MCP servers

What it is#

This is the outbound half of MCP: servers your workspace registers so its automations can call somebody else's tools. Each registration is discovered on save, its tools are listed with a switch each, and every call made on your behalf is recorded.

The inbound half — an MCP client such as an editor calling Projectri — has no screen here. You set that up from Connect IDE on Security.

How to get there#

Work → Integrations → MCP servers, or the MCP servers link from the integrations page. All integrations takes you back.

• Reading servers, tools and the call log needs mcpserver.view. It is withheld from the blanket viewer grant, because the call log holds the arguments sent on this workspace's behalf.
• Registering a server, storing a credential and enabling a tool need mcpserver.manage. Registering decides what our servers will connect to, which is why it is an administrator's key.
• Invoking an enabled tool needs mcpserver.call, kept separate because calling a third-party tool acts on your other systems under your credential.

The feature is gated on the platform.mcp_outbound entitlement (ENTERPRISE). Without it the page still loads and existing servers are read-only — the Register server button is disabled and a note says why.

How to use it#

Register a server#

1. Register server.
2. Name it — it has to be unique in this workspace.
3. Endpoint URL must start with http:// or https:// and must be reachable from the public internet. An address that resolves inside our network is refused, with the reason.
4. Choose the transport, then the authentication: no credential, a bearer token, or a custom header (which asks for the header name).
5. The credential is stored encrypted and never shown again.
6. Register and discover saves the row and performs the handshake in one call. They can disagree: the row is saved even when the handshake fails, and the message tells you which happened.

Choose which tools are callable#

1. Expand a server with Manage and stay on the Tools tab.
2. Each tool lists its description and its parameters, with required ones marked. Toggle the ones automations may call.
3. Save applies the changes; Discard returns to what was stored.
4. Rediscover re-reads the server's tool list — use it when the vendor adds or renames one.

Read what was sent#

The Log tab records every call: the tool, the caller, how long it took, and the request and response. The status is one of Succeeded, Failed, Blocked, Timed out or Replayed. Blocked means our own egress guard refused to dial that address — the third party is not at fault, and going to debug their server is the wrong move.

Pause or unregister a server#

1. The Settings tab has an Enabled switch. Turning it off stops automations calling the server without unregistering it.
2. Unregister server asks you to type the name. It removes the server and its tools; the call log is kept, because what was sent on a customer's behalf outlives the registration that sent it.

What it affects#

• Automations. An enabled tool becomes callable from a rule. Disabling the tool, disabling the server or unregistering it all stop those rules working.
• Your other systems. A call acts under the credential you stored, against the vendor you named. That is why calling is its own permission.
• Nothing inside Projectri reads these tools by itself. A registered server sits idle until something calls it.

Limits and gotchas#

• The circuit breaker cuts a server off after five consecutive failures. The health line distinguishes failing — still retrying, with the number of failures left before the cut-off — from cut off, which resumes only at the stated time. They need opposite actions, so they are never collapsed into one message.
• You cannot set a server to Active or Failed. Those are conclusions the discovery and call paths draw. The only statuses you may assert are Pending and Disabled.
• The credential is write-only throughout. Leave the box blank when editing to keep the stored one.
• A server registered but never discovered shows as *Registered, but its tools have never been discovered* — that is a normal state on the way to working, not a fault.
• The endpoint has to be publicly reachable. An internal address is refused at registration, which is the same guard that stops the console being turned into a request forger.

On mobile#

This screen is web only.

Related#

• Integrations — the connectors Projectri wrote.
• Automation — the rules that call these tools.
• Security — Connect IDE, the inbound half of MCP.
• MCP and the API — both directions explained.