huazhx/OpenCollab：See https://github.com/RISE-X-Lab/OpenCollab

A multi-agent coding framework with programmable collaboration and a controllable runtime.

_{English · Chinese}

OpenCollab follows a rigorous and restrained software engineering philosophy. In version 0.9.3, its harness core in the OpenCollab/opencollab directory, written in under 50,000 lines, provides comprehensive research instrumentation and statistics while outperforming Claude Code, Codex CLI and OpenHands across several coding benchmarks. The core implementations of these mature competitors have codebases spanning well over 100,000 lines. OpenCollab also won first place at the OAIC Agentic Software Factory Hackathon, which featured a $24,000 prize pool.

_{Full results}

News

⭐ 2026-11-07 — OpenCollab reached 1,000 stars on GitHub.
🏆 2026-10-11 — Won first place at the OAIC Agentic Software Factory Hackathon using OpenCollab! 🏆
🎉 2026-10-08 — OpenCollab 0.9.3 (v1.0.0 PreRelease) includes the latest Duo V8 workflow and all the statistical features described in our paper to support research experiments. Following extensive testing and bug fixes, the framework now offers the complete feature set planned for 1.0.0. We will release version 1.0.0 once our paper is accepted. See the changelog.
2026-09-29 — Our paper is on arXiv: OpenCollab: A Multi-Agent Coding Framework with Programmable Collaboration and Controllable Runtime.
2026-09-27 — OpenCollab 0.8 ships Duo built in and makes Single2 the default Base agent.
🎉 OpenCollab has been selected for support by the Seed Program of the Youth Open Source Special Fund.

Quick start

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

configs/.env accepts any OpenAI-compatible or Anthropic endpoint. Never commit real API keys. The command starts the built-in Self-Collaboration team: the Analyst delegates to independent Coder and Tester workspaces and adopts the Coder’s changes before testing. See cooperative teams for adoption, conflicts, and recovery. Custom teams choose their own tools. Type /help in the terminal for navigation, saving and exit controls. To define additional roles, create and select an explicit team file as shown below. To run Duo on a Git repository, use this command.

uv run opencollab workflow run duo --workspace /path/to/repository \
  --task 'Fix the public issue described here.' \
  --args '{"allow_unisolated_shell":true}'

Place run options such as --workspace, --model and --budget after workflow run NAME. To list workflows in another workspace, use workflow list --workspace PATH. Options placed before the workflow subcommand are interpreted as belonging to the root interactive mode and will produce a usage error.

Program the collaboration

Team. A YAML file declares each role’s prompt, model, tools, and token allowance, the context policy, and who may message whom. Start from the built-in team and edit it:

uv run opencollab team init team.yaml   # edit roles, prompts, tools, and topology
uv run opencollab team show --team-config team.yaml
uv run opencollab --team-config team.yaml --workspace .

Workflow. A Python module issues every handoff. Duo is one; write your own and run it with opencollab workflow run NAME.

Examples

First place at the OAIC Hackathon. OpenCollab powered the first-place solution at the OAIC Agentic Software Factory Hackathon, which offered a $24,000 prize pool. Explore the winning approach in the ARC-Bench r6 example and build on it for your own applications.

From 24,000 lines to a 239-line example. Edict is a multi-agent project of roughly 24,000 lines. Mini Edict’s original implementation recreates its core collaboration protocol in 239 lines of team and workflow code, showing how OpenCollab makes sophisticated multi-agent designs compact and easy to extend.

98.3% on a NeurIPS 2026 benchmark. IO2CodeBench was introduced in FromIO, accepted at NeurIPS 2026. Built on OpenCollab, this example reports a resolve rate of 98.3% (113/115), up from its 78.3% (90/115) fixed-example baseline with the same gpt-6-luna model, a 20-percentage-point gain across the benchmark’s 115 non-multimodal tasks. See the results.

Verify it happened

In a Team, the agents decide the handoffs, and a declared team does not guarantee that they happen. In our runs, the lead agent often briefed its teammates and then did the work itself. OpenCollab enforces the declaration and records what each agent did:

a message along an undeclared edge is refused, and the refusal is recorded;
each agent’s token budget is checked before every model call;
--trace writes one JSONL record per step, stamped with the run, agent, role, event type, and token usage.

From these records we measure Adherence, the share of runs in which the declared organization actually happened. Changing one setting of the same team raised it from 47.2% to as much as 97.2%. See how Adherence is measured.

Learn more

Guide	What it covers
Package guide	CLI, Python API, architecture, and runtime behavior
Configuration guide	Providers, models, and team files
Duo guide	Running the dual-coder workflow from the CLI and SDK
Weave guide	Ordered sessions, executable checks, bounded repair and continuation
Benchmark results	Tokens, cost, the paired Duo–Base test, and how to run an evaluation
Adherence	What the runtime records, and how we measure Adherence
OpenCollab-Eval README	Running agents on software-engineering benchmarks
Documentation index	Skills, migration guides, and design records
CONTRIBUTING.md	Development checks, testing, and releases

Citation

If you find OpenCollab useful, please give it a ⭐ and cite our work. Our paper is OpenCollab: A Multi-Agent Coding Framework with Programmable Collaboration and Controllable Runtime. OpenCollab builds on Self-Collaboration:

@misc{hsu2026opencollab,
  author        = {Hsu, Chun-Wah and Gong, Kai and Wu, Yu and Chen, Xianhe and
                   Li, Hanyu and Li, Jie and Liu, Mengyang and Liu, Zhixuan and
                   Tang, Naisheng and Chi, Jiaying and Fan, Ziheng and He, Xuning and
                   Yang, Xiaokang and Jiang, Xue and Dong, Yihong},
  title         = {OpenCollab: A Multi-Agent Coding Framework with Programmable Collaboration and Controllable Runtime},
  year          = {2026},
  eprint        = {2609.38345},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2609.38345}
}

@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}
}

License

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