MCP servers

Some features on this page are new in the upcoming release and aren't in your current version yet.

Your agents come with built-in tools — reading files, running commands, fetching web pages. MCP servers let you add more: tools built by someone else, for a specific job. Connect one, and its tools show up for your agents right alongside Cerevisor's own.

Some MCP servers run on your computer (a program Cerevisor starts for you). Others connect online — they live on someone else's server, and you sign in once so Cerevisor can talk to them on your behalf. Either kind works the same way once it's connected.

License gate: MCP servers are a Trial+/Paid feature.

Adding a server

Settings → MCP Servers → Add server. You have three ways in:

• Popular — a one-click list of well-known servers: Filesystem, Memory, Web Fetch, Git, GitHub, Linear, Notion, Sentry, Slack, Browser (Playwright), and Brave Search. Pick one, fill in anything it asks for (a folder to share, a search key), and it's ready to connect. Servers marked Sign in open your browser to connect your account — see Signing in to an online server below.
• Paste — already have a server's setup instructions from somewhere else? Paste them in as-is: a plain URL, a command line, or the JSON block some tools hand you (including the format Claude Desktop uses). Cerevisor reads it and fills in the fields for you.
• Manual — fill in the fields yourself. Choose Runs on your computer (a command to launch, plus any arguments and environment variables it needs) or Connects online (just a web address).

Whichever door you use, you land on the same confirm step: give the server a short name, hit Test connection to make sure it actually works, then Save. New servers are enabled by default.

One thing to know: once a server is saved, you can't switch it between "runs on your computer" and "connects online" — if you set it up the wrong way, remove it and add it again the right way.

Signing in to an online server

Servers that connect online usually need you to sign in — the same way you'd sign in to any web account. Click Sign in on the server's row and it opens your normal web browser to that service's own sign-in page. You log in there, approve the connection, and the browser hands control back to Cerevisor. Cerevisor never sees your password — it only receives permission to act on your behalf, and only for what you approved.

If a sign-in fails or expires, the row shows Sign in again — nothing else changes; your other servers keep working.

Editing, testing, and removing a server

Click a server's row to edit it — same three-part form (name, connection details, test/save) as adding one. Test connection always checks the exact setup you're about to save, not a cached version, so what you test is what you get.

Remove asks you to confirm first — removing a server takes its tools away from every workflow that was using it, along with its saved sign-in.

What each server row tells you

Every saved server shows its name, whether it's on or off, and once connected, how many tools it offers. A Trust / Untrust toggle controls how carefully Cerevisor treats it:

• Untrusted (the default) — in Auto approval mode, the server's tools ask for your OK before running.
• Trusted — the server's tools run without asking, in Auto mode.

If a server fails to connect — a bad command, a missing dependency, an expired sign-in — the reason shows right on its row instead of failing silently.

Choosing which agents get which servers

Not every agent needs every server. Cerevisor gives you two layers of control:

Workflow-level, in a workflow's settings: choose All (every connected server is available by default), Selected (only the servers you check), or None (no server by default). Any agent can still make its own choice on top of this.

Per-agent, in an agent's settings: by default an agent inherits whatever the workflow decided. Click Customize to give that one agent its own list of servers — useful when one agent needs GitHub access and the rest of the workflow shouldn't touch it. Reset puts the agent back to inheriting from the workflow.

Before you run a workflow, the launcher shows a one-line summary of which servers the run will actually use and how many agents have their own custom list — so there's no guessing.

A note on quarantined-reader agents: if you've marked an agent as one that only reads outside content (a trust-role setting in the agent's config, aimed at content that might try to trick an agent), it never gets MCP servers, no matter what's assigned to it. That agent's job is to read things safely, not to act on them.

The permission switch

Settings → Permissions has an MCP tools switch for each workflow. Turn it off and none of a workflow's agents can use MCP servers at all, regardless of what's assigned to them — a hard stop for workflows you want to keep fully self-contained. It's on by default, matching how things worked before this switch existed.

What agents actually see

An MCP server's tools show up in an agent's toolbox labeled with the server's name, so if two servers both offer a "search" tool, your agent can tell them apart. Cerevisor's own built-in tools are unaffected — everything just sits together in one list.

When servers start and stop

Enabled servers start as soon as you open Cerevisor, and stay running across every workflow you run afterward — so only the first run pays the connection cost. Editing a server and saving it reconnects it with the new settings. Turning a server off, removing it, or quitting Cerevisor stops it.

Where MCP doesn't reach

Codex CLI and Claude Code agents run their own separate tool systems — Cerevisor's MCP servers aren't available to them. Use Cerevisor's regular agents for MCP-dependent work, or give those agents Cerevisor's built-in tools instead.

Building your own MCP server

If you're comfortable writing code, the MCP TypeScript SDK is a straightforward way to build a server: define your tools, describe what they take and return, and the SDK handles the rest. Once it's built, add it to Cerevisor the same way as any other server — through Manual, or by pasting its setup details in.