Connectors
Named MCP server connections that give agents access to external tools and data sources.
A KomputerConnector is a named MCP (Model Context Protocol) server connection. Connectors give agents access to external tools and data sources — GitHub repositories, Slack channels, Linear issues, and any service that exposes an MCP endpoint.
When to use it
- External integrations — Let agents read and write to services like GitHub, Slack, or Linear without writing custom tools
- Custom MCP servers — Point to any MCP-compatible endpoint, self-hosted or remote
- Shared credentials — One connector definition can be attached to many agents; credentials are stored once as a K8s Secret
How it works
- Create a
KomputerConnectorCR with a URL and an optional auth secret reference - The UI can auto-create the K8s Secret from a token you paste in — you never handle the secret directly
- Reference the connector by name in
spec.connectorson aKomputerAgent - When the agent pod starts, the operator injects the MCP server config as
KOMPUTER_MCP_SERVERSand mounts the auth token as aCONNECTOR_<NAME>_TOKENenv var - The agent runtime configures the Claude SDK with the MCP server, making all its tools available as
mcp__<name>__*slash commands - If you attach or remove a connector from a running agent via PATCH, the change takes effect on the next task — no pod restart needed
- The
.status.attachedAgentsfield tracks how many agents reference each connector
Example
apiVersion: komputer.komputer.ai/v1alpha1
kind: KomputerConnector
metadata:
name: github
namespace: default
spec:
service: github
url: "https://api.githubcopilot.com/mcp/"
authSecretKeyRef:
name: github-credentials
key: tokenAttach to an agent:
spec:
connectors:
- githubThe agent can then use tools like mcp__github__create_pull_request, mcp__github__search_code, etc.
Compatibility
Which third-party services work as connectors today, which need OAuth, and which require running your own MCP server.
Supported out of the box
These services work with a static token. Paste it into the connector dialog and you're done.
| Service | MCP URL | Token type |
|---|---|---|
| GitHub | https://api.githubcopilot.com/mcp/ | Personal Access Token (ghp_) |
| Gmail | https://mcp.google.com/a/gmail/mcp | Google OAuth Access Token (ya29.) |
| Google Calendar | https://mcp.google.com/a/calendar/mcp | Google OAuth Access Token (ya29.) |
| Linear | https://mcp.linear.app/mcp | API Key (lin_api_) |
| Slack | https://mcp.slack.com/mcp | User OAuth Token (xoxp-) |
Note: Gmail and Google Calendar tokens are OAuth access tokens, but you can grab one manually via the OAuth Playground — no full OAuth flow needed in the komputer.ai UI.
Requires a full OAuth flow
These services reject static tokens — you need to run them through OAuth 2.0 to get a usable access token. The current UI doesn't support this yet.
| Service | MCP URL | Why static tokens don't work |
|---|---|---|
| Notion | https://mcp.notion.com/mcp | Integration tokens (ntn_) are rejected with invalid_token — only OAuth 2.0 access tokens are accepted |
| Atlassian (Rovo) | https://mcp.atlassian.com/v1/mcp | PAT tokens only expose 2 limited tools (getTeamworkGraphContext, getTeamworkGraphObject) — full Jira/Confluence access requires OAuth |
Requires a self-hosted MCP server
These services don't expose a remote MCP endpoint that fits our model. To use them, deploy an MCP server inside your cluster and point a custom connector at it.
| Service | Recommended server | Notes |
|---|---|---|
| Atlassian (full Jira + Confluence) | sooperset/mcp-atlassian | Exposes ~72 Jira + Confluence tools via PAT — deploy in cluster, point custom connector at pod URL |
| Notion (full) | modelcontextprotocol/servers/notion | Accepts ntn_ integration tokens — deploy in cluster |
Adding your own
Any MCP-compatible endpoint can be added as a custom connector — just create a KomputerConnector pointing at the URL and (optionally) referencing an auth secret. See the example above.

