yangheng95/opencorvus: DIY your own agent team harness for long-horizon agent work

The open-source harness for long-horizon agent work

Expert squads carry long, complex tasks forward, retain progress, and produce results you can check.

English | 简体中文

Website · Quickstart · Download · Expert Squads · Long-horizon · Composition · Evolution

OpenCorvus is an open-source harness for long, complex agent tasks. It brings models, tools, context and expert-squad coordination into one runtime.

• Carry work forward: research, implementation and independent review connect through clear responsibilities.
• Retain progress: task state and context support continued work and recovery after interruption.
• Check the result: tool calls, artifacts and review evidence stay in the work record.

Task gallery

Explore task inputs, models and coordination graphs: CUDA training and a research paper with GPT-5.6 Sol, and a tank game with DeepSeek V4 Flash.

The tank-game case involved 19 acceptance requirements, source data for 35 stages and two correction rounds.

!Tank battle execution graph

Roles and repair themes come from task records; edges illustrate the workflow and repair origins do not attribute events. The task included an operator decision and has test-report and fidelity limits: read the evidence.

Quick Start

Download the desktop app → configure a reachable model → choose a project and optionally a squad → submit a goal and acceptance criteria.

Without an explicit Squad selection, the creator searches installed Squads first and checks promising candidates. A very well matched Squad is used directly; otherwise Dynamic composes Task-local experts from real capability envelopes. Manually selected Squads retain their identities and capabilities. Scheduling can combine DAGs, local refinement loops and conditional branches; package workflows are optional guidance. SDK and scheduling contract.

Or start from source:

git clone https://github.com/yangheng95/opencorvus.git
cd opencorvus
bun install
bun run --cwd packages/opencorvus build
bun packages/opencorvus/src/index.ts doctor

Installation guide · Expert Squads

Work requires an online runtime and reachable model provider. Results depend on the task, model and available evidence. Important permissions and publishing decisions remain yours.

Further reading

The detailed material is now organized in the blog:

• Cases: coordinating expert squads
• Mechanisms: the long-horizon harness · Squad revisions
• Research: AutomationBench · The paper
• Practice: deployment and integrations · Choosing and boundaries

Open-source acknowledgements

OpenCorvus evolved from the OpenCode codebase and still carries explicitly synchronized OpenCode work in its model provider, GitHub Copilot, and provider-plugin surfaces. We are grateful to the OpenCode maintainers and contributors for that foundation.

Major runtime and distribution dependencies include:

• Runtime and agent core: Bun,

• Open interoperability: the official

• Desktop application: Tauri,

• Execution and evidence: Playwright,

• Packaged command-line runtime: Node.js and

• Interactive workbench: CodeMirror,

• Built-in capability sources: the bundled design and interview Skills adapt ideas

• Documentation: Astro and

The repository manifests and THIRD_PARTY_NOTICES.md contain the complete dependency and notice records. Each upstream project keeps its own license and trademarks.

License

MIT