Accessibility-first browser & computer use for AI agents — one MCP server for every harness (Codex, Claude Code, OpenCode, Antigravity). Instruments the DOM/accessibility tree via CDP instead of screenshot loops: diffed snapshots, stable element refs, trusted input, typed errors. Zero dependencies
axuse vs Playwright MCP — benchmark result
Same harness (Codex CLI), same model (gpt-5.6-luna, reasoning medium), same 23 tasks,
same machine. Only the browser layer differs: axuse MCP vs @playwright/mcp 0.0.83
(--headless --isolated --caps vision). Both arms ran in an isolated Codex home with no
built-in cua / node_repl / bundled browser plugins, so neither arm could reach another
browser layer. Scoring is server-side ground truth, as in README.md.
Result (71 runs per arm)
| Playwright MCP | axuse | |
|---|---|---|
| Passed (strict rules) | 71/71 | 71/71 |
| Median time per task | 16 s | 13 s |
| Total time | 1493 s | 1183 s (-21%) |
| Median total tokens | 168k | 144k |
| Total tokens | 15.38 M | 11.84 M (-23%) |
| Median tool calls | 8 | 5 |
| Total tool calls | 673 | 402 (-40%) |
Per task, axuse is equal or lower than Playwright on 21 of 23 tasks for time and 23 of 23 for tool calls. It is behind on tokens in 9 tasks (t1, t3, t8, t9, t12, t13, t14, t15, t17, mostly by small margins) and behind on time in 2 (t5, t10).
Provenance (read before quoting)
The table merges two runs, as agreed, instead of re-running everything:
- 22 tasks (all but t7): 3 repetitions each, build v3 (post-action settling, stale-ref
recovery, slimmer tool text, screenshot pixel-ratio fix).
results/pw-v3-full2.jsonl. - t7 (infinite-scroll feed): 5 repetitions, build v4 = v3 + the
ax_observe findfix.results/pw-t7-v4.jsonl. The other 22 tasks were not re-run on v4. - Merged file:
results/final-vs-playwright.jsonl(each row carries itsbuild).
The raw per-run logs (results/*.jsonl) record local file paths and are not committed; they are available on request ([email protected]). The harness that produced them is in this folder (see README.md, "Comparing against another MCP browser tool").
On v3, axuse failed 2 of 3 t7 runs (it liked the neighbouring post because find
printed the previous article's Like button next to the match; the task cannot recover from a
wrong like). Those two failures are the reason the t7 rows were replaced; the same v3 run had
Playwright at 3/3 on t7. After the fix, t7 is 5/5 for both arms. 5 runs is a small sample.
Scoring notes
- Strict rule: a run is disqualified if any call reaches another arm's browser layer.
rescore.mjsalso reports a narrower rule (disqualify only if the foreign call touched the task site). On this merged set strict and narrow agree: 71/71 for both arms. - An earlier axuse run was disqualified under the strict rule for one stray
cua.getState()call (correct answer). The isolated Codex home removes that possibility for both arms. - An invalid first full run (Playwright arm never started because of a config error in the
new isolated home) is kept only as
results/INVALID-pw-v3-full-playwright-arm-never-started.jsonl.
Not covered
Public fixture pages with no login and no bot protection: this measures speed, tokens and accuracy on generic browser tasks, not axuse's other advantages (the user's signed-in Chrome, challenge handling, workflow plugins). Playwright MCP is also not slow here; the gap is a median of 3 s and 24k tokens per task.