内容
# caveman-browse
**A token-efficient alternative to Playwright MCP for coding agents.** One Go
binary that attaches to the Chrome you already run, reads the real
accessibility tree, compresses it, and gives your agent four tools that cost
**297 tokens of catalog** — versus 3,422 for Playwright MCP (0.0.79) and 4,507
for Chrome DevTools MCP (1.7.0), measured live 2026-08-14.
| | caveman-browse | Playwright MCP | Chrome DevTools MCP |
|---|---|---|---|
| Tool-definition cost (measured 2026-08-14) | **297 tokens** (4 tools) | 3,422 tokens (24 tools) | 4,507 tokens (29 tools) |
| Page representation | compressed a11y tree | aria-YAML (lossy) | a11y snapshot |
| Focused query on a 200-row page | **~98 tokens** | ~15.7k (full ARIA) | full snapshot |
| Byte-exact recovery of the original tree | **yes (CCR)** | no | no |
| Runtime | 1 static Go binary + your Chrome | Node + Chromium | Node + Puppeteer + Chromium |
Numbers are `inferred` token counts measured with an offline `o200k_base`
counter on pinned fixtures — see [BENCHMARK.md](BENCHMARK.md) for methodology,
reproduction commands, and the cases where Playwright's bare ARIA text is
*smaller* (tiny pages; we print those too).
**Full-task head-to-head** (same headless agent, same task, server swapped;
provider-reported session tokens; 2026-08-14 pilot, n=1/cell, all arms
passed): finding one row in a 200-row table cost **150k** total input tokens
through caveman-browse vs 203k (Playwright MCP) / 223k (Chrome DevTools MCP);
counting matching rows **168k vs 358k / 223k**. On the action-heavy checkout
task caveman-browse *lost* (668k vs 362k / 312k) — its settle-proof loop
re-snapshots after every action instead of trusting dispatch. Full tables,
causes, and claim boundaries: [benchmarks/agentloop/RESULTS.md](benchmarks/agentloop/RESULTS.md),
[benchmarks/toolcatalog/RESULTS.md](benchmarks/toolcatalog/RESULTS.md),
[benchmarks/snapshot/RESULTS.md](benchmarks/snapshot/RESULTS.md).
## Install
**Claude Code (plugin — server + usage skill):**
```
/plugin marketplace add JuliusBrussee/caveman-browse
/plugin install caveman-browse@caveman-browse
```
**Claude Code (MCP only):**
```bash
claude mcp add caveman-browse -- npx -y caveman-browse
```
**Cursor / any MCP host** (`.mcp.json` / `mcp.json`):
```json
{ "mcpServers": { "caveman-browse": { "command": "npx", "args": ["-y", "caveman-browse"] } } }
```
**Gemini CLI:**
```bash
gemini extensions install https://github.com/JuliusBrussee/caveman-browse
```
The npm shim downloads a signature-verified prebuilt binary on first run. No
Node runtime is involved after launch; no Chromium is bundled — it drives your
installed Chrome with a dedicated automation profile.
## The four tools
- `browser_snapshot(url?, wait?, query?, interactive?)` — compact accessibility
tree with `[uid]` action handles plus one accounting line
(`caveman before=… after=… ratio=… basis=inferred handle=ccr_…`).
`query` keeps the ≤12 best-matching nodes plus ancestors — the
token-efficient path on large pages. `interactive: true` keeps only
actionable nodes.
- `browser_act(action, uid, text?, option?)` — `click`, `type`, `select`,
`scroll`, `wait`. Returns `settled:false` until a focused re-snapshot proves
the app updated; the driver never claims success it hasn't seen.
- `browser_eval(expression)` — JS escape hatch.
- `browser_recover(recovery_handle, query?)` — the **byte-exact** original
`Accessibility.getFullAXTree` payload behind any snapshot (CCR-backed), or
BM25-narrowed sections with a query. Compression here is never a one-way
door.
## Direct CLI
```bash
npx caveman-browse snapshot http://127.0.0.1:3000
npx caveman-browse snapshot http://127.0.0.1:3000 "save settings" # focused
npx caveman-browse snapshot -i http://127.0.0.1:3000 # actionable only
npx caveman-browse act <uid> click
npx caveman-browse recover <handle>
npx caveman-browse close
```
Direct commands share one detached, isolated Chrome until `close`. Navigation
allows `http(s)`, `about:blank`, and bounded `data:text/html`; `file:`,
`javascript:`, and privileged schemes fail closed with `cave_snake_code`
errors.
## Honesty model
- Every savings figure is **`inferred`** (offline BPE count), never provider
usage, never billing, never re-projected to a period.
- Lossy snapshots are emitted **only** when the byte-exact original is stored
for recovery (CCR-or-passthrough).
- Unknown roles, uids, or CDP responses fail closed — no fabricated handles,
no claimed success without settlement proof.
- Scope: same-origin pages with predictable controls. Cross-origin iframes
appear as leaves; OOPIFs, dialogs, downloads, and arbitrary-open-web
actionability are out of scope (see the claim boundary in BENCHMARK.md).
## Build from source
```bash
go build ./cmd/caveman-browse
CAVEMAN_BROWSE_CHROME="/path/to/Chrome" go test -tags=integration ./...
```
## License
Source and binaries ship under the [Apache License 2.0](LICENSE). Third-party
notices: [NOTICE](NOTICE). Part of the [Caveman](https://github.com/JuliusBrussee/caveman)
efficiency stack; the a11y compressor and CCR live in the Caveman engine.
MCP 市场