Bob Ross 🎨

"We don't make mistakes, just happy little servers."

An MCP server for managing a Canonical Landscape estate from an AI agent (Claude Desktop / Claude Code). Bob Ross wraps the Landscape API as MCP tools — inventory, alerts, activities, patching, script execution — with a safety layer built in so an LLM can't accidentally reboot your whole fleet.

Jira: BR

Why it's not just a dumb API wrapper

• Blast-radius preview — resolve_query shows how many machines a query matches (and a sample) before you act.

• Dry-run → confirm handshake — every destructive tool returns a short-lived confirm_token on the first call; you must call again with the token to execute. If the matched machine set drifts in between, the token is rejected.

• Secure by default — read-only mode is on out of the box. Writes need two explicit switches flipped. TLS is verified. Secrets never hit logs.

• Full audit log — every action (dry runs, executes, denials) is appended to a redacted JSONL audit trail.

• Dual auth — REST bearer token or legacy HMAC query API, auto-detected.

Related MCP server: MCP Process Server

Status

MVP scaffold. Read tools + gated write tools are implemented against both API backends. Exact REST endpoint paths are set as constants in client.py and get finalized once pointed at a live instance — run the ping tool to validate.

Setup

python3.14 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
cp .env.example .env      # then fill in URL + credentials
pytest                    # run the safety/signing/config tests

Configuration

All via BOBROSS_* env vars (or .env). See .env.example. Key ones:

Run

bob-ross            # stdio transport (Claude Desktop / Code)
# or
python -m bob_ross

Claude Code (CLI)

Install once (user scope → available in every project). The cd wrapper makes it location-independent and keeps credentials in .env (never in Claude's config):

claude mcp add bob-ross --scope user -- \
  bash -lc 'cd /path/to/bob-ross && exec ./.venv/bin/bob-ross'

Starts in read-only mode. To enable actions, flip BOBROSS_READ_ONLY=false and BOBROSS_ALLOW_WRITES=true in .env, then restart Claude Code.

Claude Desktop / Code (stdio)

{
  "mcpServers": {
    "bob-ross": {
      "command": "/path/to/bob-ross/.venv/bin/bob-ross",
      "env": { "BOBROSS_LANDSCAPE_URL": "https://landscape.example.com",
               "BOBROSS_API_TOKEN": "..." }
    }
  }
}

Tools

Read: ping, estate_health, list_computers, get_computer, resolve_query, pending_updates, list_alerts, list_activities, get_activity, wait_for_activity, list_scripts Write (gated): execute_script, reboot_computers, apply_security_upgrades, upgrade_packages, install_packages, remove_packages, add_tags, remove_tags

Activity-creating write tools accept wait=true to poll the resulting Landscape activity to a terminal status and return a completion summary (succeeded / failed / still-incomplete), so the agent knows the real outcome instead of just "queued".

Resources: landscape://computers, landscape://alerts, landscape://health, landscape://computer/{computer_id} (template) Prompts: patch_security_updates, triage_estate, reboot_reboot_required, patch_machine