Multi-agent systems

Multi-agent MCP: give each sub-agent its own tools, call limit and key

Your sub-agents share one workspace connection URL, and a looping one hits only its own limit while the rest keep working.

Try for free Quick tour video

One orchestrator, four sub-agents: a multi-agent MCP setup in Meandr

In this multi-agent MCP setup, an orchestrator builds a weekly sales report in Notion, and each sub-agent connects with its own key. Research reads the web with Fetch, billing reads Stripe, writer edits the Notion page and qa checks it in Playwright.

  1. 01 Tools per sub-agent

    Each sub-agent sees only its own tools: billing sees Stripe, and qa sees Playwright

    In Policies, add one allow rule per agent tag and set the default policy to Deny.

    • Research can read Notion pages, and only writer can edit them
    • Tools outside a sub-agent's rules never appear in its tool list
    • A new sub-agent with the same tag gets the same tools

    For example: Fetch, Notion, Stripe, Playwright

    Choose which tools each agent can call

  2. 02 Call limit per sub-agent

    A looping sub-agent can't flood Stripe past its own call limit

    The per-agent default in Settings → Rate limits gives every sub-agent its own count, and Edit agent can set a lower limit for billing.

    • Calls over the limit never reach Stripe
    • Research, writer and qa keep their own counts and keep working
    • An override can only tighten a limit, never loosen it

    For example: Stripe

    Real caseMarch 2026: an AI agent kept calling one failing tool on a design service until the account's limit ran out. The service locked the account out until the next day.

    MCP rate limiting per workspace, agent and server

  3. 03 Which sub-agent did it

    Every Notion edit shows which sub-agent made it

    Meandr records each call under the sub-agent's own name, so you can filter Traffic to writer and read its edits in order.

    • A denied call is recorded too, with the policy that stopped it

    For example: Notion

    Real caseJune 2026 survey of 107 enterprises: 69% run AI agents on shared credentials somewhere, and only 32% give every agent its own identity. VentureBeat research

    Trace one agent's calls in the audit log

  4. 04 Revoke one key

    Revoke the qa sub-agent's key, and research, billing and writer keep working

    Revoking a key in Keys stops Meandr accepting it, and every other sub-agent key stays valid.

    • The other sub-agents keep the same workspace connection URL
    • Rotate gives a sub-agent a new key if the old one leaks
    Rotate and revoke agent keys

Multi-agent system security: questions

Does a call limit on one sub-agent slow the others down?
Each agent has its own count, so the others keep working, and only the workspace and server limits cover all of them together.
Can I limit only the expensive tool?
Rate limits apply per workspace, agent or server, so for one tool add a policy that blocks it or holds it for your approval.
What happens to a call over the limit?
The sub-agent gets an error until the time window has room again, and the call never reaches the server.
Does revoking one sub-agent's key affect the others?
Every other key stays valid, so research, billing and writer carry on while Meandr no longer accepts the qa key.

Run multi-agent MCP on one URL, with a key for each sub-agent

Run any MCP server in Meandr or connect the ones you already have, then point every agent at one MCP gateway. Policy, approvals and audit cover every call, with no code changes.

Try for free

Questions? Email us