Project knowledge base

Everything in Touchgrass

Architecture, setup, tools, APIs, data, security, operations, decisions, status, and roadmap.

Documents20Curated and generated
Generated7Built from source
MCP tools84Catalogued automatically
HTTP routes158Discovered automatically
Start hereSource-generated

Documentation home

System overview, topology, implemented capabilities, and recommended reading order.

Source Generated from current repository sourcesWords 295Reading 2 minSections 5

Touchgrass documentation#

Touchgrass is a self-hosted autonomous developer-agent execution and observability platform. ChatGPT or another MCP client is the conversational control surface; Touchgrass supplies isolated workspaces, tools, durable state, browser and terminal execution, checkpoints, evidence, and an operator dashboard.

Current project at a glance#

AreaCurrent implementation
Applications4 independently built services
Shared packages4 workspace packages
MCP tools84 source-discovered tools
HTTP routes158 source-discovered Fastify routes
Durable tables12 source-discovered PostgreSQL tables
RuntimeNode.js 22+, pnpm 11, Docker, PostgreSQL 17, Nuxt 5, Fastify 5
Safety modelReversible isolated execution; external and irreversible effects denied by default

System topology#

text
MCP client / ChatGPT
        |
        v
MCP server :4311
        |
        v
Control API :4310 ---- PostgreSQL :55432
        |                    |
        |                    +-- workspaces, tasks, processes, leases, events
        v
Runner :4312
  |       |        |
  |       |        +-- isolated Chromium / Playwright
  |       +----------- detached Docker PTY processes and previews
  +------------------- isolated workspaces, files, Git, shell, checkpoints

Nuxt dashboard :3000 ---- same-origin control and preview proxy

What is implemented#

  • Isolated Git worktrees or filtered snapshot workspaces.
  • Repository analysis and command detection.
  • Durable dependency-aware task plans.
  • Bounded file reads, search, exact edits, writes, and atomic multi-file patches.
  • Git status and bounded diffs.
  • Automatic validation loops with side-effect restoration.
  • One-shot Docker shell execution with internet access and resource limits.
  • Detached PTY processes with retained output, stdin, resize, cancellation, port detection, restart recovery, and exclusive expiring control leases.
  • Isolated HTTP/HTTPS process previews without publishing Docker ports.
  • Playwright browser contexts, accessibility snapshots, refs, actions, screenshots, console/network telemetry, and adaptive frame streaming.
  • Content-addressed browser artifacts and immutable retrieval.
  • Ordered durable events, session projections, usage/cost reporting, and analytics.
  • Responsive task, process, browser, analytics, login, and documentation interfaces.
  1. Easy start — install and run the complete stack.
  2. Current implementation status — understand what is actually complete today.
  3. Architecture and data model — understand service boundaries and persistence.
  4. MCP tool reference — see the client-facing capability surface.
  5. Security model — understand isolation, effects, leases, and remaining launch blockers.
  6. Detailed implementation plan — see the full roadmap and ticket structure.