MCP for coding agents

Operational context without mutation

InferCrane includes an MCP server over standard input/output:

infercrane mcp

Configure your MCP-compatible client to launch that command with the same InferCrane context and credential it would use for read-only CLI inspection.

{
  "mcpServers": {
    "infercrane": {
      "command": "infercrane",
      "args": ["mcp"],
      "env": {
        "INFERCRANE_CONTEXT": "production"
      }
    }
  }
}

The server exposes six bounded tools:

Tool	Evidence
`infercrane_list_deployments`	Logical deployments and lifecycle state
`infercrane_inspect_deployment`	Desired/observed state, revisions, replicas, operations, measurements, and Guard evidence
`infercrane_inspect_endpoint`	Logical model, environment, bindings, active plan, and candidate plan
`infercrane_inspect_request`	Content-free routing, latency, tokens, retries, and fallback evidence
`infercrane_inspect_operation`	Durable operation state and resume identity
`infercrane_list_curated_recipes`	Reviewed, commit-pinned configuration recipes

All tools declare read-only, non-destructive, closed-world annotations. The server has no deploy, scale, promote, delete, budget, secret, or provider mutation tool.

Read-only access can still reveal operational metadata. Use a least-privilege InferCrane principal, scope it to the appropriate tenant, and do not place credentials in the client configuration when the client supports secret references or environment injection.

Mintlify also publishes the navigable documentation as llms.txt and llms-full.txt for agents that need product knowledge rather than live operational evidence.

Start here

Gateways and external systems

SDKs and automation

Providers

Runtimes and evidence tools

MCP for coding agents

Operational context without mutation

​Operational context without mutation

Operational context without mutation