kodo

kodo coordinates installed coding agents so developers can leave a defined task running for hours. It runs locally on macOS, Linux and Windows and is open source under the MIT license. It's aimed at work with a clear goal that can progress without frequent human decisions.

An orchestrator assigns implementation to workers and can call separate architect and tester agents to assess the results. Work spans multiple cycles and context windows, with automatic checkpoints and logs that let interrupted runs resume. Supported backends include Claude Code, Cursor, OpenAI Codex, Gemini CLI, Kimi and Kiro. Teams can mix backends and include specialist reviewers for areas such as security or interface design.

Alongside implementation, kodo has a testing mode that exercises user workflows and error cases, then writes regression tests and fixes confirmed bugs. Its code review mode examines unnecessary complexity, usability problems and architecture, fixes clear issues, and flags ambiguous findings for a decision. A live dashboard and log viewer help developers inspect ongoing and past runs.

Local orchestration doesn't mean local inference: kodo can coordinate through an installed CLI backend or use the Gemini or Anthropic API, which sends orchestrator requests to a cloud service. API use requires a key, and workers use their backend's account arrangements.

Verification has a concrete limit: the orchestrator decides when to request and accept review, so a success status doesn't enforce independent approval. Agents run with full permissions and can access files beyond the project directory.