Bob Ross
# Bob Ross 🎨
[](https://pypi.org/project/bob-ross-landscape/)
[](https://github.com/just-an-oldsalt/bob-ross/actions/workflows/ci.yml)
[](LICENSE)
> *"We don't make mistakes, just happy little servers."*
An MCP server for managing a **Canonical Landscape** estate from an AI agent
(Claude Code / Desktop). Bob Ross wraps the Landscape API as calm, friendly MCP
tools — inventory, alerts, activities, patching, script execution, reboots — with
a **safety layer built right in**, so your agent can tend a whole fleet of happy
little servers without ever beating the devil out of the wrong one.
Everybody needs a friend. Even your Ubuntu estate.
---
## 🖌️ Why it's not just a dumb API wrapper
A blank canvas is a beautiful thing — and so is a fleet that's still standing.
Bob Ross paints carefully:
- **Blast-radius preview** — before you touch anything, `resolve_query` steps back
from the easel and shows you *how many* machines a query matches (and a sample).
Always know how big the canvas is.
- **Dry-run → confirm handshake** — every destructive tool returns a short-lived
`confirm_token` on the first call. You have to call again *with* the token to
actually do it. If the set of matched machines drifts in between, the token is
refused. No happy little accident reboots 500 boxes.
- **Secure by default** — read-only mode is **on** out of the box (we all start
with a clean canvas). Writes need *two* switches flipped on purpose. TLS is
verified. Secrets never touch the logs.
- **Full audit log** — every stroke (dry runs, executes, denials) is appended to a
redacted JSONL trail. You can always see what the brush did.
- **Activity-aware** — write actions in Landscape run later, asynchronously. Pass
`wait=true` and Bob Ross watches the paint dry, then tells you succeeded /
failed / still-going per machine — not just "queued."
- **Dual auth** — legacy HMAC query API *or* REST bearer token, auto-detected.
---
## 🎨 Get the paints out (install)
From PyPI — the whole studio in one command:
```bash
pip install bob-ross-landscape
```
That gives you the `bob-ross` command (the import package is `bob_ross`).
<details>
<summary>Or install from source (for hacking on it)</summary>
```bash
git clone https://github.com/just-an-oldsalt/bob-ross
cd bob-ross
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
pytest # unit tests (safety / signing / config / health / ...)
python scripts/smoke.py # safe live smoke test vs your instance (read-only, no writes)
```
</details>
---
## 🖼️ Set up your palette (configuration)
Everything is `BOBROSS_*` env vars (or a `.env` file — see `.env.example`):
| Var | Default | Meaning |
|---|---|---|
| `BOBROSS_LANDSCAPE_URL` | — | Landscape base URL (no trailing `/api/`) |
| `BOBROSS_ACCESS_KEY` / `BOBROSS_SECRET_KEY` | — | Legacy HMAC keys (mode A) |
| `BOBROSS_API_TOKEN` | — | REST bearer token (mode B) |
| `BOBROSS_READ_ONLY` | `true` | Blocks all writes — the safe blank canvas |
| `BOBROSS_ALLOW_WRITES` | `false` | The *second* switch needed to enable writes |
| `BOBROSS_TLS_VERIFY` | `true` | Verify TLS certs (set `false` only for dev self-signed) |
> To let Bob Ross pick up a brush, flip **both** `BOBROSS_READ_ONLY=false` **and**
> `BOBROSS_ALLOW_WRITES=true`. Even then, every destructive action still needs the
> dry-run → confirm handshake. Belt and suspenders. Happy and safe.
---
## 🪄 Add it to Claude Code
```bash
pip install bob-ross-landscape
claude mcp add bob-ross --scope user \
--env BOBROSS_LANDSCAPE_URL=https://landscape.example.com \
--env BOBROSS_ACCESS_KEY=YOUR_KEY \
--env BOBROSS_SECRET_KEY=YOUR_SECRET \
-- bob-ross
```
Prefer to keep secrets out of Claude's config? Put them in a `.env` file and run
from that directory instead:
```bash
claude mcp add bob-ross --scope user -- \
bash -lc 'cd /path/to/your/bobross-env && exec bob-ross'
```
### Claude Desktop (stdio)
```json
{
"mcpServers": {
"bob-ross": {
"command": "bob-ross",
"env": {
"BOBROSS_LANDSCAPE_URL": "https://landscape.example.com",
"BOBROSS_ACCESS_KEY": "YOUR_KEY",
"BOBROSS_SECRET_KEY": "YOUR_SECRET"
}
}
}
}
```
Bob Ross always wakes up in read-only mode. Start there. Get comfortable. Then,
when you're ready, let's get a little crazy.
---
## 🌲 The brushes (tools)
**Read** (always safe — look all you like):
`ping` · `estate_health` · `list_computers` · `get_computer` · `resolve_query` ·
`pending_updates` · `list_alerts` · `list_activities` · `get_activity` ·
`wait_for_activity` · `list_scripts`
**Write** (gated — dry-run → confirm every time):
`execute_script` · `reboot_computers` · `apply_security_upgrades` ·
`upgrade_packages` · `install_packages` · `remove_packages` · `add_tags` ·
`remove_tags`
> Activity-creating write tools take `wait=true` to poll the resulting Landscape
> activity to a terminal status and hand back a `completion` summary
> (succeeded / failed / still-incomplete) — so your agent knows the real outcome,
> not 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`
---
## 🏔️ A happy little workflow
You have unlimited power here. Move mountains — one confirmed step at a time:
```
you: "what needs my attention across the fleet?"
→ estate_health → "kaylee-mc: 138 pending upgrades, 3 boxes need reboots"
you: "what would patching kaylee-mc actually change?"
→ pending_updates title:kaylee-mc → the list, per package
you: "apply the security upgrades there"
→ apply_security_upgrades → dry-run shows blast radius + a confirm_token
→ (you approve) → re-run with the token + wait=true
→ completion: succeeded ✅ no failed patches, just happy little servers
```
---
## 📦 Publishing & links
- **PyPI:** https://pypi.org/project/bob-ross-landscape/
- **Releases:** tag `vX.Y.Z`, `gh release create` → GitHub Actions publishes to PyPI
via OIDC (no tokens). See [`PUBLISHING.md`](PUBLISHING.md).
- **MCP Registry manifest:** [`server.json`](server.json)
---
<div align="center">
*"Talent is a pursued interest. Anything you're willing to practice, you can do."*
Now go tend some happy little servers. 🎨
</div>
TDQS
Scored across 19 tools
Each tool has a clearly distinct purpose, from package management to query resolution and system health. Even similar operations like apply_security_upgrades and upgrade_packages are differentiated by description and scope, leaving no ambiguity for an agent.
Most tool names follow a verb_noun pattern (e.g., install_packages, list_computers), with a few exceptions like pending_updates (adjective_noun) and estate_health (noun_noun). The pattern is largely consistent, aiding predictability.
With 19 tools covering a broad IT estate management domain, the count is slightly above the ideal range but justified by the need for granular operations (e.g., separate security upgrades, pending updates, scripts). No tool feels superfluous.
The tool surface covers core CRUD and lifecycle operations for computers, packages, tags, scripts, and activities. Minor gaps exist, such as lack of computer creation/deletion or patch scheduling, but the set is sufficient for common management tasks.