package is not executed); not countedn/a
Auth qualitylocal package, no credentials required12/15
Maintenancelast push 58 days ago12/15
Maintainer identityregistry namespace matches repository owner; GitHub account older than a year8/10

Overall 88/100. Components that don't apply are left out of the denominator. Any critical finding is an F.Rubric Appeal a finding JSON

How Context Autopilot compares

Against every graded MCP server, from our own scans and uptime checks. Updated after each scan run.

Safety score: 88/100; higher than 75% of 48,255 servers
Tools: 7; the typical server has 7
Tools that change things: 14%; typical: 0%. Fewer write tools means less an agent can break
Downloads a week: 48; more than 40% of 12,191 servers with a package

What its tools let an agent do: 5 read · 1 change · 1 run or deploy

Install Context Autopilot in Claude Code, Cursor or VS Code

Runs npx -y context-autopilot on your machine. Read the scan report first; the gateway never runs local packages.

claude mcp add context-autopilot -- npx -y context-autopilot

Add to Cursor

What the publisher says

From the Context Autopilot repository's README, as published. We do not edit it. Read it on GitHub

Context Autopilot

Automated context collection for coding agents. Mines your real agent sessions — every instruction you repeated, every correction you made, every tool call you rejected — and distills them into CLAUDE.md / AGENTS.md rules you approve.

Part of The Context Layer.

Run in your terminal (it's a plain CLI — don't paste this into a chat):

$ npx context-autopilot scan

Scanned 3 session(s) for ~/projects/my-app
Found 26 signal(s):

  [CORRECTION] ×2 across 2 session(s)  (score 9)
      "There are still so many buttons that dont work, like the publish…"
  [REPEATED] ×4 across 3 session(s)  (score 10)
      "Do not reference the legacy directory. Only work within…"

$ npx context-autopilot distill

[1/8] Perform click-and-type tests before reporting UI work complete  (confidence: high)
    + Before declaring any screen done, click every button and verify it works.
    evidence:
      · 2026-06-27 — "There are still so many buttons that dont work…"

$ npx context-autopilot apply

Why

Every session starts blank, so you re-teach your agent the same conventions — and when you forget, it repeats the same mistakes. Hand-writing context files works but nobody keeps them current. And naive auto-generation is worse: research on LLM-generated context files found they reduce task success and raise cost, because repo scans produce generic filler.

Context Autopilot takes a third path: evidence. Your session history is a literal record of what the agent got wrong and what you said to fix it. Autopilot mines that record and only proposes rules your own words support — each one shipped with the quotes that justify it.

How it works

Observe — ctxlayer scan parses your local Claude Code transcripts (~/.claude/projects) and Cursor sessions and extracts three signal types: instructions repeated across sessions, corrections after the agent went wrong, and rejected tool calls. Runs 100% locally.
Distill — ctxlayer distill sends the signals (not your history) through Claude — via your existing claude CLI, no API key needed — and gets back imperative, project-specific rules with evidence and confidence ratings.
Approve — ctxlayer apply walks you through each proposal. Accepted rules land in a managed block:

<!-- ctxlayer:begin -->
## Learned conventions (Context Autopilot)

- **Staff login cannot access admin view** — When authenticated as staff, the admin role toggle must be hidden or disabled.
<!-- ctxlayer:end -->

Hand-written content is never touched; re-runs update the block idempotently. Rules are written to both CLAUDE.md and AGENTS.md, so Claude Code, Cursor, Copilot, Codex, and every AGENTS.md-aware agent benefits.

Install

npm install -g context-autopilot   # or use npx, no install

Commands

Global mode

ctxlayer distill --global

Shortened. The full README is on GitHub.

Nothing above is checked by us. What we check is on the safety report.

Context Autopilot: common questions

Is Context Autopilot MCP server safe?: Yes, by our scan: it is graded A (88/100). Read the Context Autopilot safety report
How do I install Context Autopilot?: It runs on your machine. Copy the Claude Code, Cursor, VS Code or Claude Desktop config from the install section.
Does Context Autopilot need an API key?: Not as far as the registry entry and our scan can tell: no credentials are declared or required.
Is Context Autopilot maintained?: The last commit was 97 days ago (2026-07-30). The latest release is v0.6.0.
How does Context Autopilot compare with other MCP servers?: safety score 88/100 (higher than 75% of 48,255 servers); tools 7 (the typical server has 7); tools that change things 14% (typical: 0%. Fewer write tools means less an agent can break). Full comparison.