Skip to page content
Make Agent Fast
  • Introduction
  • Quickstart
  • Core concepts
  • Dashboard tour
  • Build your agent
  • Add it to your website
  • Telegram: bots & groups
  • WhatsApp
  • Instagram: DMs & comments
  • Your inbox
  • Leads & sales pipeline
  • Analytics, team & billing
  • Messenger, Discord & KakaoTalk
  • Connect your AI assistant
  • Create an agent
  • Persona and instructions
  • Knowledge sources
  • FAQs and products
  • Lookups
  • Change actions
  • Booking with Cal.com
  • Google Calendar
  • Follow-ups and reminders
  • Page actions
  • Your own tools (MCP)
  • Voice and live calls
  • Display
  • Agent link
  • Custom domains
  • Embed overview
  • HTML and JavaScript
  • React and Next.js
  • Vue and Nuxt
  • SvelteKit
  • Astro
  • WordPress
  • Shopify
  • Webflow
  • Framer
  • Wix
  • Squarespace
  • Ghost
  • Notion and Linktree
  • Tilda
  • Imweb
  • Cafe24
  • Security and CSP
  • Connector overview
  • Telegram
  • WhatsApp
  • Messenger
  • Instagram
  • Discord
  • KakaoTalk
  • Conversations and leads
  • Analytics
  • Broadcasts
  • Notifications
  • Monetization
  • Use Claude and ChatGPT
  • Credits and usage
  • API overview
  • Credentials and tokens
  • Authentication
  • Scopes and errors
  • API quickstart
  • API recipes
  • API reference
  • Webhooks
  • SDK status
  • Rate limits
  • Plans and pricing
  • Security
  • AI provider keys
  • Privacy and data
  • Troubleshooting
  • Changelog
Docs
Log inAPI
Make Agent Fast documentation

Your own tools (MCP)

Connect a remote MCP server and choose which tools customers may use through your agent.

At a glance

Connect a remote MCP server and choose which tools customers may use through your agent.

Confirmations and approvalsCustomer identity

Connect and discover#

Open Agent → Actions → Your tools (MCP). Connect your remote server using its OAuth flow or bearer credential, or use an existing MCP connection. Choose Discover to read the tool catalog. Review each tool's description, parameters and annotations before enabling it. Discovery does not execute a tool; newly discovered tools are disabled.

You can configure three tool servers and enable 20 tools in total. Credentials remain server-side. Removing a tool mapping stops that mapping, but does not revoke the shared knowledge connection or its OAuth grant; manage that connection separately. Make Agent Fast's own owner-management MCP server cannot be exposed as a customer tool server. To manage your agents from Claude, see Use Claude and ChatGPT.

Choose read or write#

EffectBehavior
ReadCalls the tool directly and treats returned text/structured data as untrusted business data
WriteCreates a proposal; the customer must confirm the exact inputs with the card's buttons before execution

Read is available only if the server declares readOnlyHint: true and does not declare destructiveHint: true, and you keep the tool classified as read. Otherwise treat it as write. Review what the server actually does; annotations are not an authorization system.

Writes default to owner approval every time. You can choose no additional owner approval or approval above a numeric parameter limit. Customer confirmation is still required in every mode. Owner approval uses the existing Approvals list and receipts. Settings autosave. Owner rehearsal never calls the remote server.

Bind the signed-in customer#

Tools require a verified customer by default. Use signed customer identity on your website; an ID typed in chat is not verification. For customer-specific records, choose the tool's string parameter that identifies the customer. The server fills it from the verified customer ID and removes it from the model's inputs. Enum parameters cannot be used for this binding.

For example, bind customer_id on a tool that reads an order. The model can supply the declared order reference, but cannot substitute another customer ID. The remote server must still check that the order belongs to that customer. Requiring sign-in without a parameter binding does not by itself scope every record the remote tool can access.

Limits and uncertain results#

Supported schemas are flat objects with up to ten string, number, integer, boolean or enum parameters. Nested objects, arrays and unsupported schema rules stay unavailable. Strings are bounded to 200 characters; remote calls have a 10-second timeout and model-visible results are capped at 4 KB. Result content is data, never instructions to the agent.

MCP provides no general idempotency guarantee. A write is called once after confirmation/approval. If its outcome is unknown, Make Agent Fast does not retry automatically and notifies the owner. Check the remote system before creating another proposal. Changing the reviewed tool configuration invalidates an older proposal.

Test each intended read and write with your own server before customers use it.

PreviousPage actionsRegister and approve website actions for the widget.NextVoice and live callsConfigure voices, voice messages, cloning, and live calls.
Was this page helpful?
On this page
  • Connect and discover
  • Choose read or write
  • Bind the signed-in customer
  • Limits and uncertain results
For AI systemsllms.txtllms-full.txt