mcpField
C
ContextLib

ContextLib

Local
huggingface.coMemory & Knowledge1Updated 5w ago

About ContextLib

ContextLib stores project decisions, requirements, facts, and lessons as Markdown files with metadata for people and agents to share across sessions. Files remain the source of truth while search indexes can be rebuilt; its CLI and MCP server support recording, retrieval, and review.

Key features

  1. 01Stores project knowledge as Markdown files
  2. 02Searches records with a rebuildable index
  3. 03Manages proposed, reviewed, and retired records
  4. 04Supports import, export, and health checks
  5. 05Provides a CLI and MCP tools

Use cases

  1. 01Carry project decisions across agent sessions
  2. 02Review and accept lessons proposed by agents
  3. 03Move or export project records
  4. 04Search and update project records through MCP

MCP support

It provides a stdio MCP server with tools for project briefs, search, retrieval, recording, review, import, and export. The server uses an identity fixed at startup, and tools cannot write as another identity.

Configuration

Local server config is the same for every client. Copy it into your client’s MCP config file.

mcp.json · Universal config
{
"mcpServers": {
"willykeenan-agentbrain-contextlib": {
"command": "docker",
"args": [
"run",
"-i",
"--rm",
"ghcr.io/willykeenan/agentbrain-contextlib:0.1.0"
]
}
}
}
Run commanddocker run -i --rm ghcr.io/willykeenan/agentbrain-contextlib:0.1.0

Local servers need a runtime on your machine (e.g. Node.js 18+). Config file locations differ by client; see your client’s docs.

Packages
ocighcr.io/willykeenan/agentbrain-contextlib:0.1.0

Tools

No tool list yet

Local servers expose tools only once running on your machine, and the maintainer has not provided a tool list.

Details

From the project README

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).

A ctxlib session: an owner's decision lands current, an agent's lesson waits in the inbox, a secret is refused, and doctor passes
A ctxlib session: an owner's decision lands current, an agent's lesson waits in the inbox, a secret is refused, and doctor passes.

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.

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
CommandWhat 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 listList projects
add PROJECT TYPE TITLE (--body TEXT | --body-file F) --author ID [options]Write a record (--evidence, --tag, --review-by, --approved-by)
get PROJECT IDPrint one record, with evidence state and supersession chain
search PROJECT QUERY [--type] [--all]Search the rebuildable index
brief PROJECTPrint the always-loaded summary (≤ 8192 bytes)
supersede PROJECT OLD NEW --reason TEXT --author IDMark old superseded, move the file, keep the body hash
review PROJECT ID confirm|retire --note TEXT --author IDConfirm or retire a record
inbox PROJECTList proposed records
accept PROJECT ID --author IDMove inbox → current
reject PROJECT ID --reason TEXT --author IDMove inbox → rejected
due PROJECTRecords whose review_by is due
regen PROJECT|--allRewrite BRIEF, INDEX, GLOSSARY, TIMELINE, README
rebuildRebuild the SQLite search index from files
capture PROJECT RESULT.jsonTurn a result file into a return (and proposed lessons)
import-md PROJECT FILE --type TYPE --author IDMarkdown file → record
export PROJECT OUT_DIR [--include-private]Zip a project
import BUNDLEImport a zip
doctor [PROJECT] [--ten-minute]Plain-English health check
syncgit pull / git push when the library is a git repo; otherwise a no-op with a message
mcp --identity IDStdio 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

Comments

Markdown supported

New to MCP?

From what the protocol is to installing your first server, these answers cover most first questions.

An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.