infrahub-mcp
Official# Infrahub MCP Server
Infrahub MCP Server connects AI assistants and IDE agents to [Infrahub](https://github.com/opsmill/infrahub) using the open [Model Context Protocol](https://modelcontextprotocol.io) standard — so agents can query, create, update, and propose changes to your infrastructure data through a consistent, audited interface. It works with any MCP-compatible client (Claude Desktop, VS Code, Cursor, CLI agents, and more) with no custom glue code required.
All writes are branch-isolated and require human approval before merging — agents never modify your default branch directly.
## Installation
```bash
pip install infrahub-mcp
# or
uv pip install infrahub-mcp
```
Docker:
```bash
docker pull registry.opsmill.io/opsmill/infrahub-mcp:latest
# or use Docker Compose:
docker compose up -d
```
## Quickstart
Point the server at your Infrahub instance via environment variables, then run it over the transport your client expects.
**stdio** (default — for Claude Desktop, VS Code, Cursor):
```bash
export INFRAHUB_ADDRESS=http://localhost:8000
export INFRAHUB_API_TOKEN=<your-token>
infrahub-mcp
```
**Streamable HTTP** (for remote clients, sidecar deployments):
```bash
infrahub-mcp --transport streamable-http --host 0.0.0.0 --port 8001
```
## What you can do with it
- **Query your infrastructure data from natural language** — find devices, interfaces, IP addresses, or any kind in your schema, with attribute filtering and partial-match search.
- **Explore your schema without leaving the conversation** — the server exposes your catalog, per-kind attribute/filter maps, and the GraphQL SDL as MCP resources.
- **Make changes on isolated branches** — writes land on an auto-created session branch (`mcp/session-YYYYMMDD-<hex>`); the default branch is never touched directly.
- **Submit changes for human review** — call `propose_changes` to open a Proposed Change for approval before merging.
- **Run arbitrary GraphQL** — execute any query or mutation against the Infrahub API when you need full control.
## Documentation
Full documentation, including client configuration for Cursor, VS Code, Claude Desktop, and Claude Code, is available at the [Infrahub MCP Server docs site](https://docs.infrahub.app/mcp).
- [Installation and client setup](https://docs.infrahub.app/mcp/getting-started/installation)
- [Docker / sidecar deployment](https://docs.infrahub.app/mcp/guides/docker-compose)
- [Authentication modes](https://docs.infrahub.app/mcp/references/authentication)
- [Configuration reference](https://docs.infrahub.app/mcp/references/configuration)
- [Methods reference — tools, resources, prompts](https://docs.infrahub.app/mcp/references/methods)
## About Infrahub
[Infrahub](https://github.com/opsmill/infrahub) is an open source infrastructure data management and automation platform (AGPLv3), developed by [OpsMill](https://opsmill.com). It gives infrastructure and network teams a unified, schema-driven source of truth — devices, topology, IP space, configuration — with built-in version control, a generator framework for automation, and native integrations with Git, Ansible, Terraform, and CI/CD pipelines.
## License
Apache 2.0 — see [LICENSE](./LICENSE.txt).
TDQS
Scored across 12 tools
Most tools have distinct purposes: typed reads (get_nodes/search_nodes), generic GraphQL reads (query_graphql), typed writes (node_upsert/node_delete), complex writes (mutate_graphql), session management, schema discovery, and graph traversal. The main overlap is between query_graphql and the typed read tools, but the descriptions clearly guide when to prefer each, so boundaries remain workable.
Most tools follow a verb_noun pattern (get_nodes, search_nodes, get_schema, find_paths, propose_changes), but node_upsert and node_delete use resource_action ordering, and some names like reset_session_branch are more phrase-like. The set is still readable, but the mixed conventions are noticeable.
Twelve tools is well-scoped for an Infrahub server covering schema discovery, typed reads/writes, GraphQL escape hatches, sessions, proposed changes, and graph traversal. Each tool appears to earn its place without excessive surface area.
The surface covers the core lifecycle: schema discovery, typed reads and searches, create/update via upsert, delete, complex mutations, proposed changes, and session control. Relationship edits require the generic mutate_graphql escape hatch, and there is no direct merge/approve tool, but these are reasonable design choices rather than critical gaps.