Cowork Forge

What it adds beyond the summary

The summary names the architecture and features; it omits the concrete integration surface, run commands, licence, and operational constraints an engineer needs to decide.

Consumption model. cowork-core is meant to be embedded, not just run: it's callable from Python/Java/NodeJS via FFI as a high-performance library. Three entry points — the core lib, cowork-cli, and the optional Tauri+React GUI. License is MIT. Built on Rust edition 2024.

How to use it

git clone https://github.com/sopaco/cowork-forge.git
cd cowork-forge
cargo build --release
# CLI binary at target/release/cowork

CLI workflow:

cowork init --name "My Project"
cowork iter --project "my-project" "Build a REST API for task management"
cowork iter --project "my-project" --base iter-1 --inherit partial "Add user profiles"
cowork continue iter-1-1234567890

GUI:

cd crates/cowork-gui
npm install   # or: bun install
cargo tauri dev

Config lives in a config.toml in the OS app-data dir. The example uses model_name = "gpt-5" against an OpenAI-compatible api_base_url; an OpenAI-compatible LLM key is required, embedding config is optional. External coding agent is off by default (enabled = false); enabling it swaps the built-in adk-rust agent for OpenCode/iFlow/Codex/Gemini CLI/Claude CLI over stdio or websocket, with fallback to built-in.

Details worth knowing before adopting

• Iteration inheritance has three modes, which the summary doesn't distinguish: None (fresh), Full (copies code + artifacts), Partial (artifacts only, regenerates code — the default). This determines whether existing code survives an evolution iteration.
• LLM calls are rate-limited to 30 req/min with concurrency=1 in the core. This is a hard throughput ceiling for the whole pipeline, not tunable per the page.
• Sandboxing is real, not aspirational for adoption decisions: command validation blocks rm -rf/sudo, path access control restricts sensitive dirs, a build-tool whitelist gates execution, timeouts cap runtime, and a watchdog flags agent deviation from objectives. File ops are confined to project boundaries.
• Persistence is JSON-based with workspace isolation; ~30+ ADK tools handle file/command/validation/memory ops.

Limits

Everything here is the project's own README — no independent benchmarks, no cost figures, no evidence of output quality on real projects. The gpt-5 model name and concurrency=1 rate limit mean cost and wall-clock time for a full 7-stage run are unstated but likely nontrivial. Whether FFI bindings for Python/Java/NodeJS are packaged or must be built yourself is not stated.