Loading ToolBench
Agentic memory, skills, agentic execution, automation, and MCP tools for DeepSeek
Server provides 20 tools with complete input schemas and descriptions. Naming follows verb_noun convention (artifact_create, browser_navigate, browser_click). Descriptions are present and substantive (avg ~120 chars), exceeding the 10-char minimum. However, several critical gaps reduce quality: (1) artifact_bundle_create lacks detailed schema for the 'files' array items (no mimeType or content type constraints documented); (2) browser tools have minimal parameter descriptions (e.g., browser_click's 'button' param lacks enum constraint in description, only in schema); (3) no output schemas documented for any tool, LLMs cannot predict return structure; (4) error handling guidance is absent, tools do not indicate what to do on failure; (5) no idempotency or retry guidance for destructive operations (browser_close_tab, artifact_create). Artifact tools are well-described; browser tools are functional but sparse on parameter guidance.
Create a ZIP bundle from multiple generated text files. Use this for multi-file projects and scaffolds.
Create a downloadable UTF-8 text artifact or runnable code artifact. Use it for HTML previews and JavaScript, TypeScript, or Python runners. Never embed binary or base64 Office, PDF, image, or archive data; keep files produced by Shell or OfficeCLI at their local path and report that path.
Set local file paths to a page file input.
Click a page element by snapshot uid or CSS selector.
Close the specified tab, defaulting to the current controlled tab.
Execute a JavaScript expression in the controlled page and return a serializable result.
Fill an input control by snapshot uid or CSS selector.
No output schemas documented for any of the 20 tools. LLMs cannot predict return structure, field names, or types. This forces agents to guess what data is available after each call, increasing errors and wasted context.
Error handling guidance absent. Tools do not indicate what to do on failure (e.g., 'If selector not found, try browser_snapshot first'). Agents have no recovery path when calls fail.
artifact_bundle_create 'files' array lacks schema detail. No description of what each file object should contain (path, content, mimeType). LLMs cannot construct valid requests.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | 2026-07-28+ | v2 |
Fill multiple input controls at once, each supporting snapshot uid or CSS selector.
Let the controlled tab go back to the previous browsing history.
Let the controlled tab go forward to the next browsing history.
Accept or dismiss the current JavaScript alert/confirm/prompt dialog.
Move the mouse to a page element by snapshot uid or CSS selector.
Send a keyboard key to the page or a specified element.
List the controllable and non-controllable tabs in the current browser.
Open a URL in a new tab by default; set newTab=false only to replace the current controlled tab.
Refresh the current controlled tab.
Select a tab as the control target for subsequent browser tools.
Return the Accessibility Tree text snapshot of the current page and numbered interactive elements.
Insert text into the current focus or a specified element.
Wait until a selector, text, or JavaScript expression is satisfied.
browser_click and browser_hover have identical parameter descriptions despite different semantics. 'button' and 'clickCount' are irrelevant to hover; descriptions should clarify which params apply to which action.
Destructive operations (artifact_create, browser_close_tab) lack idempotency or confirmation guidance. No indication whether repeated calls are safe or if they risk duplicate artifacts/closed tabs.