One workspace
Sessions, files, terminals, Git, and runtime state stay in one context.
Checking sign-in status
01AI CODING AGENT / WORKSPACE 01
An AI coding agent that can move a real task forward
From understanding a repository to editing files, running commands, reviewing diffs, and checking results, Cocode turns a development task into an inspectable workflow.
Capabilities and limits follow the current release
02Direct answer
Start with what it is.
Cocode is an AI coding agent for real repositories. It keeps models, files, terminals, tools, Sessions, diffs, and tests in one inspectable development workflow. It is built for developers who need to finish and verify cross-file tasks, not just chat about pasted snippets.
- Product type
- Desktop GUI and terminal TUI AI coding client
- Primary work
- Understand repositories, edit code, run commands, review diffs, verify results
- Model access
- Cocode Nut, DeepSeek-compatible APIs, self-hosted endpoints, or gateways
- Safety boundary
- Permissions, confirmation, Git diffs, and real tests constrain impact
Confirm before impact
Review file writes, commands, and network actions before they continue.
Two clients, one session
The desktop GUI and terminal TUI connect to the same Host and Session.
02Why it matters
More than a chat box with autocomplete
Cocode is built for complete tasks: understand the repository, make a plan, run tools, show the diff, and verify the result.
Choose the model and runtime you trust
Use Cocode Nut, a DeepSeek-compatible API key, a self-hosted endpoint, or your own gateway. Cocode is not tied to one model vendor.
↗
Clear boundaries make real development work easier to trust.
Start at your desk, continue over SSH
Begin in the GUI and keep going from the TUI or an SSH session. Presentation state stays out of the session log so another client can rebuild the task.
05Working flow
Move from setup to verification with a clear result at every step.
This is an executable, reviewable path through a real repository—not an abstract feature list.
- 01
Connect a real repository
Choose a workspace and let the agent read its structure, dependencies, conventions, and current Git state.
- 02
Define goals and constraints
State acceptance criteria, protected areas, and runtime limits before the agent fills in missing context.
- 03
Plan and run tools
Break work into reviewable steps, then read files, edit code, run commands, or use the network within permission boundaries.
- 04
Review diffs and tests
Compare actual file changes, command output, and test results with the goal instead of trusting a prose summary.
- 05
Deliver or repair
Ship a verified result or keep the Session context and continue repairing failed checks.
06Use cases
Decide by the work you actually need to finish.
Cross-file bug fixes
Trace a call path, change the implementation, add regression coverage, and verify with real commands.
APIs and integrations
Update routes, data structures, client calls, documentation, and tests as one task.
Refactors and migrations
Lock down behavior first, then change the code in stages while running type checks and tests.
Remote diagnosis
Continue through TUI and SSH on servers, containers, or machines without a graphical environment.
07Before you start
Runtime requirements and compatibility boundaries
- An accessible local or remote workspace
- Cocode GUI or Cocode TUI in a TTY-capable terminal
- Cocode Nut, a DeepSeek-compatible API, or a self-hosted model endpoint
- Optional: Git for reviewable diffs and recoverable history
08Try it
Shortest start path
Check the Host, configuration, and terminal before entering the workspace.
cocode://workspaceREADY
cocode doctor
cocode09Facts and boundaries
An AI coding agent works only when every layer supports the task.
Client behavior, model capability, and workspace permissions must align. These claims stay within boundaries the Cocode workflow can verify.
| Capability | Status | Conditions and evidence scope |
|---|---|---|
| Repository reading | Supported | Limited to workspaces available to the current user and sandbox. |
| Edits and commands | Permission-aware | Writes and commands may require confirmation and remain subject to OS permissions. |
| GUI / TUI continuity | Same-scope support | Both clients must use the same Host, Profile, and Session data. |
| Fully offline use | Not a default promise | Inference, sign-in, updates, or external tools may still need a network. |
10Cocode perspective
A reliable agent does not just answer; it proves the result.
Cocode treats actual file changes, command exit status, and test output as completion evidence. A model summary explains the work but cannot replace repository and runtime proof.
- Maintained by
- Cocode Agency
- Last reviewed
- 2026-08-25
- Scope
- Public Cocode GUI, TUI, Host, Session, and model-entry workflows
- Verification
- Maintained against current page behavior, product boundaries, and repository-level checks
06FAQ
Before you start, make the boundaries clear.
What is Cocode?
Cocode is an independent distribution built on the DeepSeek Harness developer preview, with a desktop GUI, terminal TUI, and model-service options.
Does Cocode change files automatically?
Tool actions follow permission and confirmation flows. You can inspect the impact before writing files, running commands, or going online.
Do I have to use Cocode Nut?
No. You can bring a DeepSeek-compatible API key, use a self-hosted endpoint, or use Cocode Nut.
How is an AI coding agent different from a code chat tool?
The practical difference is whether it can read a real workspace, use tools, expose file changes, and run verification. Producing code text alone does not complete a repository task.
How do I know an agent task is actually complete?
Inspect the real diff, relevant tests or builds, command exit status, and remaining risks instead of relying on the model’s completion statement.