# MCP Servers

MCP servers (Model Context Protocol) give the agent its tools: driving the browser, exploring your code, working with your Git host, or anything else you connect. Settings → MCP Servers lists them with their current state.

# The list

Column	What it shows
Server	The tool's name, as the agent sees it
Connection	Where a remote server is reached
State	See below
Source	Managed by deployment for tools that come with the workspace, User for tools you added
Actions	Connect or Disconnect, and for your own tools, edit and remove

State	Meaning
Connected	The agent can use the tool.
Starting	The tool is being started.
Not started	The tool has not been started. Click Connect.
Disabled	The tool is switched off, or waits for something it needs. The reason is shown underneath.
Needs authentication	The tool needs you to sign in or provide a key. See Credentials → MCP authorizations.
Failed	The tool could not start or connect.

# Tools every workspace has

Tool	What it does
`browser`	Drives the Browser tab's web browser.
`codegraph`	Answers structural questions about your code from the CodeGraph index of each project. See Projects.
`forgejo`	Works with repositories, issues and pull requests on the PHOENIQS Git service, `git.kvant.cloud`. It stays Disabled until you store a Git credential for that host under Credentials.

Tools that come with the deployment cannot be changed here. To use your own key or account for one of them, see Credentials → MCP authorizations.

# Adding a tool

Click Add server.
Enter a Name: lowercase letters, digits, dots and dashes.
Choose the Type:
- Local, a program started inside your workspace. Enter its Command, and optionally a Working directory, an Environment (one KEY=value per line) and a Timeout in milliseconds.
- Remote, a server reached over the network. Enter its Server URL. The connection uses Streamable HTTP, with SSE as fallback.
For a remote server, choose its Authentication:
- No auth
- API key header: the header and the key or token to send
- OAuth: optionally a client id, client secret and scope. Without a client id, the server registers itself. There is no redirect URI to set up: the workspace receives the sign-in callback itself.
Leave Enabled on start ticked to connect the tool whenever the workspace starts.

Changes to MCP servers apply immediately. Your tools are stored in your own settings and kept across restarts.

# Whose account a tool uses

When you sign a tool in, the authorization is stored in the workspace, and every session uses it. Because each workspace has a single owner, that is your account. Credentials → MCP authorizations shows which identity each tool acts as.

Credentials

../credentials/

Terminal

../../terminal/

ai coding phoeniqs ai build settings mcp tools