Loading ToolBench
Persistent agent desktop with browser automation, shell execution, and authentication attempt orchestration. Provides computer control tools (navigate, eval, action, exec) and MCP-based authentication/handoff management.
Four tools with clear action verbs and reasonable descriptions (avg 180 chars), but critical gaps in schema completeness and parameter documentation. computer_navigate, computer_eval, and computer_action have input schemas visible in case-tools.mjs, but parameter descriptions are minimal or missing. computer_exec lacks visible schema entirely. No output schemas documented. Error handling guidance absent. Naming is strong (verb_noun pattern), but parameter constraints and recovery guidance are missing.
UI action on the desktop (1280x800 by default): click|double_click|move|drag|scroll|type|key|wait. Coordinates are pixels, origin top-left. keys uses xdotool syntax (ctrl+l, Return). For elements INSIDE a web page prefer computer_snapshot + computer_click_element; use this for the desktop itself, canvas, shortcuts, and scrolling.
Evaluate JS in the active tab (CDP, promises awaited). Prefer this over screenshots for page content. Return plain values — DOM nodes are not serialisable. To read a page as prose, document.body.innerText. Do not drive location.assign from here; use computer_navigate.
Run a shell command on the computer (bash, as user agent).
Point the computer browser at url and block until the page has loaded. Returns {ok, url, title, text, snapshot} — text is the first 2000 chars of the page; snapshot holds the numbered elements, so do NOT call computer_snapshot after this. Same-page #anchor jumps are not navigations.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields without knowing response structure.
Parameter descriptions are minimal or absent. 'timeout_s' in computer_navigate and computer_eval lack descriptions; computer_action parameters (x, y, text, keys, dy, ms, from_x, from_y, to_x, to_y) have minimal guidance on when to use each.
No error handling guidance. Tools do not explain what to do on timeout, invalid URL, JS error, or command failure. Agents cannot self-correct.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 56 | 2026-07-28+ | v2 |
computer_action parameter 'type' is an enum but lacks description of what each action does (click vs double_click vs drag). LLMs must infer from name alone.