cocode

Cocode

A ready-to-run DeepSeek Harness distribution.

English · 简体中文

Most coding agents make the choices for you and take the choice away with them: which tools are available, how work gets scheduled, what the interface looks like — all settled upstream. Cocode wants the other half of that bargain. It makes the choices for you and leaves the screwdriver within reach: the default setup works the moment you install it, with nothing to study first, and when it doesn't fit you can open it up, see exactly how it was put together, and change it to your liking. Everything else in Cocode follows from that.

The runtime underneath is DeepSeek Harness, which makes an argument of its own: everything is a plugin. Models, tools, skills, sessions, sandboxes, and scheduling all come together through configuration, and Harness takes no position on what you should assemble. Cocode assembles one of those, and puts the assembly out in the open in the product instead of burying it in a config file. Project status: Developer preview. Cocode is built on the DeepSeek Harness developer preview. It is an independent distribution, not an official DeepSeek product, and upstream compatibility may change. The repository supports source builds on macOS, Windows, and Linux; the current release scripts target macOS and Windows artifacts, and this repository does not include a hosted backend or a vendored Harness checkout. Install and try it in five minutes The published TUI package installs the matching @cocode-agency/host-supervisor automatically. For every release, keep @cocode-agency/tui and @cocode-agency/host-supervisor on the same version line; do not mix versions from different releases. GUI installers are downloaded separately from the matching GitHub Release and are not installed through n

1. Install the GUI from GitHub Releases, or run npm install --global @cocode-agency/tui. 2. Start the GUI, or run cocode doctor followed by cocode in a real TTY. 3. Choose Cocode Pro / Cocode Nut, or enter your own DeepSeek-compatible API key. 4. Open a workspace and ask the agent to inspect a file or explain the project. 5. If startup fails, save the output of cocode doctor and cocode host status --json for troubleshooting. Remove credentials and private session data before sharing diagnostics. Why Cocode

The agent itself comes apart. Which tools a session runs, what prompts it carries, which capabilities are on — that is a plugin assembly called a preset, and you can read its agent.cordis.yml right in the interface. Four presets ship built in — Standard, PTC, Minimal, and Creator — and you copy one to make it yours. Elsewhere you get to swap the model. Here you get to swap the agent.

A workspace, not a chat box. Files, Git, a terminal, a built-in browser, and diff previews sit alongside the session. What the agent changed and what it ran is right there next to you, and you can take over by hand at any point instead of carrying context between windows.

The desktop and the terminal pick up the same session. GUI and TUI connect to the same Host and read and write the same session record. Push a task forward at your desk, then SSH in and continue from the terminal — presentation state never enters the session log, so any client can rebuild the full conversation from it.

The model is your call too. An official DeepSeek key, any OpenAI-compatible self-hosted or gateway endpoint, or Cocode Pro's hosted Cocode Nut service — all three are available, and you can switch between them.

Cocode is not trying to be another chat window. It won't parade a full chain of thought at you as a feature, and it won't hand the model every tool by default. What it wants to be is a workspace you trust with real work — and one you can reshape to your own liking. Two entry points, one session Both attach to the sam