Connect

By host

How you call a role depends on the host. Claude Code and VS Code turn each role into a command; Codex and OpenClaw read the role node instead.

Roles as MCP prompts

When a host supports MCP prompts, Vinc lists one prompt per role. The prompt name is the role id without concept:, so concept:agent-reviewer becomes agent-reviewer. Calling it returns the role's contract as the graph holds it right now, not a copy.

A prompt takes one optional argument, focus. Give it a node id and the nodes linked to that node come along; give it a phrase and the search results for it come along.

Only roles in the agent topic or one level below it (such as agent/release) are listed. Older spaces keep their roles under vinc/agent, which is read the same way. A role filed deeper does not show up.

Claude Code

Connect Vinc first (Connect over MCP). Each role then appears as a slash command. Type the command and, if you like, a focus after it.

/mcp__vinc__agent-reviewer concept:design-0

The connection name (vinc here) is the one you gave when adding the server.

VS Code

After connecting (Connect over MCP), type the prompt in chat as /mcp.vinc.agent-reviewer. VS Code asks for the focus in a box.

Codex

Codex does not support MCP prompts yet, so roles do not appear as commands there. Ask the assistant to read the role node and work from it. The contract is the same text either way.

Read concept:agent-reviewer from Vinc with vinc_graph and work as that role.

OpenClaw

Add Vinc under mcp.servers with the streamable-http transport written out (Connect over MCP). Without that line OpenClaw tries the older SSE transport, and the server is added but never connects.

When OpenClaw hands work to a sub-agent, that sub-agent's session may not have the Vinc tools even though the gateway is signed in. If a delegated run says it could not write to the graph, first check whether it had the tools at all, before signing in again.

When no roles show up