Skip to content
Flow AI logo

Flow AI MCP server

Which MCP servers are Flow AI's own, from the official registry and from asking its own domain, whether they answer, and what they offer. A program checked each of these, nobody copied them from a product page.

read 2026-10-24

← Back to Flow AI

Whose server it is
  • Registry namecom.flowaiapi/agent-busverified in the MCP registry

The registry name sits under flowaiapi.com. The registry only lets a publisher use that namespace after proving control of the domain through DNS or HTTP.

When we asked itchecked 2026-10-24
  • Stateanswerschecked 2026-10-24
  • Tools25
  • Server calls itselfflow-ai
  • Server version1.0.0+acf3c95d483f57d6
  • Protocol version2025-06-18

On 2026-10-24 the server answered initialize and tools/list without a sign-in. First answer in 300 ms.

Should an agent connect itNot proven either way yet
  • Its tools do route agent model calls to the cheapest model that finishes tasks, but not get web-grounded answers with normalized source citations.
  • Checked on 3 of the 7 days it takes to judge reliability.

Once connected

  • bus_send, bus_cancel change data without a sign-in: anyone with the address can call them.
  • Parameters are untyped or undescribed in 10 of 25 tools: check arguments before calling.

Said by rules from the measurements above. The likely first call is a model's pick from the tool list; read its description before relying on it.

Measured dailydaily checks: 3, the first on 2026-10-24
  • INITE MCP scoreafter a weeka score needs 7 daily checks: 3 so far
  • What an agent can do with it1 of 2 use cases
  • Answeredmeasuring3 of the 7 daily checks it takes to print a share
  • Median first answer363 ms
  • Space in an agent's context≈ 5,016 tokensthe definitions of all its tools (25), which an agent carries on every turn
  • Tool definitionsA · 93 of 100
  • Typed results13 of 25 toolsdeclared outputSchema: results come typed, read once
  • Parameters without a description: in 10 of 25

What an agent can do with it

  • can do · route agent model calls to the cheapest model that finishes tasks
  • not sure · run one prompt across several models in parallel
  • cannot · get web-grounded answers with normalized source citations

Each of the product's use cases, put to Jev against this server's tool list: can an agent do it with these tools?

Worth a look

  • Tools that change data, on a server that answers without a sign-in: bus_send, bus_cancel

A program checks the server once a day: initialize and tools/list, no model. Tokens are the definitions' length at about four characters a token. The grade averages six checks of every tool's name, description, schema and annotations; it says how well the tools are described, not whether they work. The score weighs what was measured and nothing else: answering (25%), doing the product's job (25%), definitions (20%), context (10%), use (10%) and how readily a client connects (10%). A part that was not measured is left out, and a hidden instruction caps the score at 40. The state of vendors' MCP servers: method and numbers →

Tools it offers25 tools · read 2026-10-24
  • search_models

    Search Flow AI's live model catalog by name/provider/family. Returns id, provider, context window, per-token prices and verified capabilities for up to 20 matches.

  • get_live_prices

    Flow AI's live market book: clearing prices vs published list prices per model, with savings percentage. Sorted by savings. These are the prices requests actually clear at.

  • list_free_models

    The canary-verified FREE models currently promoted into Flow AI's default routing pool. Each passed a live tool-calling + output-quality canary; they serve at $0 per token.

  • about_flow_ai

    What Flow AI is, how routing and pass-through pricing work, and how to point an agent or SDK at it (base URLs for OpenAI, Anthropic, and Codex protocols).

  • convene_council

    Convene a council of DIVERSE AI models to critique a proposal or decision you are uncertain about. Each model independently assesses it, then a synthesis merges agreements, disagreements, and a recommendation. Requires your Flow AI API…

  • delegate_task

    Hand a self-contained subtask to the cheapest model that can complete it — offload work that doesn't need your own (expensive) model, to save tokens and move faster. Returns the result, which model served it, and the exact cost. Requires…

  • bus_agents

    Flow Agent Bus: directory of your agents — address, harness, machine, status. Addresses look like agent://project/name.

  • bus_send

    Flow Agent Bus: send a durable message to another agent on your account (agent://team/name). The recipient is woken by its doorbell/webhook if it has one; otherwise the message waits — the response's recipient_status/delivery_note say…

  • bus_inbox

    Flow Agent Bus: claim the next message addressed to you as a LEASE (at most one at a time, strict FIFO). Settle it with bus_reply (or bus_ack) before the next is offered; if you crash, the lease expires and the message is re-offered. Pass…

  • bus_ack

    Flow Agent Bus: acknowledge durable receipt of a leased message WITHOUT replying (pass message_id + lease_id, read from the bus_inbox response as messages[0].message_id and messages[0].lease.lease_id). CAUTION: if the message expects a…

  • bus_nack

    Flow Agent Bus: return a leased message you could NOT process (pass message_id + lease_id and a short error). retryable:true re-offers it on the next poll (four failures dead-letter it); retryable:false dead-letters it now. Never nack…

  • bus_cancel

    Flow Agent Bus: withdraw a message YOU sent that is still queued (not yet leased). Returns its final status; a message already leased or replied cannot be cancelled.

  • bus_mint

    Flow Agent Bus: create bus keys for NEW teammates. With a bus-scope key you may mint only under your own project (agent://<yourproject>/<name>); an account-owner (full-scope) key may mint into any project on the account. Up to 8 per call…

  • bus_manage

    Flow Agent Bus: manage your account's agents — the owner/self-service surface. actions: remove_agent (address) | set_accept_from (address, accept_from[]) | set_harness (address, harness) | set_notify (address or '*', webhook_url…

  • bus_health

    Flow Agent Bus: your ACCOUNT's health — median reply latency (are agents listening?), who is listening now, push coverage, stuck mailboxes (mail nobody is collecting), volume. The diagnostic an orchestrator needs before blaming a peer.

  • bus_configure

    Flow Agent Bus: change your OWN mailbox settings WITHOUT claiming a message — accept_from, harness, webhook_url/webhook_format/webhook_headers (or "" to clear). Same fields as bus_inbox's settings, but this tool can never take a lease, so…

  • bus_renew

    Flow Agent Bus: extend the lease on a message you are still working on (pass message_id + lease_id). Each renewal adds the standard lease (15 min); total lease time is capped at 6h from first claim. Use it for hour-long work instead of…

  • bus_reply

    Flow Agent Bus: reply to a message you received via bus_inbox (creates a first-class reply message and implies acknowledgment). Pass the lease_id (messages[0].lease.lease_id in the bus_inbox response) to fence a stale worker.

  • bus_check

    Flow Agent Bus: status of a message you sent — queued | leased | delivered | replied | dead_letter | expired | cancelled — with the reply once it exists. terminal:true means stop waiting; reply_declined:true means the recipient closed it…

  • bus_me

    Flow Agent Bus: inspect your OWN mailbox — who may message you (accept_from), unread queue depth, the lease you are holding (and whether it blocks your mailbox), webhook state. Call this FIRST when the bus seems stuck.

  • bus_signup

    Flow Agent Bus: create a NEW, ISOLATED team account (free, no registration). FIRST agent of a team only. If your user already has agents on the bus, use bus_join (with a code from a teammate's bus_invite) — a second signup cannot see the…

  • bus_join

    Flow Agent Bus: join an EXISTING team with an invite code (no key needed). A teammate made the code with bus_invite; you get your own bound key on their account, shown once. Then bus_configure {harness, accept_from:["*"]} to activate.

  • bus_invite

    Flow Agent Bus: mint a one-time INVITE CODE so a new teammate can join your team by itself (bus_join) — a code is not a key, so a human can relay it safely. A bus key invites into its own project; an owner key names the project. Default 1…

  • bus_history

    Flow Agent Bus: your past mail — sent or received (owner keys: the whole account), bodies within the 30-day retention, dead letters included. Filter by peer (with), thread (trace_id), status; page with before = the previous page's…

  • bus_replay

    Flow Agent Bus: re-queue a dead-lettered or expired message addressed to you (owner keys: any). Attempts reset, a fresh 24h lifetime applies, the doorbell rings. Find candidates with bus_history {status:"dead_letter"}.

How to connect

Add the server to your MCP client config. Claude Desktop, Cursor and most other clients read this shape:

{
  "mcpServers": {
    "agent-bus": {
      "type": "http",
      "url": "https://api.flowaiapi.com/mcp"
    }
  }
}

Add it

Claude Code, in a terminal

claude mcp add --transport http flowaiapi https://api.flowaiapi.com/mcp

Add to CursorAdd to VS Code

claude.ai, ChatGPT and other clients: add https://api.flowaiapi.com/mcp as a custom connector in the settings.

Each line is something a program found when it went looking: the registry entry, the server's own answer, a spec that parses. A missing line means we did not find it, not that it does not exist.