An open-source AI coding assistant: a FastAPI backend with a LangGraph agent, tool execution, streaming responses (Server-Sent Events), and a web UI. It talks to Anthropic and OpenAI APIs, runs Bash and file tools against a configurable working directory, and stores conversations in SQLite.
- Configure
.envwith a valid API key, then run:uvicorn app.main:app --reload --host 0.0.0.0 --port 8000 - Open http://localhost:8000 in your browser.
- Show a short flow: e.g. new chat → send a message → watch streaming text and tool cards (toggle theme or mode if you want extra polish).
- LangGraph agent — The model runs in a graph with an LLM node and a tools node; tool calls can require approval before execution. Responses stream to the client as SSE so you see text and tool activity in real time.
- Multi-provider models — Choose a model in the UI or in the JSON body; the server picks the right provider from the model id (and optional
provideroverride).
| Mode | Purpose |
|---|---|
| Agent | Full tool set: read/write files, Bash, web, sub-agents (Agent, Coder), notebooks, todos, etc. |
| Ask | Read-only exploration: Read, Glob, Grep, WebFetch, WebSearch — no edits, shell, or sub-agents. |
| Plan | Same read-only tools as Ask, plus TodoWrite after the plan. The model explores the repo, writes a plan, then records implementation steps as todos. The UI shows an editable plan panel and Build plan, which switches to Agent mode and sends an implementation prompt. |
- Conversation sidebar — List of chats; hover a row to reveal delete (trash). Deletion uses a confirmation dialog; if the chat is open while streaming, the stream is aborted first. Theme preference is stored in
localStorage. - Dark / light theme — Theme button in the header (sun / moon) toggles appearance; choice persists across visits.
- Model and mode — Dropdowns for model and chat mode (Agent / Ask / Plan).
- Auto-approve tools — When enabled, tools that would normally ask for approval run without prompting.
- Tool cards — Chronological tool calls with expandable inputs/outputs and status badges.
- File autocomplete — Type
/in the message box to search paths under the configured working directory. - Plan workflow — After a Plan-mode reply, edit the plan in the panel and click Build plan to run implementation in Agent mode.
| Tool | Role |
|---|---|
| Bash | Shell commands (timeout, optional approval) |
| Read | Read files (line numbers, images, binary hint) |
| Write | Create/overwrite files |
| Edit | Exact string replace in files |
| Glob | Find files by glob |
| Grep | Content search (ripgrep when available) |
| WebFetch | Fetch URL as text |
| WebSearch | Web search (DuckDuckGo) |
| NotebookEdit | Jupyter notebook cells |
| TodoWrite | In-memory task list (server process scope) |
| Agent | Nested sub-agent with the same tool surface (respects parent chat mode) |
| Coder | Coding sub-agent with a restricted tool set (no nested agents) |
- Tool approval — Destructive or sensitive tools can pause until the user approves in the UI (
POST /api/tool-approve). - Working directory — Set
WORK_DIRso file tools and/autocomplete resolve to your project root. - Conversation storage — SQLite stores threads, messages, model, and token counts.
A separate MCP package can browse a codebase over STDIO or HTTP. See MCP Explorer below.
- Python 3.11+
- At least one of: Anthropic API key, OpenAI API key
Option A — uv (recommended) — uses the pinned uv.lock for reproducible installs:
cd fast-code
uv sync # runtime deps from uv.lock
uv sync --extra dev # include dev deps (pytest, ruff, …) for contributingOption B — pip
cd fast-code
python3 -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -r requirements.txt
# or: pip install -e ".[dev]"The uv.lock file should be committed to Git (see Contributing). It records exact dependency versions; after changing pyproject.toml, run uv lock and commit the updated lockfile.
cp .env.example .env
# Set ANTHROPIC_API_KEY and/or OPENAI_API_KEY, optional WORK_DIR, etc.uvicorn app.main:app --reload --host 0.0.0.0 --port 8000Open http://localhost:8000 for the web UI.
export ANTHROPIC_API_KEY=sk-ant-...
docker compose up --build| Method | Path | Description |
|---|---|---|
| GET | /health |
Status, default models, configured keys |
| GET | / |
Web UI |
| POST | /api/chat |
Streaming chat (SSE). Body includes messages, optional model, conversation_id, mode (ask | agent | plan), auto_approve, etc. |
| POST | /api/tool-approve |
Resolve pending tool approvals |
| GET | /api/files/list |
Directory listing for UI autocomplete |
| GET | /api/conversations |
List conversations |
| POST | /api/conversations |
Create conversation |
| GET | /api/conversations/{id} |
Load conversation + messages |
| DELETE | /api/conversations/{id} |
Delete conversation (used by the UI trash control) |
| GET | /api/tools |
List tool definitions |
| GET | /docs |
OpenAPI (Swagger UI) |
curl -X POST http://localhost:8000/api/chat \
-H "Content-Type: application/json" \
-d '{
"messages": [{"role": "user", "content": "List Python files here"}],
"model": "gpt-4o-mini",
"mode": "agent",
"stream": true
}'SSE event types include: message_start (may include chat_mode), content_block_delta, tool_use_start, tool_use_end, tool_approval_request, tool_execution_start, tool_result, tool_denied, message_stop, error, and a final [DONE].
See .env.example for the full list. Common entries:
| Variable | Meaning |
|---|---|
ANTHROPIC_API_KEY |
Claude models |
OPENAI_API_KEY |
OpenAI models |
WORK_DIR |
Default project root for tools and UI / paths |
DATABASE_URL |
Async SQLite URL for conversations (see .env.example) |
AUTH_TOKEN |
Optional bearer token for API/UI access |
Issues and pull requests are welcome. For larger changes, opening an issue first helps align on direction.
- Commit
uv.lockto Git. It pins exact package versions so CI and contributors get the same dependency graph when using uv. - If you add or bump dependencies in
pyproject.toml, runuv lock(oruv sync, which updates the lockfile as needed) and include the updateduv.lockin your PR. - People who install with pip +
requirements.txtare unaffected byuv.lock; both flows are documented in Install.
Run the suite before submitting a PR:
# Using uv (uses locked versions from uv.lock)
uv sync --extra dev
uv run pytest tests/ -q# Using pip + editable install
python3 -m venv .venv && source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -e ".[dev]"
pytest tests/ -qUseful variants:
pytest tests/ -v --tb=short # verbose, shorter tracebacks
pytest tests/test_chat.py -q # single file
pytest tests/ -k "chat_modes" -q # tests whose name contains chat_modesMost tests do not call real LLM APIs; they use the in-memory ASGI client or unit assertions. If a test fails only with network or keys, check AUTH_TOKEN / env and that nothing in .env overrides expected defaults.
fast-code/
├── app/
│ ├── main.py # FastAPI app, lifespan, routes
│ ├── config.py
│ ├── dependencies.py
│ ├── agent/ # LangGraph graph, LLM, streaming, tools bridge
│ ├── api/ # chat, sessions, files, …
│ ├── core/ # prompts, chat modes, approval, registry
│ ├── llm/ # Provider routing
│ ├── tools/ # Bash, Read, Write, Agent, Coder, …
│ ├── services/ # SQLite store
│ └── models/
├── docs/ # e.g. demo.gif for README
├── mcp_explorer/ # Optional MCP server
├── web/ # Static UI (HTML, CSS, JS)
├── tests/
├── Dockerfile
├── docker-compose.yml
├── pyproject.toml
├── uv.lock # Pin dependencies (commit this; use with uv)
└── requirements.txt
Runtime artifacts (e.g. local DB, data/ workspace files, .venv) should stay untracked; see .gitignore.
- Never commit
.env— it is listed in.gitignore; keep API keys only on your machine or in CI secrets. - If a key was ever pasted into a tracked file or a public issue, revoke it in the provider dashboard and create a new one.
Optional codebase exploration over MCP:
python -m mcp_explorer.stdio
# or HTTP/SSE: python -m mcp_explorer.httpSet SRC_ROOT to the tree you want to expose.
MIT