Claude Agent Runner App

Small app that combines:

• the Claude Agent SDK for orchestration
• the local AutonomyX Model Runner as a callable tool
• a packaged macOS desktop shell that can be built into a .dmg

What it does

The backend starts a Claude agent request and exposes the selected AutonomyX runner model as an in-process MCP tool. Claude can call that tool during the turn, inspect the local-model output, and then produce the final answer.

The frontend lets you:

• choose a Claude model
• refresh and choose runner models from RUNNER_BASE_URL
• load the selected runner model
• edit the model selection config from the UI
• cancel a running request and switch to another model immediately
• choose from 9 memory types for the agent run
• inject dynamic context per run without changing saved config
• attach a per-run decision log for explicit prior decisions and rationale
• attach evaluation criteria for quality and acceptance checks
• attach per-run feedback and guardrails
• attach prompt registry, skill registry, model registry, and LangGraph components
• attach decide components
• emit optional Langfuse traces for each agent run
• submit a prompt to the Claude agent

Files

• server.js: Express app and Claude Agent SDK integration
• public/: static frontend
• .env.example: required environment variables
• config.json: runtime-persisted UI config for model selection and runner URL

The app exposes 9 built-in memory types: working, episodic, semantic, procedural, preference, project, task, decision, and artifact.

The app treats AutonomyX Model Runner as the runtime source of truth. The built-in defaults are limited to the local AutonomyX runner models.

Run In Dev

1. Start the AutonomyX Model Runner:

   cd /Users/apple/organization/products/model-runner/repo
   ./run.sh

2. Install app dependencies:

   cd /Users/apple/organization/products/claude-agent-runner-app/repo
   npm install

3. Create .env:

   cp .env.example .env

4. Set at least:

   ANTHROPIC_API_KEY=...
   RUNNER_BASE_URL=http://127.0.0.1:8765

   Optional Langfuse tracing:

   LANGFUSE_PUBLIC_KEY=...
   LANGFUSE_SECRET_KEY=...
   LANGFUSE_BASE_URL=https://cloud.langfuse.com

5. Start the app:

   npm run dev

This launches the Electron desktop app and starts the local Express server inside it.

If you only want the browser version:

npm run dev:web

Build DMG

cd /Users/apple/organization/products/claude-agent-runner-app/repo
npm install
npm run dist:dmg

The generated .dmg will be written to dist/.

Docker

Build the production image:

cd /Users/apple/organization/products/claude-agent-runner-app/repo
docker build -t your-dockerhub-user/claude-agent-runner-app:latest .

Run it:

docker run --rm -p 3000:3000 \
  -e ANTHROPIC_API_KEY=... \
  -e RUNNER_BASE_URL=http://host.docker.internal:8765 \
  your-dockerhub-user/claude-agent-runner-app:latest

Notes

• The runner model must be loaded before the Claude agent can use it.
• The app defaults to a small list of Claude models. Change CLAUDE_MODELS in .env if you want a different set.
• The integration keeps built-in Claude tools disabled and only auto-approves the two MCP tools defined in server.js.
• Langfuse tracing is only enabled when both Langfuse keys are present.
• The DMG build is unsigned by default. macOS may warn before first launch unless you sign and notarize it.