Context

Context is a documentation cache system for coding agents. It maintains project knowledge as markdown files with automatic invalidation when source files change. It can be used as a CLI tool or MCP server.

Install

brew install rchowell/tap/context

Usage

Documentation is stored in .context/ as markdown files. You (and agents) can view and edit with any tools. You can use context sync to mark documentation as fresh, and you can use context status to get a list of stale documents. This information can be fed to agents (and you) to determine which documentation should be updated.

# Initialize .context/ within the project root
context init

Claude Code

# Adds a 'context' MCP server with stdio
claude mcp add --transport stdio --scope project context context serve

# Add the skill (optional, in-progress)
.claude/skills/using-documentation/SKILL.md

via @

# It's just markdown
You can do @.context/guides/doc.md

via MCP

# MCP via stdio protocol
context serve

via CLI

# You can use it manually or within a bash tool call
context --help

Command	Purpose
context init [dir]	Scaffold directory structure
context status	Report valid/stale/orphaned docs
context sync [path]	Update hashes, mark as reviewed
context find [path]	Find all references to the given path

How It Works

Each project gets a root .context directory somewhat like the .git directory.

context init  # initialiazes .context/

Author markdown documentation within ./context and reference other files in backticks. Here is an example doc:

---
slug: auth
description: Authentication and JWT handling
---

The auth system lives in `src/auth/mod.rs` and `src/auth/jwt.rs`.

Then run context sync to generate all reference hashes in the frontmatter.

references:
  src/auth/mod.rs: 8a3b2c1
  src/auth/jwt.rs: f4e5d6a

As you write code, the documentation can become stale (bad) — you can find invalidated documents by doing context status, ex:

context status

modified:  .context/guides/auth.md

Directory Structure

The idea here is that an index is like a layered cache, and agents should try to read index > guides > references.

context/
├── index.md                # Project overview
├── guides/
│   ├── index.md            # Guide listing
│   └── {topic}.md          # Conceptual docs
└── references/
    ├── index.md            # Reference listing  
    └── {topic}.md          # Technical docs

• Guides: explain concepts, workflows, and architecture. They have broad references and invalidate slowly.
• References: document specific modules and components. They have narrow references and invalidate quickly.
• Index files: aggregate references from their children, invalidating when any child document's dependencies change.

Release

I will be manually releasing prebuilt Apple Silicon binaries for early versions (~1MB).

# Define the version and tag
export VERSION="0.1.0"
export TAG="v$VERSION"

# Build a release
cargo build --release --target aarch64-apple-darwin

# Bundle as a tar.gz
tar -czvf context-$VERSION-aarch64-apple-darwin.tar.gz -C target/aarch64-apple-darwin/release context

# Upload the release
gh release create $TAG context-$VERSION-aarch64-apple-darwin.tar.gz

# Compute the new sha245
shasum -a 256 context-$VERSION-*.tar.gz