MCP Tools

is an open standard for connecting AI agents to external systems — databases, APIs, filesystems, and services. Friday manages MCP server processes: declare a server in your space and every llm or user agent in that space can call its tools.

Declaring a server

workspace.yml

tools:
  mcp:
    servers:
      github:
        transport:
          type: stdio
          command: npx
          args: ["-y", "@modelcontextprotocol/server-github"]
        env:
          GITHUB_PERSONAL_ACCESS_TOKEN: from_environment
      time:
        transport:
          type: stdio
          command: uvx
          args: ["mcp-server-time", "--local-timezone", "UTC"]

Each key under servers is the server ID you’ll use to prefix tool names. The transport block tells Friday how to start the server. Use from_environment for credentials — the daemon injects them at process start.

Tool naming

Tools must be referenced as serverId/toolName in an agent’s tools array:

# Server ID is "github", tool name is "create_pull_request_review"
tools:
  - github/create_pull_request_review
  - github/get_pull_request

Run friday agent list -w <space> to see the exact tool names available from each configured server. Built-in platform tools (memory_save, memory_read) have no prefix and require no MCP server declaration — they’re always available:

tools:
  - memory_save
  - memory_read
  - github/search_issues

Using tools in an llm agent

workspace.yml

agents:
  pr-reviewer:
    type: llm
    description: "Reviews pull requests using GitHub tools"
    config:
      provider: anthropic
      model: claude-sonnet-4-6
      prompt: |
        Review the pull request and post a summary comment.
      tools:
        - github/create_pull_request_review
        - github/get_pull_request
        - github/get_pull_request_files

The LLM decides when and how to call tools based on the task and prompt.

Using tools in a custom Python agent

Python agents call tools via ctx.tools — tool names are bare (no prefix) when calling from code:

result = ctx.tools.call("get_pull_request", {"owner": "myorg", "repo": "myrepo", "pull_number": 42})

See the MCP tools guide for the full Python API.

atlas agents and MCP

type: atlas built-in agents are self-contained — they have their own built-in tool surfaces and ignore the tools array entirely. If you need a specific MCP tool called, use type: llm instead.

Common servers

Server	Install	What it provides
GitHub	`npx -y @modelcontextprotocol/server-github`	Repos, PRs, issues, commits
Postgres	`npx -y @modelcontextprotocol/server-postgres`	SQL queries on Postgres
Filesystem	`npx -y @modelcontextprotocol/server-filesystem`	Read/write local files
Time	`uvx mcp-server-time`	Current time, timezone conversion
Fetch	`uvx mcp-server-fetch`	HTTP fetch as a tool

Browse the full registry at github.com/modelcontextprotocol/servers.

Transport

stdio and streamable HTTP transports are supported. SSE transport is planned for a future release.

Getting started

Core concepts

Guides

Agent SDK

Reference

Resources

Declaring a server

Tool naming

Using tools in an llm agent

Using tools in a custom Python agent

atlas agents and MCP

Common servers

Transport

​Declaring a server

​Tool naming

​Using tools in an llm agent

​Using tools in a custom Python agent

​atlas agents and MCP

​Common servers

​Transport

Declaring a server

Tool naming

Using tools in an llm agent

Using tools in a custom Python agent

atlas agents and MCP

Common servers

Transport