Overview
MCP & Plugins
Build and manage your Relevance AI agents programmatically using Claude Code or any MCP-compatible AI client.
Connect your AI coding environment to Relevance AI to build agents, tools, and workforces directly from your terminal — using natural language. Instead of clicking through a UI, you build, test, and iterate through clients like Claude Code, Cursor, or any MCP-compatible AI client.
Getting started
Pick the setup that matches your AI client.
OpenAI Codex
Connect Codex to Relevance AI with the MCP server and agent skills.
Claude Code plugin
The fastest way to get started. Install the Relevance AI plugin for Claude Code and build agents from your terminal.
MCP server
Connect from any MCP-compatible client — Claude Desktop, Cursor, VS Code, ChatGPT, and more.
Agent skills
Clone the agent skills repository to give your AI coding assistant built-in knowledge of Relevance AI.
This page focuses on what builders can create and manage via MCP. MCP connections also work for team members who only need to run existing agents — not build or edit them. Users with Viewer or Chat project roles are automatically placed into a restricted access mode when they authenticate, so they can view and run agents without any write access. See Access control for details.
What you can do
Once connected, your AI client gets full access to your Relevance AI project. This goes far beyond running existing tools — you can build and manage everything in your project from clients like Claude Code.
Create Agents
Design and configure new Agents, set their instructions, assign Tools, and configure Triggers.
Build Tools
Create new Tools with custom steps, inputs, and outputs.
Set up Workforces
Build multi-agent workflows with Triggers, conditions, and agent-to-agent handoffs.
Trigger Agents
Start conversations with your Agents and get responses.
Execute Tools
Run any of your Relevance AI Tools directly from your AI client.
Troubleshoot Agents
Diagnose issues with your Agents by reviewing conversation logs and tool outputs.
Refine Agents
Iterate on Agent instructions, Tool configurations, and behavior based on real results.
Evaluate runs
Review previous agent runs, identify failures, and improve performance over time.
Update configurations
Modify agent instructions, tool settings, and workflow logic.
Use cases
Create and configure agents end-to-end from your AI client. Describe what you want in natural language and let your AI client handle the setup.
Example prompts:
Customer support agent
“Create a new agent called ‘Customer Support Bot’ that answers questions using our FAQ knowledge base. Give it a friendly tone and make sure it escalates to a human when it can’t answer.”
Lead qualification agent
“Build me an agent that qualifies inbound leads from HubSpot. It should check the company size and industry, then send a personalized follow-up email via Gmail.”
Slack triage agent
“Set up an agent that monitors our Slack support channel, categorizes messages by urgency, and assigns them to the right team member.”
Scheduled reporting agent
“Create an agent with a scheduled trigger that runs every morning, pulls yesterday’s data from Google Sheets, and posts a summary to Slack.”
Review how your agents have been performing by looking at previous conversation runs, identifying where they succeeded or failed, and making targeted improvements.
Example prompts:
Identify failures
“Pull the last 20 conversations for my Support Agent. Identify any where the agent gave an incorrect answer or failed to resolve the issue.”
Measure qualification rates
“Look at my Sales Agent’s recent runs. How often is it successfully qualifying leads vs. letting unqualified ones through?”
Find common struggles
“Review the last week of conversations for my Onboarding Agent. Are there any common questions it struggles with? Suggest improvements to its instructions.”
Compare before & after
“Compare the performance of my Sales Agent before and after I updated its prompt last Tuesday. Is it doing better at objection handling?”
Create custom tools that your agents can use, combining API calls, code steps, LLM processing, and integrations — all from your AI client.
Example prompts:
Company research tool
“Create a tool that takes a company URL, scrapes the homepage, and returns a one-paragraph summary of what the company does.”
Knowledge search tool
“Build a tool that searches our knowledge base for the top 3 most relevant articles given a customer question, and formats them as a numbered list with links.”
Data enrichment tool
“Make a tool that takes a CSV of records, enriches each one with LinkedIn data, and outputs a Google Sheet with the results.”
Email generator
“Create a tool that generates a personalized email based on a contact’s LinkedIn profile and our product’s value props.”
When something isn’t working right, dig into agent behavior, tool failures, and configuration problems.
Example prompts:
Debug triggers
“My Support Agent stopped responding to Slack messages yesterday. Check its trigger configuration and recent conversation logs to figure out what happened.”
Fix tool errors
“The data enrichment tool is returning empty results. Look at the tool steps and check if the API call is configured correctly.”
Fix hallucination
“My agent keeps hallucinating answers instead of using the knowledge base. Review its instructions and knowledge configuration and suggest fixes.”
Audit configurations
“List all my agents and their triggers. I think one of them has a broken webhook — find it and show me the configuration.”
Best practices
Plan before you build
Before asking your AI client to create or modify anything, start by having it plan the work first. In Claude Code, you can type /plan to enter plan mode — this lets you and Claude align on the approach before any changes are made.
Review before you approve
When your AI client proposes changes — like updating an agent’s instructions or modifying a tool — read through what it’s about to do before confirming. This is especially important for agents that are already live and handling real conversations.
Use conversation history for context
When troubleshooting or refining an agent, ask your AI client to pull recent conversation logs first. This gives it real context to work with rather than guessing.
Test with real scenarios
After building or updating an agent, trigger a test conversation to see how it actually behaves. Don’t just review the configuration — run it. Ask your AI client to “Send a test message to my Support Agent asking about refund policies” and review the response.
Work across multiple projects deliberately
If you have separate projects for development and production, connect to both via separate MCP entries. Build and test in your dev project, then once you’re happy, recreate or promote the agent in production. This keeps your live agents safe while you experiment.
Frequently asked questions (FAQs)
What is MCP?
The Model Context Protocol (MCP) is an open standard that allows AI clients to connect to external tools and data sources. It provides a standardized way for AI assistants to access your Relevance AI workspace, so you can build and manage agents from clients like Claude Code, Cursor, or VS Code instead of the Relevance AI web interface.
Do I need to use Claude Code?
No. Claude Code with the Relevance AI plugin provides the richest experience, but you can use any MCP-compatible client — Claude Desktop, ChatGPT, Cursor, VS Code, Windsurf, and more. See the MCP server page for all supported clients.
Is it free to use?
The MCP server and Claude Code plugin are free. You will be billed for any Relevance AI usage (agent runs, tool executions, etc.) according to your plan.
Can I use multiple AI clients at the same time?
Yes. You can connect to the Relevance AI MCP server from as many clients as you like simultaneously. Each client authenticates independently.
Does authentication expire?
Authentication tokens may expire after a period of inactivity. If you are prompted to re-authenticate, simply follow the login flow again.
Can I restrict which tools are available?
Run-Only mode is the primary way to restrict MCP access. Assign a user the Viewer or Chat project role to force them into a restricted mode automatically — Viewers get run-only access (no write or destructive operations), and Chat users are limited to an even narrower set of tools. Member, Editor, and Admin users can opt into run-only mode from the toggle on the OAuth consent screen. For finer-grained control — such as exposing only a specific subset of agents — organize your tools across separate projects and authenticate each connection to the appropriate project. See Access control on the MCP server page for the full role matrix.