Introducing Tributary: your coding agents, one workbench

Tributary is an open-source AI workbench for working with the coding agents and local models you already use. It includes a desktop app, a VS Code companion and a CLI.

The idea is straightforward: keep the conversation, real project files and review tools together, even when you change assistants.

Original project artwork. The workbench screenshot below shows the actual app.

Bring your own agent

Tributary connects to native coding agents such as Claude Code, Codex and Gemini CLI, and supports compatible Agent Client Protocol tools such as OpenCode. You can also connect model services including Ollama, LM Studio, Unsloth and llama.cpp.

Your assistant still needs its own account, installation or local service. Model access and available tools depend on that connection. Tributary gives them a shared place to work with your project.

Follow the work in real files

Open a project folder, ask for a change and review the result in the editor. You can inspect Git diffs and red/green line counts, use split editors, run commands in terminals and check the project’s formatting rules.

The workbench also has a preview browser, project checks and Node.js/TypeScript debugging support. It can help with QA workflows such as inspecting imported Playwright results, screenshots and traces.

Start a team from chat

For a task with independent pieces, start with /team, choose workers and follow their progress in Live activity. You can keep more than one background chat moving and review the resulting changes.

Background runs require the app to remain open. Shared project rooms coordinate people and tasks; they do not automatically synchronize everyone’s source files or execute agents on another person’s computer.

Keep project context useful

Project notes can hold decisions, run instructions and gotchas. Bounded excerpts help carry useful context to an assistant without sending every previous conversation. File mentions and agent handoffs help keep the work connected when you switch tools.

Local coding connections support file tools and approved commands. Deep research needs an assistant with web tools; local coding mode does not include a bundled web-search service.

A small Rust bridge, with published measurements

Tributary uses an optional Rust helper for bounded file discovery and ownership of a local model process. The intention is to give specific expensive work an explicit lifetime and keep its buffers bounded.

The recorded Linux results are deliberately scoped:

• Discovery allocation requests fell about 33–66% in the recorded cases.
• Fifteen of 18 CLI scan medians improved; three became slower.
• A six-app CPU-model comparison measured idle app/model PSS of 456.803 → 439.785 MiB.

These are different source-build workloads, not evidence of a universal editor speed ranking. The benchmark guide publishes raw observations, source snapshots, methods and unfavorable cases. Its verifier recalculates the numbers without running another benchmark.

Read the benchmark evidence.

Try the preview

Get Tributary, open a project and use Settings → Connections to connect an assistant. Preview installers are unsigned; release checksums are included.

If you prefer building from source, the README covers Node.js, npm, Git and the optional Rust helper.

I’m looking for concrete feedback: which agent do you use, what work do you want to do, and where does the workflow get in your way?

Source and setup · Report an issue