MCP server

Connect an AI agent

Kanbi exposes a Model Context Protocol server so coding agents — Codex, Claude Code, Grok Build, opencode, or another MCP client — can turn a finished discussion into reviewed Kanbi tasks from your editor or terminal. Auth is OAuth with a browser consent screen; there are no API keys to copy.

Claude Code

Add the server, then run any kanbi tool (or /mcp) — Claude opens the consent screen, and after you approve the tools are live.

claude mcp add --transport http kanbi http://localhost:3333/api/mcp

Codex

Add the remote server, then start the OAuth login. The same MCP setup is shared by the Codex CLI, desktop app, and IDE extension.

codex mcp add kanbi --url http://localhost:3333/api/mcp
codex mcp login kanbi

opencode

Add the server to opencode.json (project) or ~/.config/opencode/opencode.json (global):

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "kanbi": {
      "type": "remote",
      "url": "http://localhost:3333/api/mcp",
      "enabled": true
    }
  }
}

Then trigger the login (it also starts automatically the first time it hits the server):

opencode mcp auth kanbi

Grok Build

Add Kanbi as a remote server. Grok opens the OAuth flow when the server is first used.

grok mcp add --transport http kanbi http://localhost:3333/api/mcp

For Grok's hosted custom connector, use a publicly reachable HTTPS deployment, then add it at grok.com/connectors. A local development URL cannot be reached by the hosted service.

What an agent can do

Each tool wraps the same procedures the web app uses, so permissions, validation, and realtime updates apply identically. The agent acts as you, limited to the projects you can access.

• list_projectsreadResolve the target project when it is not unambiguous.
• get_project_contextreadGet canonical authoring IDs, guidance, and priorities.
• search_tasksreadSearch task titles for likely duplicates.
• create_taskswriteCreate only the exact task set the user approved.

Capture approved work

The connected agent writes the task content; Kanbi does not make a second model call. The agent searches for likely duplicates, shows the exact project, column, title, description, checklist, priority, labels, assignee, and due date, then waits for your explicit approval before it creates anything.

One useful task with a description and checklist is the default. Work is split only when each item can be delivered independently, and the full split is shown before approval. Task management after creation is not available over MCP.

Scopes

Two scopes are enforced on every tool call: kanbi:read for the read tools and kanbi:write for the write tools. A read-only grant cannot create anything. Add offline_access and the agent renews silently instead of reopening the browser.

Managing access

Every agent you authorize is listed under Profile → Connected apps with its scopes and when it was approved. Revoking takes effect immediately; the agent must re-authorize to reconnect.

Troubleshooting

Tools 401, or an endless re-auth loop. The token came back opaque instead of a JWT. Re-authorize with a clean token — for Claude Code, claude mcp remove kanbi then add it again.

Host mismatch. Point the agent at the same host this page is served from, so OAuth discovery, the token audience, and your browser session all line up on one origin.