Rocannon runs on your Ansible control node and turns it into an MCP server. At startup it reads ansible-doc and exposes every module you have installed (plus any role with an argument spec) as a typed tool, so an MCP client like Claude Code, Cursor, or your own agent can drive your real environment in plain English. The tool surface is whatever you have installed, one collection or a hundred. Each tool's name is its Ansible FQCN with dots swapped for underscores (e.g. ansible.builtin.copy becomes ansible_builtin_copy), the form OpenAI-style function calling requires; the original FQCN stays in the tool's description.

demo

Install

pip install rocannon

This brings ansible-core and ansible-runner with it. rocannon doctor checks the environment for anything missing.

Quickstart

rocannon quickstart

Scaffolds a localhost profile under .rocannon/ and prints the wiring for your MCP client (Claude Code, Claude Desktop, Cursor) plus a command to confirm the tools registered. Then ask your assistant something like "Gather facts from localhost and tell me the OS and kernel version."

You don't need an LLM, though. The same tools are a shell:

rocannon mcp doctor --profile .rocannon/quickstart.yml   # list registered tools
rocannon repl       --profile .rocannon/quickstart.yml   # operator shell

What it does

  • Reflects your modules. Each installed module becomes a typed MCP tool, with parameters, types, defaults, and choices read from ansible-doc. Whatever you install shows up automatically.
  • Scales to whole collections. Set discovery: progressive in a profile and the module tools start hidden behind two: ansible_search_modules ranks the catalog by capability, and ansible_use_module reveals a match for the session. A revealed module is the real typed tool, same parameters, choices, safety hints, dry-run flags, and approval gate as in static mode, just surfaced on demand so hundreds of modules don't flood the client's context. The default (static) registers every module up front, which is best for small profiles.
  • Reflects your roles. A role with a meta/argument_specs.yml becomes a typed tool too; its arguments are the parameters, validated by ansible at run time.
  • Carries the metadata. Tools get safety hints (read-only vs destructive), collection and namespace tags, and a meta block with the module's documented requirements, return keys, and version, straight from ansible-doc.
  • No lock-in. Record any session, from an agent or the CLI, to a standard Ansible playbook under .rocannon/playbooks/ that runs with plain ansible-playbook. Saved sessions also load back as MCP prompts.
  • Dry runs. Modules that support check mode expose --check and --diff, both on the CLI and as MCP tool parameters.
  • Approval gates. Set ROCANNON_APPROVAL=destructive (gate the command/shell/script/raw family) or =writes (gate everything that changes state) and Rocannon asks the human to confirm each gated call through the MCP client before it runs, via the protocol's elicitation request. Dry-runs are never gated. If approval is required but the client can't elicit, the call is refused rather than run unattended.
  • A CLI, too. Every module is a subcommand: rocannon ansible.builtin.copy --target h1 -i hosts --src a --dest b.

Examples

Profiles

A profile is a YAML file declaring an inventory plus the modules and roles to expose:

inventories:
  - ./hosts
modules:
  - ansible.builtin
  - community.docker
roles:                      # optional
  - my_ns.my_coll.setup_web
roles_path: ./roles         # optional, for standalone (non-collection) roles
  • modules takes a module (ansible.builtin.copy), a collection (ansible.builtin), or a namespace (community). Only modules become tools; filter and lookup plugins are skipped.
  • roles takes a collection role by FQCN, or a standalone role by its directory name together with roles_path (which resolves against the profile's own directory). Roles without an argument spec are skipped.
  • Optional keys: ansible_cfg, vault_password_file, extra_envvars.
  • execution_environment dispatches module and role execution into a container image (built with ansible-builder, or a plain Dockerfile) instead of Rocannon's own process, via ansible-runner's native process_isolation/container_image support, the same mechanism AWX and ansible-navigator use. execution_environment_engine picks the container engine (podman, default, or docker), and execution_environment_container_options passes extra args straight to it (for example ["--network", "my-compose-net"] so the container can reach other services by name). Collection Python dependencies (psycopg2 for community.postgresql, the docker SDK for community.docker, and so on) then live in the image, baked in from each collection's own declared requirements, not in Rocannon's control-side environment. See examples/execution-environment-dispatch for a working image + profile.
    • Discovery matches execution. At startup, Rocannon reflects each profile's modules and collection roles by running ansible-doc inside that same image (ansible_runner.get_plugin_list/get_plugin_docs, the doc-side sibling of the execution dispatch), so the tool surface it registers is exactly what the image can run, not whatever (if anything) happens to be installed locally.
    • Roles are asymmetric. A collection role (FQCN) is expected to already live in the image, same as a module. A standalone roles_path role is host-local: introspected locally (that directory is never visible inside the container), but still staged into the container and executed there.
    • rocannon doctor checks the configured engine binary is on PATH when execution_environment is set, so a typo'd engine fails fast instead of surfacing as an opaque error mid-call.

Drop multiple profiles in .rocannon/profiles/ (with a default.yml) and switch at runtime via the rocannon_list_profiles, rocannon_current_profile, and rocannon_use_profile tools.

Dependencies: modules with third-party Python libraries (for example community.crypto needs cryptography, community.docker needs the docker SDK) need them installed in the same environment as Rocannon. The quickstart inventory pins ansible_python_interpreter so localhost runs use it.

Rocannon vs dedicated MCP servers

Most MCP servers target one layer: the service API. The MongoDB MCP server queries documents. The AWS MCP server describes EC2 instances. The Kubernetes MCP server inspects pods. None of them can touch the OS underneath.

Rocannon operates at the OS and configuration layer — the same layer Ansible has always owned. That makes it complementary to service-layer MCP servers, and the only MCP option for domains that have no official server at all.

Collection What it does that service MCP can't Equivalent service MCP
amazon.aws Configure the OS on EC2 — packages, users, services, files AWS MCP (awslabs)
azure.azcollection Configure VMs after provisioning; multi-cloud plays Azure MCP (Microsoft, official)
google.cloud Configure GCE VMs; self-managed DBs on GCE Google Cloud MCP (50+ official servers)
kubernetes.core Configure nodes before the API exists; bootstrap kubeadm kubernetes-mcp-server (Red Hat)
community.mongodb Install mongod, write mongod.conf, build replica sets across hosts MongoDB MCP (official)
community.postgresql Install PostgreSQL, configure pg_hba.conf, set up streaming replication No official MCP server
community.mysql Install MySQL, configure my.cnf, manage replication No official MCP server
community.hashi_vault Install and initialize Vault on hosts; pull Vault secrets into module args Vault MCP (HashiCorp, beta)
cisco.ios / arista.eos Idempotent device config via NETCONF/SSH; --check/--diff before pushing Cisco/Arista management-plane MCP only
ansible.builtin / ansible.posix Package, service, file, user management — abstracted across distros SSH MCP (community only, raw shell)

A few things stand out from this table:

  • Self-managed PostgreSQL and MySQL have no official MCP server. The Anthropic reference Postgres server was deprecated in July 2025. For those, Rocannon is the only MCP path.
  • Network automation is where MCP coverage is thinnest. cisco.ios, arista.eos, and junipernetworks.junos have decades of vendor investment. The network MCP servers cover management planes (Catalyst Center, CloudVision), not direct device configuration.
  • Dedicated service MCP servers do some things better. MongoDB MCP's Atlas integration, AWS MCP's CloudWatch log correlation, and Google Cloud MCP's managed remote endpoints are genuinely useful for their specific domains. Use them alongside Rocannon when you need both layers.

The typical workflow:

  1. Use a service MCP server to query and explore (MongoDB MCP to inspect a collection, AWS MCP to find a misconfigured security group).
  2. Use Rocannon to act at the OS or config layer (fix the EC2 user data, restart a service, push a patched config file).
  3. Run commit_session to save the successful steps as a standard ansible-playbook-runnable YAML — no Rocannon needed for the next run.

What plain Ansible still does better than Rocannon: multi-host sequencing with serial, throttle, and run_once; change-management pipelines; scheduled drift enforcement; versioned roles and collections. Rocannon is for exploration and targeted action. When a task needs to be repeatable and auditable, commit it to a playbook.

CLI

rocannon quickstart    scaffold a localhost profile and print client wiring
rocannon <fqcn>        invoke a module: rocannon ansible.builtin.copy ...
                       optional --record FILE appends each call to a playbook
rocannon mcp serve     start the MCP server (stdio or http)
rocannon mcp doctor    list registered tools, resources, prompts
rocannon repl          interactive shell on the same MCP server
rocannon doctor        system health (binaries, env, inventory)
rocannon doc <module>  print parsed schema for a module
rocannon search <q>    find modules by name or description
rocannon ls <kind>     list hosts, groups, or modules from a profile
rocannon playbook      list/show/run saved playbooks

Each module invocation needs an inventory: pass -i/--inventory, --profile, or run where a .rocannon/profiles/ is discovered. rocannon <fqcn> --help shows that module's typed flags.

MCP clients

A working .mcp.json ships at the repo root; per-client snippets are in examples/clients/.

Client Config location
Claude Code .mcp.json at project root, or claude mcp add
Claude Desktop macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Cursor .cursor/mcp.json or ~/.cursor/mcp.json
mcphost ~/.mcphost.yml or --config <path>
IBM Bob .bob/mcp.json or ~/.bob/mcp_settings.json

All use the standard mcpServers envelope pointing at rocannon mcp serve --profile <your-profile.yml>.

Development

git clone https://github.com/msradam/rocannon.git
cd rocannon
uv sync
./tests/check.sh                # ruff format + lint + mypy + pytest
uv run pytest -m integration    # opt-in: real Ansible against localhost / a UBI9 container

See ARCHITECTURE.md for how the pieces fit together. Rocannon is developed with AI assistance.

The name

Ursula K. Le Guin coined the word "ansible" in her 1966 novel Rocannon's World. The gryphon is a nod to the Windsteeds that Rocannon and his companions ride.

Credits