Skip to main content
The Bland MCP server lets any AI agent work with your Bland account through natural language: place and inspect calls, send texts, build and validate pathways, manage agents, query analytics, run evals, and search the Bland docs. It’s a remote Model Context Protocol server hosted by Bland. There’s nothing to run locally: point your client at the endpoint, sign in with Bland or use your API key, and your agent has the full toolset.

​
Which path is yours?

Personal AI agent (Grok Bot, Meta Muse, OpenClaw, Instinct, Poke, or similar) asked to integrate with Bland? Go straight to Personal AI Agents. The rest of this page is for developers and coding agents.

I'm building voice agents from a coding agent

Install the Norm plugin (norm@bland) for Bland Agents. It bundles this server with Bland Agents skills and /norm:* commands on supported hosts. Other clients, or the server alone, are on the MCP clients page.

I want my personal AI agent to have a phone

Setup guides and compatibility requirements for Grok Bot, Meta Muse, OpenClaw, Instinct, and Poke. Connect through MCP or REST and get a dedicated number on the Agent Phone Plan.

​
How it works

  • Transport: Streamable HTTP (the current MCP standard). The server is stateless, so there are no sessions to manage and any request can hit any node.
  • Authentication: sign in with Bland (OAuth; ChatGPT, Claude, and Claude Code today, other clients once dynamic client registration is available), or your Bland API key in the Authorization header (Bearer org_...), which works with every client. Either way, one organization scopes every tool: an agent connected to your org sees only that org’s calls, pathways, and agents. Get a key from the dashboard, or let an agent obtain its own through the Agent onboarding API.
  • Rate limits: 120 requests per minute per organization.
  • Safety: every tool is annotated read-only or destructive, so well-behaved clients ask before the agent does anything with side effects (like placing a real phone call). Mutating tools additionally require an explicit confirmation argument.

​
Discovering the server from a 401

A request that reaches /v1/mcp with no credential gets a 401 that tells the client what to do next. Clients that already send a valid Authorization header never see it. Two pieces:
  • A WWW-Authenticate header (RFC 6750):
    A credential that was refused gets error="invalid_token" on the same header.
  • A JSON-RPC error body whose error.message and error.data name the header to send, where a human gets a key, and any self-serve signup flows an agent can run on its own: currently headless onboarding and the device-code flow. Each is listed only while it’s enabled.
Which piece a client uses depends on what it supports:
  • OAuth clients follow resource_metadata to the protected resource document (RFC 9728). It names the authorization server, https://api.bland.ai/authorization, and the scopes mcp:read and mcp:write. The authorization server’s own metadata is at https://api.bland.ai/.well-known/oauth-authorization-server/authorization and https://api.bland.ai/authorization/.well-known/openid-configuration. From there the client runs the flow on Sign in with Bland. Clients that only register dynamically can’t complete it yet and need an API key.
  • Agents with no key read the body, pick a signup flow, obtain their own key, and retry the same MCP call.

​
What your agent can do

See the full tool reference. The server also ships MCP prompts: build_pathway and edit_pathway appear as slash commands in clients that support them (like Claude Code) and load Bland’s pathway-authoring doctrine into the session.

​
One endpoint, three MCP surfaces

Bland has three MCP-related surfaces. If you’re unsure, use the first one.

​
Security notes

  • Treat your API key like a password. Use environment variables or your client’s secret storage. Never commit it to a repo.
  • A sign-in connection acts as you, in the organization you chose, with the permissions of your role there. Review the consent screen before you approve, and disconnect apps you no longer use.
  • OAuth tokens are only valid on /v1/mcp. They don’t work on the REST API.
  • Tools that spend money or touch live traffic (placing calls, publishing agents) are marked destructive and confirmation-gated, but review what your agent proposes before approving.
  • MCP tool results can contain untrusted content (e.g. call transcripts of real conversations). Prompt-injection-aware clients like Claude Code treat tool output as data, but keep it in mind when building your own agent loops.

​
Next steps

MCP clients

Connect the server directly to VS Code, Windsurf, Gemini CLI, or any client, without the plugin.

Agent Phone Plan

A dedicated number with unlimited US and Canada calling and texting, built for one agent.

Tool reference

Every tool the server exposes, grouped by what it does.

Send your first call

The 5-minute API quickstart.
Docs for agents: llms.txt