Cadre for VS Code

Cadre runs an organisation of AI agents — builders, reviewers, verifiers, deciders, managers — declared in one YAML file, on whichever free model APIs you have added. This extension puts Cadre inside the editor: start a run on the folder you have open, watch it work, answer its approvals, and review the branch it wrote.

It is a thin client. All routing, quota arithmetic, runs and approvals live in the Cadre engine (cadre serve, a local server on 127.0.0.1); the extension talks to its API and re-implements none of it. It is published to the VS Code Marketplace and to Open VSX, the registry Cursor, Windsurf and Antigravity install from.

What it does

• Runs view (the Cadre icon in the activity bar): recent runs with a status icon each — running, waiting for you, parked on a daily limit, succeeded, failed.
• Run view: a live timeline of the run's events (model calls, tool calls, checks, review rounds, votes), its result and the files it wrote. Updates as the run works.
• Status bar: today's usage against the tightest daily limit across your models — the largest of requests/RPD and tokens/TPD — and how many approvals are waiting.
• Approvals: a gate or an agent's question arrives as a notification (Approve / Reject, or Answer…). A request to execute code arrives as a notification with Review…, which opens a modal dialog; see Security below.
• Commands (Command Palette, category Cadre):
  • Start run on this folder — project mode on the workspace root: pick an organisation, type a goal. The run works on its own branch cadre/<run-id> in a git worktree; your working tree is not touched. Cadre refuses a folder with uncommitted changes and says why; you can then start from HEAD anyway.
  • Forecast — whether a goal fits in today's free quota, from measured history when there is some.
  • Review branch — the files a project run changed between its base commit and cadre/<run-id>, each opened as a diff.
  • Add provider — opens a terminal running cadre provider add <id>.
  • Open dashboard — the full web dashboard in your browser.
  • Review pending approvals, Cancel run, Resume run, Start server.

With no model key yet, a run can use Cadre's offline demo model, which exercises the whole workflow (not real intelligence) so you can see how it behaves.

Requirements

• uv on PATH (the extension then runs uvx cadre-ai serve), or cadre itself on PATH (uv tool install cadre-ai, or pip install cadre-ai).
• git, for project mode and Review branch.
• At least one free model key for real work (Groq and Google AI Studio both give one without a card). Add it with Cadre: Add provider.

When a Cadre view or command needs the server and none answers on the configured port, the extension starts cadre serve (or uvx cadre-ai serve) in the background, detached, so runs keep going after the editor closes. Its output goes to server.log in the extension's global storage folder. Nothing is started merely because the editor opened.

Settings

The server's home is CADRE_HOME (default ~/.cadre). If you run Cadre with a different CADRE_HOME, start VS Code with the same environment variable.

Security

• The API token. Every API call needs the bearer token that the server keeps in CADRE_HOME/token. The extension reads that file at the moment of each request, inside the extension host, and keeps it nowhere: not in settings (which Settings Sync would upload), not in VS Code's SecretStorage, not in logs, not in any message or webview. It is never displayed.
• Model keys never pass through the extension. Add provider runs the Cadre CLI in the integrated terminal, where you type the key into its hidden prompt; the CLI stores it in your operating system's credential store.
• Executing code. A run's checks (tests, linters) execute code the agents wrote, as you, on this machine — it is not a sandbox. The extension never starts a run with execution pre-approved. When a run first wants to run its checks, a notification says so; it never takes keyboard focus, so nothing typed into another window can answer it. Review… opens a modal dialog that shows each check's name and its exact command from the organisation file; only an explicit Allow execution click approves. Closing the dialog decides nothing (the run keeps waiting); Reject rejects.
• The run view is a webview with a strict Content Security Policy: one script, identified by a per-load nonce, from the extension itself; no remote resources; no network access. All text an agent wrote is inserted as text (textContent), never parsed as HTML. The webview receives run data by message from the extension and never sees the token.
• Notifications that quote an agent (a question, a gate) have Markdown link syntax broken up, so model text cannot become a clickable command link.
• Open dashboard runs cadre ui, which opens the browser with the token in the URL fragment — the part after # that browsers never send to a server — and the dashboard removes it from the address bar at once. The extension itself never handles that URL. Without the CLI on PATH it opens the plain address and the dashboard asks for the token.
• Untrusted folders. In Restricted Mode the extension does not start runs on the folder or run git in it; watching runs and answering approvals still work.

Troubleshooting

• "the server rejected the token" — the server is running with another CADRE_HOME than the one VS Code sees.
• The server did not start — open the server log from the error message. A port already in use is the usual cause; change cadre.port.
• A provider added while the server was running is not used — restart the server; it reads its providers when it starts.

Licence

Apache-2.0. Source: github.com/Daemon-VI/cadre, under editors/vscode.