Borrowing it
Nothing to install: this file belongs to JuliusBrussee/caveman-browse. Take a copy, put it at the same path in your own repository, and replace the rules that are about this project with yours.
curl -O https://raw.githubusercontent.com/JuliusBrussee/caveman-browse/main/CLAUDE.mdgit clone --depth 1 https://github.com/JuliusBrussee/caveman-browseWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/instructions/juliusbrussee/caveman-browse/claude-md)<a href="https://agentmods.dev/instructions/juliusbrussee/caveman-browse/claude-md"><img src="https://agentmods.dev/badge/instructions/juliusbrussee/caveman-browse/claude-md/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/instructions/juliusbrussee/caveman-browse/claude-md"><img src="https://agentmods.dev/badge/instructions/juliusbrussee/caveman-browse/claude-md.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.01244 | $0.01244 |
| Opus 5.5 | $0.00498 | $0.00498 |
| Sonnet 5.5 | $0.00249 | $0.00249 |
| Haiku 4.5 | $0.00124 | $0.00124 |
Grade A, and why
caveman-browse CLAUDE.md scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2mo ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 79 lines — stays where its author put it; the contents beside it link to each section on GitHub.
caveman-browse — token-efficient browser driver
Standalone repo. Serves
stdio MCP tools that read a real Chrome accessibility tree, compress it with
the Caveman engine's forced-only a11y compressor, act on uid handles, and
recover the byte-exact original AX payload through CCR. Every saving here is
inferred; Browse never emits verified.
Layout
session.go— MCP tool handlers, engine/CCR integration, UID target cache.cdp.go— chromedp-backed Mode-A dedicated Chrome driver and bounded actionability recipe.cmd/caveman-browse/— stdio MCP binary + direct CLI subcommands.bin/— npm shim (launcher + generated installer);scripts/gen-installer.mjsregenerates the generated files fromBINARY_RELEASE+BINARY_SIGNING_PUBKEY.pub..claude-plugin/,.codex-plugin/,gemini-extension.json,skills/— installable plugin surfaces;.mcp.jsonis the server entry they share.benchmarks/— head-to-head suite vs Playwright MCP / Chrome DevTools MCP.
Module boundary
- This module pins
github.com/JuliusBrussee/caveman(engine, engine/ccr, engine/tokens, mcp framing) ingo.mod. Bumps are explicit commits. Thea11ycompressor (engine/compressors/axtree.go) lives in that repo, not here. - Signed binaries are still published on the caveman monorepo releases
(
scripts/gen-installer.mjsrelease base); repo-owned signed releases are a tracked follow-up.
Gotchas
go test ./...is setup-free.CAVEMAN_BROWSE_CHROME=<chrome> go test -tags=integration ./...runs the CDP contract across the package and direct CLI.- The actionability layer is intentionally bounded: same-origin dashboards and predictable design-system controls, not arbitrary-open-web parity.
- Unknown handles/actions fail closed with
cave_snake_codeerrors. - Snapshot output is compact indented text, not JSON-lines.
querykeeps at most 12 highest-scoring task matches plus ancestors; recovery metadata exposes only UIDs actually shown. Keeptokens_afterequal to exact agent-visible result cost andview_tokensequal to serializer output cost. - The whole payload is text, not a JSON envelope: the AX view plus one
trailing
caveman before= view= after= ratio= basis= handle=line (renderSnapshot). Re-wrapping it in JSON escapes every already-strconv.Quoted accessible name a second time — a measured 12.5% tax plus 49 envelope tokens.finalizeSnapshotPayloadstill settles the self-reference (the line states the count that stating it changes);ratiois rounded to the 4 decimals printed,before/afterstay exact. interactive: truekeeps uid-bearing lines plus the ancestors that place them and nothing else. It is a flag and never a default — it hides the text an agent has to READ — and fails open to the full view when no uid survives. It never narrows the uid map: the target cache always comes from the full curate pass.- The four-tool catalog is at 297 of its 300-token budget. A fifth tool does not
fit, and neither does a verbose description. That budget is the moat versus
3,422 (@playwright/mcp 0.0.79) and 4,507 (chrome-devtools-mcp 1.7.0)
tool-definition tokens, measured 2026-08-14 (
benchmarks/toolcatalog/); earlier Playwright MCP builds measured ~13.7k — always re-measure, never quote stale competitor numbers. - Direct CLI Chrome is detached so separate
snapshot/act/evalprocesses can share a target.closemust terminate it. FreshCAVEMAN_HOMEmust work; state writes stay atomic and mode0600. - CDP action acknowledgement is not application settlement. Non-wait actions
return
settled:falseand require a focused resnapshot for proof. - Navigation denies
file:,javascript:, and privileged Chrome schemes. - Benchmark contract and reproducible comparisons live in
BENCHMARK.md; keep token budgets executable in tests. - The uid map is
browser_snapshot's contract, not a side effect of a compression ratio. On engine pass-through (no recovery handle — a tree that did not get smaller, or no CCR store)snapshotToolfails closed withcave_browser_snapshot_uncompressedand returns the prior page's uids intact; it MUST NEVER dump the raw AX tree intouids(strictly worse than not using Browse) nor wipe the target cache. Regressing this reopens a known acting-dead failure mode. - An
<iframe>is a leaf, not a broken tree.Accessibility.getFullAXTreereturns one frame at a time, so an iframe node'schildIdpoints at a child document absent from this payload; thea11ycompressor (engine/compressors/axtree.go) treats an unresolvablechildIdas a leaf and still curates the frame-visible nodes. Because of that the CDP driver keeps Chrome's default Site Isolation (site-per-process) — do not disable it to "fix" cross-origin iframes.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 2mo ago First seen · 79 lines · 1,244 tokens per session scan A 2fc76a89a097
caveman-browse CLAUDE.md is an instructions file published in the GitHub repository JuliusBrussee/caveman-browse (49 stars, last pushed 1mo ago), licensed Apache-2.0. It adds 1,244 tokens to every session, about $0.0050 per session on Opus 5.5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-09-04.
caveman-browse CLAUDE.md compared with instructions from other repositories
| Mod | Type | Stars | Last push | Licence | Scan | Tokens | Works with |
|---|---|---|---|---|---|---|---|
| caveman-browse CLAUDE.mdJuliusBrussee/caveman-browse | Instructions file | 49 | 1mo ago | Apache-2.0 | A | 1,244 | Claude Code |
| vscode buildNext.instructions.mdmicrosoft/vscode | Instructions file | 195k | yesterday | MIT | A | 9,719 | GitHub Copilot |
| next.js AGENTS.mdvercel/next.js | Instructions file | 143k | today | MIT | A | 7,461 | Codex, OpenCode |
| langchain AGENTS.mdlangchain-ai/langchain | Instructions file | 149k | yesterday | MIT | A | 4,680 | Codex, OpenCode |
| vscode oss-third-party-notices.instructions.mdmicrosoft/vscode | Instructions file | 195k | yesterday | MIT | A | 5,516 | GitHub Copilot |
| vscode copilot-instructions.mdmicrosoft/vscode | Instructions file | 195k | yesterday | MIT | A | 3,093 | GitHub Copilot |
| deepseek-harness AGENTS.mddeepseek-ai/deepseek-harness | Instructions file | 274k | 2d ago | MIT | A | 4,313 | Codex, OpenCode |