MCP servers

MCP (Model Context Protocol) is an open protocol for tool/context servers. Cerevisor is an MCP client; you can connect any MCP server and its tools become available to your agents alongside Cerevisor's built-in tools.

License gate: MCP is a Trial+/Paid feature.

Common MCP servers

A few off-the-shelf servers you can drop in:

Server	What it provides
@modelcontextprotocol/server-filesystem	Scoped filesystem read/write tools.
@modelcontextprotocol/server-github	GitHub issues, PRs, repo operations.
@modelcontextprotocol/server-postgres	SQL queries against a Postgres database.
@modelcontextprotocol/server-puppeteer	Web automation / scraping.
@modelcontextprotocol/server-slack	Slack messaging and search.

Many more on the MCP servers list.

Adding a server

Settings → MCP Servers → + Add MCP server.

The form fields:

Field	Required	Example
Name	Yes	filesystem
Command	Yes	npx
Args	No	-y @modelcontextprotocol/server-filesystem /Users/me/Documents
Working directory	No	Defaults to your workflow's project folder.
Environment variables	No	Key/value pairs passed to the spawned process.
Enable on connect	Toggle	If on, Cerevisor starts the server now and lists its tools.

Click Add. Cerevisor spawns the server, waits for the MCP handshake, then lists the tools the server exposes.

How tools are exposed

MCP tools are namespaced with the server name to avoid collisions:

• A read_file tool exposed by a server named filesystem becomes mcp__filesystem__read_file in the agent's tool catalog (mcp__<server>__<tool> is the canonical prefix; the mcp__ prefix is mandatory).
• A different server's read_file would be mcp__<other-server>__read_file.

Cerevisor's built-in read_file is unchanged. The namespacing keeps everything distinct.

Permission policy per server

Each MCP server has its own policy:

• Allow: agents can call any of this server's tools without prompting.
• Ask: every call prompts you.
• Deny: block entirely.

Set it per server on its row in the MCP Servers tab. Defaults to Ask.

The server-level policy is in addition to the global mcp permission category in Settings → Permissions.

Tool catalog

On each server row, click View tools to expand a list of every tool the server exposes, with its description. Useful for confirming the server connected correctly and for understanding what your agents can now do.

Run-time integration

When a workflow runs, Cerevisor:

1. Starts every enabled MCP server (or reuses already-running ones: the client is module-singleton, shared across runs).
2. Merges each server's tools into the agent tool registry.
3. Agents call MCP tools the same way they call built-ins.

The agent doesn't know or care whether a tool is built-in or MCP. From its perspective, the tools just exist.

When MCP doesn't work

Codex CLI provider: Codex has its own tool surface. Cerevisor's MCP tools are not injected into the Codex subprocess. Use built-in tools for Codex agents.

Cursor Agent provider: Cursor accepts an mcpServers config and you can pass it through, but Cerevisor's MCP client isn't shared with Cursor. Cloud-side MCP connectivity is your deploy problem.

Restart behavior

When you add or edit a server, Cerevisor starts/restarts it immediately. If the server fails to start (bad command, missing dependency), you'll see an error in its row with the failure message.

Servers that are running survive across Cerevisor's normal idle state. They're killed when Cerevisor quits.

Building your own MCP server

The MCP TypeScript SDK makes building a custom server straightforward, define your tools, declare their inputs/outputs, and the SDK handles the protocol.

Once built, add it to Cerevisor the same way you'd add any other MCP server: command + args + env.