capsulectx

Capsule

Compact, source-linked context packs for coding agents.

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.

Quickstart

npx capsulectx init

Capsule scans your repo 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 and create all .capsules/*.md files
capsule scan	Print detected source groups without writing
capsule write <name>	Refresh one capsule (preserves your edits)
capsule get <name>	Print one capsule to stdout
capsule stale [name]	Check which source files changed since last write
capsule estimate <name>	Show estimated token savings for one capsule

Seeing the savings

$ capsule estimate architecture

Capsule: architecture

Without Capsule:
  files: 5
  estimated tokens: 9,607

With Capsule:
  capsule plus stale files: 1
  stale source files: 0
  estimated tokens: 417

Estimated discovery savings: 96%

The estimator is directional — it uses 4 characters per token as a conservative approximation.

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.

Building the Demo

The animation in this README is a generated GIF — no screen-recording tools required.

npm run demo

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.

Status

Capsule is early — v0, local-only, no paid LLM calls.

In v0.1:

• local CLI, language-agnostic repo scanning
• Markdown context packs with source fingerprints
• stale detection, token savings estimation
• write-safe re-generation (human edits preserved)
• custom group config via .capsules/config.json

Roadmap:

• capsule prompt claude|codex|cursor — inject capsule into agent context directly
• smarter write that diffs source changes and suggests updates
• better source grouping for monorepos
• before/after token reports for real agent sessions
• GitHub Action for stale capsule checks in CI
• public capsule format spec

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