Local file context
Start from your repository and directories instead of copying an entire project into another chat window.
Checking sign-in status
01LOCAL WORKSPACE / CODING CLIENT 07
A local-first AI coding client for your workspace
Cocode is a local-first AI coding client: files, terminals, and task context stay around your workspace while you choose the model connection.
Capabilities and limits follow the current release
02Direct answer
Start with what it is.
Cocode is a local-first AI coding client. Workspace, terminal, and Git operations can stay on a developer-controlled machine, while inference, sign-in, updates, or external tools may still use the network. Local-first is not fully offline.
- Workspace location
- Local or developer-controlled remote machine
- Inference location
- Determined by the selected model endpoint
- Offline capability
- Local-first is not fully offline
- Data boundary
- Depends on models, tools, logs, gateways, and deployment configuration
Local execution flow
Commands, tests, diffs, and results stay centered on your machine or a remote workspace you control.
Replaceable model access
Use a DeepSeek-compatible API, self-hosted endpoint, gateway, or Cocode Nut without locking the client to one provider.
02Why it matters
Local-first does not mean isolated
Cocode puts local files and execution at the center while allowing a remote model service. You keep continuous repository context without giving up model choice.
Built for repositories, not isolated snippets
A task starts with the directory and dependencies, moves through edits, commands, and tests, then ends with a diff and conclusion in the same session.
↗
Clear boundaries make real development work easier to trust.
One task across desktop, terminal, and SSH
Use the GUI for structure and visual review, and the TUI for keyboard-first work on remote machines. Both connect to the same Host and Session.
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
Keep the workspace on your machine or host
The client reads and edits the target directory directly without importing the whole repository into a web editor.
- 02
Choose the model data path
Use Cocode Nut, a DeepSeek-compatible API, an enterprise gateway, or a self-hosted model endpoint.
- 03
Set tool permissions
Control file, command, and network access separately, with stricter confirmation for sensitive repositories.
- 04
Verify boundaries with logs and diffs
Inspect request paths, file changes, and command results instead of treating local-first as fully offline.
06Use cases
Decide by the work you actually need to finish.
Private repositories
Keep code in your workspace and send only task-relevant context to the selected model endpoint.
Restricted networks
Route model access through an approved enterprise gateway or internal endpoint.
Local-model experiments
Connect a compatible endpoint and verify protocol, tool-use, and quality boundaries.
Controlled remote hosts
Run tasks on your own development machine or server through TUI and SSH.
07Before you start
Runtime requirements and compatibility boundaries
- Local-first means the workspace and execution environment can remain under your control; it does not guarantee offline inference
- Prompts, code excerpts, and tool results sent to a remote model still leave the machine
- A self-hosted endpoint must match the protocol and model capabilities required by the client
- Sensitive projects need network policy, credential isolation, auditability, and least privilege together
08Try it
Start in a local workspace
Launch from the target repository after checking the model endpoint and network boundary.
cocode://workspaceREADY
cd /path/to/private-repository
cocode doctor
cocode09Facts and boundaries
Local must be separated into code, inference, and control planes.
A task meets privacy and network requirements only when the data path for every layer is explicit.
| Capability | Status | Conditions and evidence scope |
|---|---|---|
| Repository and Git | Can stay local | File and Git operations run on the machine that owns the workspace. |
| Model inference | Endpoint-dependent | Cloud APIs receive required requests; a local endpoint can keep inference in a controlled environment. |
| Sign-in and updates | May use network | Accounts, model services, downloads, and updates can require connectivity. |
| Fully offline | Requires dedicated setup | Models, tools, dependencies, and network policy must all be controlled; local-first alone does not prove it. |
10Cocode perspective
Local-first describes a workflow preference, not a privacy certificate.
Cocode treats code location, inference location, credentials, external tools, and logs as separate boundaries. A credible claim about code leaving the machine requires checking every layer.
- Maintained by
- Cocode Agency
- Last reviewed
- 2026-08-25
- Scope
- Local workspaces, inference, network dependencies, and data-path boundaries
- Verification
- Maintained against current client workflows, endpoint selection, and tool permission models
06FAQ
Before you start, make the boundaries clear.
Is Cocode a fully offline AI coding client?
Cocode should not be described as fully offline by default. It supports local files and execution, while model access can be remote or self-hosted depending on your configuration.
Can I use my own API key?
Yes. Cocode supports a DeepSeek-compatible API key, self-hosted endpoint, gateway, or Cocode Nut.
Does code leave my machine automatically?
Whether model requests leave the local environment depends on the model service and configuration you select. Review the endpoint, privacy notice, and organization policy before use.
Does local-first mean code never leaves my machine?
No. Whether code reaches a model or tool depends on the endpoint, request context, tools, and deployment configuration.
How can I get closer to fully offline AI coding?
Use a local model endpoint, local dependencies, controlled tools, blocked external network access, and offline-capable client components, then verify every layer makes no external call.