MCP for AI agents integration
Last reviewed:
AI agents & API
AI agents do real work now — MCP makes them first-class actors in Kiantu, with their sessions, cost, and output attributed and budgeted like any teammate.
One MCP server turns Claude Code, Cursor, Codex CLI — any MCP client — into a tracked Kiantu worker with cost attribution.
One config line connects Claude Code, Cursor, or any MCP clientAgents report tokens + cost per model — spend shows up in reportsreport-only tokens: agents write their own record, read nothing else
What it does
- The `kiantu mcp` stdio server exposes Kiantu as MCP tools: sessions, token-usage reporting, activity logging, tasks, and the plate.
- Agents report tokens; Kiantu computes dollars from its server-side rate card and rolls cost up into sessions, goals, and reports.
- Registered agents are first-class workspace members: their sessions, activity, and spend all attribute to them.
Setup
- 1 In Kiantu, open Settings → Agents and click “Register agent” (admin/owner; Intelligence tier).
- 2 Choose the token scope: report_only (default — write its own work, read nothing) or workspace (needed for the task/plate tools).
- 3 Copy the kt_ token — it is shown exactly once.
- 4 Add the server to your agent runtime (see snippet). Tools appear immediately.
Tools
whoami confirm the actor + workspace behind the token start_session / close_session bracket a piece of work report_usage per-model token counts — Kiantu prices them log_activity discrete events (opened_pr, ran_tests, deployed) list_my_plate today’s committed work (workspace scope) get_task / update_task_status / create_task task workflow (workspace scope) Quickstart — Claude Code
claude mcp add kiantu -e KIANTU_API_TOKEN=kt_… -- kiantu mcpData & privacy
- report_only tokens cannot read teammates’ data — a leaked agent token cannot exfiltrate your workspace.
- Monthly budgets (workspace-wide and per-agent) alert admins on breach; ingestion never drops data.
- Every agent write is tagged at the database layer as agent-initiated.