Install

install — connect the existing graph engine

Connect the graph engine already shipped by this plugin. Do not copy, regenerate, or fork the engine: mcp/broker.mjs, mcp/graph.mjs, mcp/prompts.mjs, and the bundled Codex adapter remain plugin-owned and are updated with the graph plugin.

Install modes

Prefer the marketplace plugin. Its .mcp.json already registers the graph-engineering stdio server, so installation should not add project files. Ask the user to install or update graph@newkayak12-claude-skills, reload Claude Code if needed, then verify that graph_open, graph_next, graph_run, graph_submit, graph_retry, and graph_status are available.

Use a project-local connection only when the user explicitly wants to run from a source checkout instead of the marketplace plugin. Merge this entry into the target project's existing .mcp.json; preserve every unrelated server and use an absolute path:

{
  "mcpServers": {
    "graph-engineering": {
      "command": "node",
      "args": ["/absolute/path/to/graph/mcp/broker.mjs"]
    }
  }
}

Do not use ${CLAUDE_PLUGIN_ROOT} in a project-owned .mcp.json; that variable belongs to the plugin's own manifest. Do not register both marketplace and project-local copies, because two servers exposing the same graph_* tools make routing ambiguous.

Verification

1. Confirm node --version is 18 or newer.
2. Validate the selected server path exists when using project-local mode.
3. After Claude Code reloads the MCP configuration, confirm all six graph_* tools are present. Tool discovery is the install gate; do not open a real run just to test setup.
4. If a later run uses a named vendor, verify that vendor separately. The graph:orchestrate skill uses balanced allocation: reasoning prefers the driving host/model, Implement/Test prefer the other vendor's efficient model. It declares host_vendor, host_model, and supported native_models; external executors pass readiness checks. Fable/Astra require explicit model requests. Verify that role isolation is available before running work; an MCP connection alone does not provide new AI sessions. Both hosts use explicit task working directories and shared artifact paths supplied by the orchestration loop. Optional named vendor policies forward the selected model into readiness probes; a failed probe blocks the node. A bare direct graph_open({vendor: "auto"}) call still stays on self; a named vendor fails instead of silently degrading.

Report which mode is active, which files changed, whether a reload is still required, and the observed tool list. Installation ends there; use graph:orchestrate to run work.

Related

• orchestrate — drive a request through the connected graph
• harness:harness — the six-stage reasoning contract represented by the graph
• ../../.mcp.json — plugin-owned MCP registration