caveman-setup Skill for Claude: Route Your LLM Calls Through the Caveman Gateway
caveman-setup makes your assistant route every LLM call in your repo through the Caveman Cloud gateway so cost is measured, with no behavior change.
- Installs
- 110.2K
- on skills.sh
- Stars
- 116.7K
- on GitHub
- Updated
- Nov 2026
- last repo commit
caveman-setup finds each live LLM callsite (OpenAI, Anthropic, Vercel AI SDK, LangChain, LiteLLM or raw HTTP), picks an app slug and points the base URL at the gateway with one auth header. It keeps the key in your env file, adds that file to .gitignore if needed, and sends one small real request to confirm HTTP 200 with usage numbers.
It expects four values from a setup prompt: the gateway URL, the Caveman API key, whether provider keys are stored or sent per request, and the dashboard URL. If any is missing it stops and asks. It runs in record mode only, so it measures and does not turn on any optimizer. It bundles no scripts.
Good for
- Developers who ship LLM features and want cost visibility
- Teams starting with Caveman Cloud who need the first integration done
- Anyone who wants a minimal change with a verified first request
Install caveman-setup
Claude Code
- 1
Copy this prompt.
Install the caveman-setup skill in Claude Code. Run this command: npx skills add https://github.com/juliusbrussee/caveman --skill caveman-setup Then check that it loaded and tell me how to use it.
- 2
Paste it into Claude Code. It shows the change and asks before running it.
Requires Node.js.
From its listing on skills.sh. Source.
What's in the skill
When it loads
Wire a repository through the Caveman Cloud gateway so every LLM request is measured, with no behavior change. Use for "set up caveman" or adding LLM spend observability.
Sections in SKILL.md
- Rules (non-negotiable)
- Step 1 — Find every live LLM callsite
- Step 2 — Pick the app slug
- Step 3 — Wire each callsite
- Step 4 — Verify with one real request
- Step 5 — Report
- Caveman is live in this repo
- Failure templates (use verbatim, filled in — never soften)
Key facts
- Author
- juliusbrussee (GitHub)
- Installs on skills.sh
- 110,239 (as of 2026-11-02)
- License
- Apache-2.0
- Last repo commit
- 2026-11-01
- Files in the skill folder
- 1 file: SKILL.md
- skills.sh audits
- Gen Agent Trust Hub: pass, Socket: warn, Snyk: pass
What it includes
List SDK imports, raw HTTP calls and base-URL variables
Searches dependency files and source; if it finds none, it stops with a "nothing to wire" message.
Read ·
Step 1: Find every live LLM callsiteOne slug names the app in the gateway path
Derives it from the package name: lowercase, up to 64 characters.
Step 2: Pick the app slugBase URL to the gateway plus one auth header
Edits each client for its SDK, adds CAVE_GATEWAY_URL and CAVE_API_KEY to your env file, and adds that file to .gitignore if it is not ignored.
Write ·
Step 3: Wire each callsiteOne small billable request through the new wiring
Sends a request with a small token cap and reads the response. Success is HTTP 200 with a usage block.
Runs code ·
Step 4: Verify with one real requestWired callsites, status and tokens from the real response
Reports only what it saw, with set failure messages for an unreachable gateway, a rejected key or a wrong route.
Step 5: Report
From the official docs
Permissions
Medium risk
It rewires your LLM calls through a third-party gateway and sends one real, billable test request. Use a branch, check where your env file is stored, and read what the gateway receives before you roll it out.
What it can do
- Read your data
- Create and change things
- Delete things
- Read-only mode
Watch out for
It sends a real, billable request
Verification is one small request through your own provider account. The SKILL.md says it is real and billable and that it does not pause to ask for it.
It routes your LLM traffic through Caveman
After setup, requests from your app go through the Caveman gateway. In byok mode the SKILL.md has the app send its provider key to the gateway in an x-cave-upstream-key header. The README's Privacy section covers the CLI and Auto, not this gateway, so read Caveman's own terms first.
It writes a secret to your env file
The key goes into your env file and the skill adds that file to .gitignore if it is not ignored. Check that it is not committed.
Technical details
| Talks to | Caveman gateway |
|---|---|
| Scopes |
|
Questions
What does caveman-setup do?
It points every LLM call in your repo at the Caveman Cloud gateway, adds the gateway key to your env file, and sends one small test request. The result is that Caveman can measure your LLM cost without changing what your app does.
How do I install caveman-setup?
Use the install box at the top of the page. The SKILL.md expects a setup prompt that supplies the gateway URL, the Caveman key, the provider-key mode and the dashboard URL. Without those values it stops and asks.
Does caveman-setup need an API key?
Yes. It needs a Caveman gateway key. In byok mode your app also keeps sending its own provider key. In stored mode the provider key lives encrypted in Caveman Cloud and the skill never sees it.
Will it enable optimizations or claim savings?
No. The SKILL.md says record mode only: it measures and enables no optimizer, and verified savings stay at $0 until you turn one on.
Is caveman-setup safe?
It edits code and env files and sends one real, billable request, so use a branch and review the diff. The SKILL.md says the key stays in env vars and is never hardcoded or printed in full. After setup your app's LLM traffic goes through the Caveman gateway, so read Caveman's own terms for the gateway before you roll it out.
What should I run after caveman-setup?
The report suggests caveman-discover to split spend by workflow instead of only by app. See that skill for how it labels each job.
Information
updated Nov 2, 2026
- Repository
- juliusbrussee/caveman
- Folder
skills/caveman-setup
Sources
Docs, issue trackers and discussions this page is built from.
Show all 3Hide
- caveman-setup SKILL.md · github.com
- caveman README · github.com
- caveman-discover SKILL.md · github.com