flevinsky-ai

Cross-agent dev service runner. Declare your project's long-running services once and run, observe, and tail logs from any supported AI coding agent — without losing state when the agent restarts.

What it does

• One config, many agents. Declare services (frontend, api, workers…) in a single JSON file.
• Persistent state. Services keep running across agent sessions. State + logs land in <root>/<prefix>/services/.
• Per-service logs. Tail or grep them from inside the agent instead of re-running commands.
• Fallback prefix chain. Picks the first of .flevinsky-ai, .services, .pi, .claude that exists — so existing pi-services or Claude Code state keeps working.

Agent compatibility

Agent	Mechanism	Status
Claude Code	Plugin (MCP + hooks)	✅
Pi	Extension	✅
CLI	flevinsky-ai runner	✅
Other MCP	bin/server.js	✅

Install

macOS / Linux (one-liner)

curl -sL https://raw.githubusercontent.com/FabianLevi/flevinsky-ai/main/scripts/install.sh | bash

Auto-detects Homebrew → pnpm → npm. Force a method with FL_METHOD=brew / FL_METHOD=pnpm / FL_METHOD=npm.

Homebrew

brew install FabianLevi/tap/flevinsky-ai

pnpm

pnpm add -g flevinsky-ai

npm

npm install -g flevinsky-ai

Windows

Use pnpm or npm inside WSL2, PowerShell, or Git Bash. Native Windows is best-effort.

Per-agent install

• Claude Code — /plugin marketplace add FabianLevi/flevinsky-ai then /plugin install flevinsky-ai
• Pi — installs as a Pi extension automatically when the package is on $PATH

Declare services

Create .flevinsky-ai/services.json at your project root:

{
  "services": {
    "frontend": {
      "kind": "server",
      "cmd": "pnpm dev",
      "cwd": "frontend",
      "readyPattern": "Local:\\s+http"
    },
    "api": {
      "kind": "server",
      "cmd": "uvicorn app:app --reload",
      "cwd": "backend"
    },
    "migrate": {
      "kind": "task",
      "cmd": "pnpm db:migrate"
    }
  }
}

Field reference: see docs/architecture.md.

Use it

From any terminal

flevinsky-ai list                          # show declared + live services
flevinsky-ai start frontend                # background-spawn
flevinsky-ai logs   frontend 200           # tail last 200 lines
flevinsky-ai stop   frontend
flevinsky-ai run    frontend -- pnpm dev   # attached: stream output to terminal, track state

From inside Claude Code

The plugin exposes MCP tools services_list / start / stop / restart / logs. Just ask: "tail frontend logs", "restart api".

From inside Pi

Same surface via the Pi extension — /services list, /services logs frontend, etc.

Docs

• Architecture — state layout, process model, plugin/extension wiring
• Release checklist — how a v* tag becomes a release
• Rollback — undoing a bad release
• AGENTS.md — skills, MCP tools, hooks index for AI agents

License

MIT