← all repositories
mdiri-ai/SoduWork

AI work that isn't done until a human signs off

SoduWork is a self-hosted workbench where agent output only counts after acceptance criteria are met, an AI review passes, and the requester approves.

★6.5k stars JavaScript AgentsCoding Assistants
SoduWork
Too new to rank. Velocity needs a few days of history.
not enough history yet
star history

What it does

SoduWork is a self-hosted, Electron-and-FastAPI platform for running team work through AI agents. Requests enter a shared issue pool: AI drafts acceptance criteria, the request gets assigned to a person or an agent, an agent executes it, a separate AI review checks the criteria, and the requester signs off. It fronts nine agent harnesses — Pi (built in), OpenCode, Claude Code, Codex, Qwen Code, Kimi Code, DeepSeek Harness, and more — through one adapter interface, so you can switch harnesses mid-conversation and keep history and files.

The interesting bit

The core bet is that the bottleneck isn’t agent capability, it’s trust and reuse. One harness can do the work while a second, on a different model, reviews it against the criteria — and bounce it back with reasons if it fails. When a workflow works, you save the chat as a plugin that, after review, becomes a team asset anyone can run and improve.

Key highlights

  • Multi-harness with mid-conversation switching; anything speaking the Agent Client Protocol can be added via a catalog declaration
  • Smart dispatch: cheap/local models for routine steps, strong models where it matters, long tasks on your own cloud runners
  • Governance baked in: plugin review before sharing, role-based grants, usage/cost tracking, audit log
  • Parallel execution across agents and issues via the Hub; agents never hand work to other agents unprompted
  • Docker-based team deployment with Hub, admin console, and PostgreSQL; fully self-hosted

Caveats

  • Public beta — the README itself says to evaluate before production use
  • Setup friction is real: Python 3.12+, Node 22.19+, and the quickstart troubleshooting section is long enough to suggest first-run pain (empty agent lists, models not appearing, connection probes that pass while inference fails)
  • Harnesses like Claude Code and Codex use their own accounts and sign-ins, not your central model catalog

Verdict

Worth a look if your team’s problem is trusting, routing, and preserving agent output rather than generating more of it. If you just want a single-agent chat client, this is considerably more machine than you need.

Frequently asked

What is mdiri-ai/SoduWork?
SoduWork is a self-hosted workbench where agent output only counts after acceptance criteria are met, an AI review passes, and the requester approves.
Is SoduWork open source?
Yes — mdiri-ai/SoduWork is open source, released under the MIT license.
What language is SoduWork written in?
mdiri-ai/SoduWork is primarily written in JavaScript.
How popular is SoduWork?
mdiri-ai/SoduWork has 6.5k stars on GitHub.
Where can I find SoduWork?
mdiri-ai/SoduWork is on GitHub at https://github.com/mdiri-ai/SoduWork.

heatdrop uses Google Analytics to see which pages get read — nothing else. Your call. How we handle data.