Darwin Code Engine

BYOK-only execution engine for Darwin's autonomous coding pipeline.

Directory Structure

• darwin-cli/ — Node.js CLI entry point
• darwin-rs/ — Rust sandbox runtime (darwin-code-exec, darwin-code-core, darwin-code-tui)
• docs/ — Documentation

Install / Run

Preferred: run from source

Use source install when you need the latest GitHub code.

Warning: build artifact size and disk usage

Rust projects can produce large incremental build caches. Long-running local development may grow target/ beyond 100 GB and put system storage at risk. This checkout is configured with a low-level Cargo command interceptor that silently sweeps redundant caches older than one day in the background. If disk space is still tight, run cargo sweep --time 1 manually.

cd /your/own/path/darwin-code
cp .env.example .env
$EDITOR .env
$EDITOR config.toml  # optional: change provider/model/API endpoint
cd darwin-rs
cargo clean
cargo run -p darwin-code-cli --bin darwin-code

config.toml is the checked-in default source-run configuration. It must stay secret-free: commit provider ids, endpoints, model metadata, and api_key_env names only. Put real keys in ${DARWIN_CODE_HOME}/.env or your parent shell environment. When running from the source checkout root or any child directory such as darwin-rs/, Darwin Code resolves the checkout root as DARWIN_CODE_HOME and loads the root .env automatically. You do not need to source .env; if you do source it manually from darwin-rs/, use source ../.env. Set DARWIN_CODE_HOME only if you need a different config directory.

Use another config directory:

mkdir -p /path/to/darwin-code-home
cp config.toml /path/to/darwin-code-home/config.toml
cp .env.example /path/to/darwin-code-home/.env
$EDITOR /path/to/darwin-code-home/.env
export DARWIN_CODE_HOME=/path/to/darwin-code-home
cargo clean
cargo run -p darwin-code-cli --bin darwin-code

Alternative: npm install

The npm package is convenient, but it may lag behind the latest GitHub source.

npm install -g @darwin-code/darwin-code
darwin-code --version
darwin-code --help

Then edit the packaged default config and set provider/key environment names:

$EDITOR "$(npm root -g)/@darwin-code/darwin-code/config.toml"
$EDITOR "$(npm root -g)/@darwin-code/darwin-code/.env"

You can also keep configuration under your user directory by copying the packaged default config and editing it there:

mkdir -p ~/.darwin
cp "$(npm root -g)/@darwin-code/darwin-code/config.toml" ~/.darwin/config.toml
$EDITOR ~/.darwin/config.toml
$EDITOR ~/.darwin/.env

At minimum, verify these fields in config.toml and put the real key in ${DARWIN_CODE_HOME}/.env:

model_provider = "qwen"        # Current provider id
model = "qwen3.6-plus"         # Current model; /model lists this provider's models

[openai_compatible.qwen]
name = "Qwen"
base_url = "https://token-plan.cn-beijing.maas.aliyuncs.com/compatible-mode/v1"
wire_api = "chat-completions"  # Use chat-completions or responses to match the endpoint
api_key_env = "QWEN_API_KEY"  # Real key lives in ${DARWIN_CODE_HOME}/.env

To add a model, add it under the matching provider:

[openai_compatible.qwen.models."your-model"]
display_name = "Your Model"
context_window = 128000
max_output_tokens = 8192
thinking = true
multimodal = true
tools = true

BYOK quick start

Darwin Code's default product path is BYOK (Bring Your Own Key): select the provider, model, base_url, and api_key_env explicitly in config.toml, then put the real key in ${DARWIN_CODE_HOME}/.env or an external environment variable. Each provider declares selectable models under its own models table. The TUI /model command shows that model list and writes the selected model_provider plus model back to the config.

The local TUI/CLI does not derive key names from provider ids. It loads ${DARWIN_CODE_HOME}/.env early, without overriding existing parent environment variables. Direct api_key remains a local backward-compatible escape hatch, but docs and templates use api_key_env. Darwin Code does not support hosted OAuth/device sign-in, workspace/cloud auth, command-backed bearer tokens, keychain storage, or implicit OpenAI login.

DeepSeek provider example

model_provider = "deepseek"
model = "deepseek-v4-flash"

[openai_compatible.deepseek]
name = "DeepSeek"
base_url = "https://api.deepseek.com"
wire_api = "chat-completions"
api_key_env = "DEEPSEEK_API_KEY"

[openai_compatible.deepseek.models."deepseek-v4-flash"]
display_name = "DeepSeek v4 Flash"
context_window = 128000
max_output_tokens = 8192
thinking = true
multimodal = true

[openai_compatible.deepseek.models."deepseek-v4-pro"]
display_name = "DeepSeek v4 Pro"
context_window = 128000
max_output_tokens = 8192
thinking = true
multimodal = true

Generic OpenAI-compatible endpoint

Any external gateway that implements an OpenAI-compatible API, such as a company gateway or a local proxy, appears in Darwin Code as a standard endpoint:

model_provider = "my-gateway"
model = "your-model"

[openai_compatible.my-gateway]
name = "My Gateway"
base_url = "http://127.0.0.1:8080/v1"
wire_api = "responses"
api_key_env = "MY_GATEWAY_API_KEY"

[openai_compatible.my-gateway.models."your-model"]
display_name = "Your Model"
context_window = 128000
max_output_tokens = 8192
thinking = true
multimodal = true

wire_api must match the protocol path that the provider actually supports. DeepSeek's official API uses chat-completions, which sends requests to /chat/completions. Gateways compatible with the OpenAI Responses API use responses, which sends requests to /responses. Do not set DeepSeek's base_url to a /v1 URL and then use Responses API mode; that would hit DeepSeek's unsupported /v1/responses path.

OpenAI / Gemini / Claude native providers

Model IDs below are current provider examples; access still depends on your account, region, and provider rollout.

model_provider = "openai"
model = "gpt-5.5"

[openai]
base_url = "https://api.openai.com/v1"
api_key_env = "OPENAI_API_KEY"

model_provider = "gemini"
model = "gemini-3.1-pro-preview"

[gemini]
base_url = "https://generativelanguage.googleapis.com/v1beta"
api_key_env = "GEMINI_API_KEY"

model_provider = "claude"
model = "claude-opus-4-7"

[claude]
base_url = "https://api.anthropic.com"
api_key_env = "CLAUDE_API_KEY"

The current runnable projection primarily covers [openai_compatible.<id>]. If Gemini or Claude native runtime adapters are not yet available, selecting those providers directly will fail with an explicit error. To use Gemini or Claude through an OpenAI-compatible gateway in the meantime, configure [openai_compatible.<id>] with the gateway base_url and api_key_env.

License and upstream

Darwin Code is a downstream derivative of OpenAI Codex. See LICENSE, NOTICE, and docs/license.md for upstream attribution, downstream modifications, and third-party license notes.