An Operating Theory of Organized Intelligence.

OpenCollab is inspired by Self-Collaboration.

News

• 🎉 Congratulations! OpenCollab has been selected for support by the Seed Program of the Youth Open Source Special Fund.

What you can run

OpenCollab supports two forms of collaboration on the same agent runtime.

Mode	Command	What it is
Team	opencollab [--team-config FILE] --workspace .	A lead plans the work and spawns specialists that collaborate until the task is done. The agents decide the division of labor.
Workflow	opencollab workflow run NAME	Python defines fan-out, pipeline, loop, and verification behavior while agents complete each step.

Model access, context handling, tool execution, orchestration, and environments have separate extension points. An experiment can change one component at a time.

OpenCollab turns ambitious multi-agent ideas into small, runnable extensions. Researchers describe the collaboration protocol while the framework provides the reusable infrastructure needed to run it.

Edict implements the Three Departments and Six Ministries as a standalone system with roughly 24,000 source lines. Mini Edict implements Edict’s core review-and-dispatch protocol in 239 lines of team and workflow code on the OpenCollab runtime. This roughly hundredfold reduction makes OpenCollab’s value concrete. A substantial collaboration design becomes a compact, readable protocol when it runs on shared infrastructure. Researchers can concentrate on the new idea and turn it into a working system with a fraction of the engineering effort. The same OpenCollab runtime can carry many more collaboration designs. The example includes a bilingual guide and tests.

Quick start

uv sync --locked
cp configs/.env.example configs/.env   # then set OPENCOLLAB_API_KEY
uv run opencollab --workspace .

Point configs/.env at an OpenAI-compatible or Anthropic endpoint. The command starts with the built-in single lead, which may spawn ad-hoc specialists. Never commit real API keys. To use declared roles and a fixed topology, select a team file explicitly.

cp configs/team.example.yaml configs/team.yaml
uv run opencollab --team-config configs/team.yaml --workspace .

For repeatable pipelines, author a Python workflow and run it by name.

uv run opencollab workflow run NAME --args '{"task": "..."}'

See Workflow authoring for a complete module.

Evaluate with OpenCollab-Eval

OpenCollab-Eval is a downstream application built on OpenCollab’s public Python API. It exercises agents, teams, workflows, tools, and environments from outside this repository.

OpenCollab-Eval runs agents on software-engineering benchmarks. It creates an isolated workspace for each task and records the Solver’s patch. It then runs the official tests and keeps the commands and reports needed to inspect the result. It currently supports SWE-bench Pro-Lite and provides a generic task runner for other evaluation workloads.

Datasets, Docker integration, benchmark adapters, experiment reports, and their execution records live in OpenCollab-Eval. This repository contains the collaboration framework.

The evaluation guide explains how to run it. The integrity guide explains how results are checked. MIGRATION.md records the boundary between the repositories.

Documentation

The package guide covers installation, the CLI, the Python API, architecture, and runtime behavior. The configuration guide covers providers, models, and teams.

Mini Edict shows a nine-role institutional workflow. The skills guide documents on-demand instructions. The scripts guide documents launchers and provider diagnostics.

Repository development is documented in CONTRIBUTING.md. Maintainers can follow RELEASING.md when preparing a release. The documentation index links design records and research notes. Benchmark users should begin with the OpenCollab-Eval README.

License

OpenCollab is licensed under the Mulan Permissive Software License v2 (MulanPSL-2.0).

Citation

If you find this project useful, please consider giving it a ⭐ and citing our paper. Our technical paper is also coming soon.

@article{dong2023self,
  author={Dong, Yihong and Jiang, Xue and Jin, Zhi and Li, Ge},
  title        = {Self-Collaboration Code Generation via ChatGPT},
  journal      = {ACM Transactions on Software Engineering and Methodology},
  volume       = {33},
  number       = {7},
  pages        = {189:1--189:38},
  year         = {2024}
}