Skip to main content
Glama
ContextLib MCP Server

ContextLib MCP Server

by willykeenan

ContextLib

A project library that outlives every agent session.

The sample library rendered read-only: project brief, current decisions, open requirements and key facts

Real screenshot of the sample library (an invented project).

Each project's knowledge — decisions, requirements, facts, lessons, glossary, sources and returned work — is stored as plain Markdown files with a small front-matter block, in one folder per project. The folders can sit on an external SSD. A person can understand a project from Finder alone. Agents read and write the same files through a CLI (ctxlib) and an MCP server (context_* tools). Any database is a search index rebuilt from the files, never the source of truth.

It is the open core of the context plugin of AgentBrain.

The Python package is agentbrain_contextlib. The CLI is ctxlib. Python 3.9+, standard library only. Git is optional.

Why

Agent sessions end. Chat logs drift. A store that only a model can parse is a dead end the first time you open Finder on an airplane.

ContextLib keeps the durable layer as files you can read, diff, copy, and take to another Mac. The search index is disposable. The Markdown is the product.

Related MCP server: Knowledge MCP Server

60-second quickstart

pip install "git+https://github.com/willykeenan/agentbrain-contextlib"

ctxlib init ~/ContextLib --project lighthouse-app \
  --name "Lighthouse App" \
  --purpose "Harbor dashboard for lamp status, fog signals, and traffic"

ctxlib add lighthouse-app decision "Use keeper logs as lamp status" \
  --body "**Decision:** Lamp status comes from keeper logs, not AIS." \
  --author mara \
  --tag lamp --tag keeper-log

ctxlib brief lighthouse-app
ctxlib doctor lighthouse-app --ten-minute

The library path is --library PATH, else env CONTEXTLIB_ROOT, else ~/ContextLib.

A finished sample you can open in Finder without installing anything lives at examples/sample-library/. Start at that README, then the project BRIEF, then decisions/INDEX.md.

Full walkthrough: docs/QUICKSTART.md.

CLI (ctxlib)

ctxlib [--library PATH] COMMAND

Command

What it does

init PATH [--project SLUG --name NAME --purpose TEXT]

Create a library (and optionally the first project)

project add SLUG NAME [--purpose]

Add a project folder

project list

List projects

add PROJECT TYPE TITLE (--body TEXT | --body-file F) --author ID [options]

Write a record (--evidence, --tag, --review-by, --approved-by)

get PROJECT ID

Print one record, with evidence state and supersession chain

search PROJECT QUERY [--type] [--all]

Search the rebuildable index

brief PROJECT

Print the always-loaded summary (≤ 8192 bytes)

supersede PROJECT OLD NEW --reason TEXT --author ID

Mark old superseded, move the file, keep the body hash

review PROJECT ID confirm|retire --note TEXT --author ID

Confirm or retire a record

inbox PROJECT

List proposed records

accept PROJECT ID --author ID

Move inbox → current

reject PROJECT ID --reason TEXT --author ID

Move inbox → rejected

due PROJECT

Records whose review_by is due

regen PROJECT|--all

Rewrite BRIEF, INDEX, GLOSSARY, TIMELINE, README

rebuild

Rebuild the SQLite search index from files

capture PROJECT RESULT.json

Turn a result file into a return (and proposed lessons)

import-md PROJECT FILE --type TYPE --author ID

Markdown file → record

export PROJECT OUT_DIR [--include-private]

Zip a project

import BUNDLE

Import a zip

doctor [PROJECT] [--ten-minute]

Plain-English health check

sync

git pull / git push when the library is a git repo; otherwise a no-op with a message

mcp --identity ID

Stdio MCP server; the identity is fixed for the process

Types: decision, requirement, fact, lesson, glossary, source, return, brief-note.

MCP (context_* tools)

ctxlib mcp --identity ID is a stdio JSON-RPC server. Protocol versions: 2024-11-05, 2025-03-26, 2025-06-18.

Tools: context_brief, context_search, context_get, context_record, context_supersede, context_review_due, context_review, context_capture, context_export, context_import, context_status.

The identity is fixed at start. Tools cannot write as another identity. Errors come back as isError results.

Claude Code

{
  "mcpServers": {
    "contextlib": {
      "command": "ctxlib",
      "args": ["mcp", "--identity", "claude:project"],
      "env": { "CONTEXTLIB_ROOT": "/Volumes/SSD/ContextLib" }
    }
  }
}

Codex

{
  "mcpServers": {
    "contextlib": {
      "command": "ctxlib",
      "args": ["mcp", "--identity", "codex:project"],
      "env": { "CONTEXTLIB_ROOT": "/Volumes/SSD/ContextLib" }
    }
  }
}

Cursor

In .cursor/mcp.json (project) or ~/.cursor/mcp.json (user):

{
  "mcpServers": {
    "contextlib": {
      "command": "ctxlib",
      "args": ["mcp", "--identity", "cursor:project"],
      "env": { "CONTEXTLIB_ROOT": "/Volumes/SSD/ContextLib" }
    }
  }
}

Details: docs/MCP.md.

Record format

A record is UTF-8 Markdown. It starts with a YAML-like front-matter block (scalars, quoted strings, lists of scalars, lists of flat maps) and a Markdown body. The body hash is stored in the front-matter; bodies are not rewritten.

---
id: dec-20260926-7f3a
type: decision
title: Use plain files as the source of truth
status: current
project: lighthouse-app
scope: {lane: workstream, team: owner}
created: 2026-09-26T14:03:00Z
author: mara
approved_by: mara — "files first, database is only an index"
evidence:
  - {ref: "project:BRIEF.md"}
  - {ref: "url:https://example.com/x"}
supersedes: []
superseded_by: null
review_by: 2027-03-25
tags: [storage, library]
sensitivity: normal
body_sha256: "…"
---
**Decision:** …
**Why:** …
**Consequences:** …
**How to verify:** …

Evidence refs are project:<relative path>, ssd:<relative path>, room:<name>#<seq>, url:<https url>, or note:<text>. Absolute home paths and secrets are refused at write time.

Ids: <prefix>-<YYYYMMDD>-<4 hex> with prefixes dec, req, fact, les, glo, src, ret, note.

Files: <YYYY-MM-DD>_<title-slug>_<id-suffix>.md. Superseded records move to superseded/ with the body hash still valid.

Library layout

<library root>/
├── README.md              how to read this in 10 minutes
├── INDEX.md               one line per project
├── library.json           {"format": "contextlib/1", "library_id", "created"}
└── projects/<slug>/
    ├── README.md BRIEF.md GLOSSARY.md TIMELINE.md project.json
    ├── decisions/ requirements/ facts/ lessons/ glossary/ sources/ returns/
    │   ├── INDEX.md
    │   ├── <YYYY-MM-DD>_<title-slug>_<id-suffix>.md
    │   └── superseded/
    ├── inbox/             proposed records
    ├── exports/
    └── .contextlib/       ledger + rebuildable index.sqlite

Keeping the library on an SSD, using it unplugged, and moving it between Macs: docs/SSD.md.

Develop

PYTHONPATH=src python3 -m unittest discover -s tests
python3 tools/check.py packaging

Python 3.9+, standard library only. See CONTRIBUTING.md and CHANGELOG.md.

Apache-2.0, copyright KE Studios. SECURITY.md for reports.

ContextLib is the open library layer of AgentBrain (agentrooms.io).

Run with Docker

The image is published at ghcr.io/willykeenan/agentbrain-contextlib for Apple silicon and Intel. Mount your library at /library:

claude mcp add contextlib -- docker run -i --rm --user "$(id -u):$(id -g)" -v ~/ContextLib:/library ghcr.io/willykeenan/agentbrain-contextlib mcp --identity you

Maintenance

ActivityNo data
ResponsivenessSyncing

Related MCP Connectors

Related MCP Servers

  • A
    license
    Not graded
    quality
    A
    maintenance
    Enables MCP agents to maintain durable, evidence-aware project knowledge, retrieve precise excerpts on demand, and track decisions, conflicts, and revisions across sessions.
    1
    Apache 2.0
  • A
    license
    Not graded
    quality
    B
    maintenance
    Enables AI agents to search, read, create, and maintain a local-first knowledge base using hybrid retrieval and Markdown note management through the Model Context Protocol.
    11 npm
    1
    MIT
  • A
    license
    A
    quality
    B
    maintenance
    Enables AI agents to save and read persistent, project-local markdown context files in a .opencontext/ directory, so durable project decisions and conventions survive across sessions.
    2
    8
    MIT
  • A
    license
    A
    quality
    B
    maintenance
    Enables AI agents and MCP clients to start each session with a pre-assembled context pack of the project's Markdown notes, decisions, rules, Git and tracker state, and to write progress and decisions back as dated logs. Context is gathered locally by lookups and queries rather than model calls, so different tools like Claude Code, Codex, and Antigravity share the same memory without spending tokens on retrieval.
    15
    1
    Apache 2.0