Mobile Code Agent

A mobile-friendly coding agent with a VS Code-like web UI, Monaco editor, indexed code search, and a tool-calling backend for a local workspace. Open it on your phone over the network, lie back, and let AI inspect, edit, and write code for you.

中文说明

What is included

• React web UI with Monaco editor
• Express API server
• coding-agent loop with tool calls
• local session persistence
• workspace-safe file tools
• local code index with:
  • heuristic semantic chunking
  • BM25-style keyword retrieval
  • optional embedding-based vector retrieval
  • hybrid RRF fusion
  • filesystem watch and auto-rebuild
• CLI entrypoints for one-shot tasks, REPL, index rebuild, and search

Model setup

This project is set up for MiniMax M2.5 by default.

cp .env.example .env

Minimum chat config:

MINIMAX_API_KEY=your_minimax_api_key
CHAT_BASE_URL=https://api.minimaxi.com/v1
CHAT_MODEL=MiniMax-M2.5

Optional embedding config for vector retrieval:

EMBEDDING_API_KEY=your_embedding_key
EMBEDDING_BASE_URL=https://your-openai-compatible-embedding-endpoint/v1
EMBEDDING_MODEL=your-embedding-model

If embedding config is missing, the index still works with keyword retrieval only.

Run the app

Install dependencies:

npm install

Start the backend:

npm run dev

Start the frontend:

npm run dev:web

Production build:

npm run build
npm start

The server defaults to http://127.0.0.1:3000.

To expose it on your LAN or a public server, set:

CODE_AGENT_HOST=0.0.0.0
CODE_AGENT_PORT=3000

Then open:

• http://<your-server-ip>:3000

For the Vite dev frontend, you can also expose the UI with:

CODE_AGENT_WEB_HOST=0.0.0.0
CODE_AGENT_WEB_PORT=5173

Then open:

• http://<your-server-ip>:5173

Public exposure safety

This app can read and write files in the configured workspace. Do not expose it publicly without protection.

Recommended minimum:

CODE_AGENT_AUTH_TOKEN=choose-a-long-random-secret

If you set CODE_AGENT_AUTH_TOKEN, the web UI can authenticate by opening it with:

• http://<your-server-ip>:3000/#token=your-secret

or in Vite dev mode:

• http://<your-server-ip>:5173/#token=your-secret

The token is stored in browser local storage after first load and the URL fragment is cleared from the address bar.

One-command Docker startup

If you want a single command instead of running backend and frontend separately:

docker compose up --build

Then open:

• http://127.0.0.1:3000
• or http://<your-server-ip>:3000 from another machine on the network

How it works:

• the container serves both the API and built web UI
• the current project directory is mounted into /workspace
• all file edits from the agent are written back to your host files
• the app inside the container reads CODE_AGENT_WORKSPACE=/workspace

CLI

The CLI still works for direct local agent usage:

npm run cli -- run "explain this repo"
npm run cli -- repl
npm run cli -- index
npm run cli -- search "index manager"

Top-level commands:

• run
• repl
• index
• search
• tools

Useful options:

--workspace <path>   Workspace root, defaults to current directory
--session <id>       Session id, defaults to main
--model <name>       Override model from env
--max-steps <n>      Max model/tool rounds, defaults to 6
--reset              Reset the session before running

Web UI

The web app includes:

• file browser
• Monaco editor with save-back-to-workspace
• indexed code search panel
• agent chat panel with tool activity stream
• index status and manual rebuild

Agent workflow helpers

The agent now has dedicated tools for safer edit workflows inside git repositories:

• git_status
• git_diff
• run_checks
• git_commit

This lets the agent inspect the current working tree, run build/test checks, and create a commit after successful changes.

The automatic commit tool is intentionally conservative:

• it refuses to auto-commit if the repository was already dirty when the agent run started
• it runs git diff --check on staged changes before committing

These safeguards help avoid accidentally sweeping unrelated local edits into an agent-generated commit.

VS Code tasks

If you open the project in local VS Code, these workspace tasks are available in .vscode/tasks.json:

• Code Agent: Dev Server
• Code Agent: Web Dev
• Code Agent: Build
• Code Agent: Test
• Code Agent: Rebuild Index

API surface

Main endpoints:

• GET /api/config
• GET /api/index/status
• POST /api/index/rebuild
• GET /api/search
• GET /api/files/list
• GET /api/files/content
• POST /api/files/content
• GET /api/sessions/:sessionId
• POST /api/agent/stream

Project structure

• src/server.ts: API server entry
• src/api/app.ts: HTTP routes and SSE streaming
• src/agent/loop.ts: tool-calling agent loop
• src/rag/index-manager.ts: indexing, watch, and search
• src/rag/chunker.ts: semantic chunking heuristics
• src/tools/definitions.ts: workspace-safe coding tools
• web/src/App.tsx: React application shell

Prompt gap report

The implementation status against coding-agent-prompt.md is documented in:

• PROMPT_GAP_REPORT.md

Validate

npm run build
npm test