MCP gives agents tools. Skills give agents procedures. Capsules give agents compact project context.
Large repositories make coding agents burn tokens rediscovering the same files, decisions, conventions, and setup details over and over. Capsule targets a 50–70% reduction in repeated repository-discovery context for large, multi-session agent work.
Install
No install required — run directly with npx (always gets the latest version):
Or install globally with your package manager:
npm install -g capsulectx
pnpm add -g capsulectx
yarn global add capsulectx
bun add -g capsulectx
Keep it up to date:
npm install -g capsulectx --prefer-online
Tip: npx is the easiest option — it always runs the latest published version with no manual update step.
Quickstart
Capsule scans your repo, optionally enriches capsules with AI (using your installed Claude CLI, API keys, or a generated prompt for any browser-based AI), and creates:
.capsules/
index.md
architecture.md
setup.md
api.md
data.md
ui.md
testing.md
deployment.md
Each capsule is a Markdown file with source fingerprints in JSON frontmatter:
---json
{
"name": "api",
"sources": ["src/api/**"],
"fingerprints": { "src/api/users.ts": "sha256:..." },
"updated_at": "2026-06-20T14:00:00.000Z"
}
---
# API Capsule
## Purpose
Request handlers, routes, controllers, and service boundaries.
## Key Files
- `src/api/users.ts`: source file matched by this capsule.
## Decisions
- Record stable decisions here so future agents do not rediscover them.
Agents read .capsules/index.md, choose the relevant capsule, check staleness, then inspect only the source files that actually matter.
Commands
| Command | What it does |
|---|
capsule init | Scan repo, run AI enrichment (optional), write all capsules |
capsule enrich | Re-run AI enrichment on existing capsules |
capsule apply | Apply AI-generated JSON to capsule Conventions and Decisions |
capsule write <name> | Refresh one capsule from source (preserves your edits) |
capsule refresh <name> | Explicitly refresh one capsule |
capsule refresh --all | Explicitly refresh all detected capsules |
capsule get <name> | Print one capsule to stdout |
capsule scan | Print detected source groups without writing |
capsule stale [name] | Check which source files changed since last write |
capsule estimate <name> | Show estimated token savings for one capsule |
capsule stats | Show token savings across all capsules |
capsule report [name] | Show reproducible source/capsule/stale token savings |
capsule report --write | Write .capsules/report.md as a proof artifact |
capsule check | CI-friendly freshness/config check; exits nonzero on stale or missing capsules |
capsule --mcp | Start MCP server for native agent integration |
MCP Server
Capsule ships an MCP server so Claude Code and other MCP-compatible agents can natively discover and consume capsules — no manual file reads needed.
Add to your project's .claude/settings.json:
{
"mcpServers": {
"capsule": {
"command": "npx",
"args": ["capsulectx", "--mcp"]
}
}
}
Or globally in ~/.claude/settings.json to enable for all projects.
Once configured, agents get these tools automatically:
| Tool | What it does |
|---|
capsule_list | List all capsules with freshness status |
capsule_get | Read a capsule's compressed context; warns when stale but does not write |
capsule_refresh | Explicitly refresh one capsule or all capsules |
capsule_stale | Check which capsules have outdated sources |
capsule_stats | Show token savings per capsule |
capsule_report | Show machine-readable discovery-savings proof |
capsule_check | CI-style stale/missing/config check |
capsule_init | Generate capsules for the current repo |
Capsules are also available as MCP resources at capsule:///index and capsule:///{name}.
Seeing the savings
$ capsule report architecture
Capsule: architecture
capsule status source capsule stale saved
architecture fresh 9,607 417 0 96%
Token counts use the cl100k_base tokenizer for accurate measurement.
For CI, run:
It exits nonzero when capsules are stale, missing, or blocked by invalid config.
Agent Workflow
Add this to AGENTS.md, CLAUDE.md, Cursor rules, or your agent instructions:
Before working in this repo:
1. Read `.capsules/index.md` if it exists.
2. Read the capsule matching the task area.
3. Run `capsule stale <name>` when the CLI is available.
4. If stale, inspect changed source files before editing.
5. Update capsules when durable project knowledge changes.
Works with any agent that can read files — Claude Code, Codex, Cursor, Windsurf, Devin, custom MCP agents.
Custom Groups
Capsule detects standard layouts out of the box. For non-standard structure, add .capsules/config.json:
{
"groups": [
{
"name": "workers",
"description": "Background jobs, queues, and scheduled tasks.",
"sources": ["src/workers/**", "src/jobs/**"]
}
],
"ignore": ["legacy/**"]
}
A groups entry with the same name as a default group overrides it. New names are appended. ignore patterns extend the default ignore list.
See docs/capsule-format.md for the public capsule file format.
Building the Demo
The animation in this README is a generated GIF — no screen-recording tools required.
Requires rsvg-convert (librsvg / librsvg2-tools) and ffmpeg. The script outputs docs/demo.gif from SVG frames rendered at 2× for retina. See scripts/build-demo.mjs.
Documentation Site
The OSS docs live in docs-site/ and are built with VitePress.
npm run docs:dev
npm run docs:build
Status
Capsule is early but functional. Current capabilities:
- Language-agnostic repo scanning with monorepo support (Turborepo, pnpm workspaces)
- Static analysis: framework detection, table/model extraction, route detection, env vars
- AI enrichment on init — launches your installed Claude CLI, calls Claude/OpenAI API, or generates a prompt for any browser-based AI (Claude.ai, ChatGPT, Cursor, Windsurf)
- Write-safe re-generation: human edits to Conventions and Decisions are preserved on refresh
- MCP server for native Claude Code / agent integration
- Real token counting (cl100k_base) with savings dashboard
- TypeScript API surface extraction
- Binary file filtering
- Stale detection and token savings estimation
- Reproducible
report and CI-friendly check
- Custom group config via
.capsules/config.json
- Auto-patches
CLAUDE.md, AGENTS.md, .cursorrules, .windsurfrules with capsule instructions
Roadmap:
- smarter
write that diffs source changes and suggests targeted updates
- before/after token reports for real agent sessions
- GitHub Action wrapper for
capsule check
Install for Local Development
git clone https://github.com/dawitlabs/capsule.git
cd capsule
npm install
npm run build
node dist/cli.js init
License
MIT
