MCP for AI agents

Point your MCP host at Kabu and operate the platform from Claude Desktop, Cursor, Grok Build, and other MCP-compatible tools—without pasting broker secrets into the model.

What MCP is for Kabu

The Model Context Protocol (MCP) lets AI hosts call tools on a remote server. Kabu’s MCP product is a standalone HTTPS service—not a local monorepo process and not mounted on the dashboard.

Connect once to https://mcp.usekabu.app. Your host talks Streamable HTTP MCP; Kabu resolves identity server-side and runs the same operator tools available to the dashboard and CLI (experiments, runs, models, deployments, and existing broker accounts).

Remote product host: URL only — OAuth Connect in the browser
Same tool registry as the rest of Kabu—no separate agent API to learn
Broker credentials stay with you; agents only reference existing account IDs
MCP Apps UI — hosts that support interactive tools show a Kabu-branded view: tables, metric charts, symbol cards, and Back navigation (not only raw JSON)

Interactive results in the host

When your host supports MCP Apps (for example Grok Build, MCPJam, or other Apps-capable clients), every tool can open a shared Kabu view—not a separate dashboard tab.

Readable lists — human labels for brokers and related entities, short IDs, status badges, and dates
Click through — open a broker, experiment, run, model, or deployment from a row; use Back to return
Action chips — from a run, jump to metric keys and series; from an experiment, open its training runs
Symbol search — on a broker account, search tradable symbols (3+ characters), browse card results, fuzzy-filter and paginate what loaded. First search after idle can take up to a minute while the catalog service wakes—retry if the host times out
Text still works — hosts without Apps still get structured JSON for the model

Connect in minutes

Add the URL, sign in in the browser, pick a workspace. The host stores a short-lived token. No secrets in host config.

{
  "mcpServers": {
    "kabu": {
      "url": "https://mcp.usekabu.app"
    }
  }
}

The host discovers authorization metadata from the MCP origin, opens a browser login on Kabu, lets you pick (or create) a workspace, and then calls tools with the issued access token. Stay logged into Kabu in the browser for the Connect flow. Do not set an Authorization header—some hosts treat a pre-set Bearer as non-OAuth.

Revoke connected hosts under Account → Sessions & connections.

What agents can do

MCP tools are operational: discover context, manage experiments and runs, inspect models and metrics, and work with deployments. They are not a path to paste or store broker secrets.

Experiments — list, get, create, and validate configuration
Training runs — start, resume, cancel; list runs and metrics
Models — list and get artifacts from training
Deployments — list and inspect live deployments; start/stop when enabled for your environment
Brokers (read-only for secrets) — list accounts you already connected and resolve symbols with a search query (or specific IDs)—not create broker credentials in the model

Connect broker accounts yourself in the dashboard or CLI first. Agents then select an existing account ID for experiments and deployments.

Training · Live trading · Broker accounts

Safety and boundaries

MCP is an operator surface for agents—not a vault and not a place for monorepo install instructions.

Config is only the MCP URL — the host obtains the OAuth token itself
No database URLs, org IDs, or vault keys on the AI host
Broker secrets are never accepted as tool arguments
Product endpoint is https://mcp.usekabu.app—a separate service from the dashboard