CodeCraft

A self-hosted, web-based coding harness for AI models – in the spirit of Claude Code and opencode, usable from any browser and installable as a PWA.

• Any model: Anthropic (native Messages API), Google Gemini, opencode Go, fal.ai, OpenRouter and self-hosted models (Ollama, vLLM, llama.cpp, LM Studio) through OpenAI-compatible endpoints.
• Agent loop with tools: read, write, edit, list, glob, grep, bash, web_fetch, todo_write, ask_user, task (subagents), skill – plus every tool of your MCP servers.
• Modes: build (full access), plan (read-only, proposes a plan), craft (study a target → build your own version → check visually → refine), and custom agents.
• Human in the loop: permission requests (with diffs) and questions from agents and MCP servers (elicitation) are forwarded to the UI; optional browser notifications.
• MCP: stdio and Streamable HTTP servers, configured in the UI or the config file.
• Statistics: every model request is recorded (tokens in/out, cache, time to first token, prefill and output speed, cost per model price) – per chat and in a dashboard with daily charts and a weekday × hour heatmap.
• Claude Code compatible project files: AGENTS.md, CLAUDE.md, .claude/{agents,commands,skills} (and .codecraft/…).
• Configuration in a commented TOML file that the web UI edits in place (comments are preserved).
• i18n: English and German; adding a language is one file.

Quick start (Docker)

cp .env.example .env          # set CODECRAFT_TOKEN, WORKSPACE and provider keys
docker compose up -d --build

Open http://<host>:8787, sign in with the token and pick a project directory. On first start the commented default config is written to data/config.toml. Put the service behind a TLS reverse proxy for remote access – the PWA install prompt requires HTTPS.

Configuration

See config.example.toml – every option is documented there. Highlights:

• [providers.<name>] – kind = "anthropic" | "openai", base_url, api_key (${ENV} references recommended), optional models with context size, effort and prices per 1M tokens. Without models, the list is fetched from the provider.
• [mcp.<name>] – type = "stdio" (command/args/env) or "http" (url/headers).
• [permissions] – allow / ask / deny per tool (globs such as mcp__*), plus [permissions.bash_rules] for shell command patterns. Compound commands are only auto-allowed when every part is allowed.
• [agents.<name>] – custom modes and subagents.
• [craft] – Chromium binary, screenshot widths, clean room vs. deep mode.

Development

Requirements: Rust (edition 2024), Node ≥ 20.19.

# backend (writes ./data/config.toml on first start)
cd server && CODECRAFT_CONFIG=../data/config.toml cargo run

# frontend with hot reload, proxies /api to 127.0.0.1:8787
cd web && npm install && npm run dev

For local development set bind = "127.0.0.1:8787" and workspace to your projects directory in data/config.toml.

Tests: cd server && cargo test, type check: cd web && npx tsc --noEmit.

Architecture

server/ (Rust, axum + tokio)
  provider/   Anthropic + OpenAI-compatible streaming clients
  runner.rs   agent loop: streaming, tool execution, permissions, subagents, compaction
  tools/      built-in tools (fs, bash, web, agent, craft)
  mcp.rs      MCP client (stdio, Streamable HTTP, elicitation)
  prompts.rs  agents/modes, system prompt, AGENTS.md, slash commands, skills
  permission.rs, config.rs (toml + toml_edit), store.rs (SQLite), api.rs (REST + WebSocket)
web/ (React + Vite, PWA)
  src/components  chat, composer, permission/question cards, settings, statistics
  src/i18n        en (reference), de

The server streams all events over one WebSocket (/api/ws); the UI reconnects and resyncs automatically, so long-running agents keep going when the browser is closed.

Roadmap

• Background shell processes, file checkpoints/undo, @file mentions
• MCP OAuth, resources and prompts; remote MCP over SSE (legacy)
• Hooks (pre/post tool), output styles, per-project config overrides
• Image generation tools (fal.ai), OpenAI Responses API
• Multi-user accounts