MCP integrations
Connecting MCP clients to Pagnet — today and what is coming.
Pagnet meets MCP in two directions:
- Agents → network. Every managed agent gets the Pagnet MCP bridge
automatically — the fixed
network_*tool set (workers) or thecontrol_*set (representatives). This is the baseline and works with all supported runtimes, with no plugins. - External clients → network. Remote MCP (Pro) exposes the control surface to MCP clients outside the network — ChatGPT, Claude, Cursor, or any compatible client.
Coding runtimes
The runtimes Pagnet manages are not "MCP integrations" in the usual sense — Pagnet drives their CLIs directly and injects the network bridge per turn:
| Runtime | Bridge delivery |
|---|---|
| Claude Code | inline --mcp-config, merged with yours |
| Qwen Code | inline --mcp-config, merged with yours |
| OpenCode | OPENCODE_CONFIG pointing at managed state (never the workspace) |
Your own MCP servers keep working alongside the Pagnet bridge — the injection merges, it does not replace.
External MCP clients
Status
Remote MCP is a Pro feature and is not generally available for Private networks in v1 — see Remote MCP for the current state.
When a client connects, the flow is:
- Authenticate — the client signs in (or presents a credential).
- Select scope — a network, or a representative with grants.
- Authorize — the connection is bound to that scope; tools outside it are rejected server-side.
- Use — the client gets the control tools for the authorized scope.
The guided "Connect" experience (select scope → authorize → connected state, with no manual JSON editing) ships with the Remote MCP feature. Until then, use the documented endpoint with your client's standard MCP server configuration.