Loading ToolBench
AI Gateway & Chat Platform with voice mode, multi-provider support, and channel integrations
Tarsee provides 14 tools with complete input schemas and descriptions for all tools and parameters. Naming follows verb_noun convention (read_file, write_file, search_memories). However, descriptions are inconsistent in depth, some are detailed (browser, exec) while others are minimal (list_files, calculator). Parameter descriptions are present but often lack format constraints, ranges, or validation guidance. Output schemas are not documented. Error handling is minimal, no recovery guidance or actionable error messages visible in tool definitions. Security concerns: exec tool allows arbitrary shell commands with only a 60-120s timeout; no permission gates or audit trail visible. The remember tool's category enum is good, but most tools lack similar constraints. Overall solid foundation but gaps in error guidance, output documentation, and security hardening.
Append content to a workspace file without overwriting existing content. Useful for adding entries to MEMORY.md, TOOLS.md, etc.
Control a stealth web browser (anti-detection enabled). Actions: navigate (go to URL), screenshot (capture page), click (click element), type (type text), evaluate (run JS), get_text (extract text), wait_for (wait for selector/navigation), scroll (scroll page), select (dropdown), solve_captcha (auto-detect and solve reCAPTCHA/hCaptcha/Turnstile via 2Captcha or Capsolver). Browser persists across calls.
Evaluate a math expression with precision. Use this for ANY math: arithmetic, percentages, unit conversions, financial calculations, etc. Never do math in your head — always use this tool. Supports: +, -, *, /, **, %, Math.sqrt(), Math.round(), Math.ceil(), Math.floor(), Math.abs(), Math.log(), Math.PI, Math.E, parentheses.
Append a timestamped note to today's daily memory log (memory/YYYY-MM-DD.md).
Edit a workspace file by replacing specific text (targeted find-and-replace). Only replaces the first occurrence.
exec tool allows arbitrary shell commands with minimal safety constraints. No permission gates, no dry-run capability, no audit trail visible. Timeout alone (60-120s) is insufficient for destructive operations.
Output schemas not documented for any tool. LLMs cannot plan downstream calls or extract required fields. E.g., search_memories returns results but structure is undocumented.
No error handling guidance. Tools lack recovery hints (e.g., 'file not found, try list_files()'), error classification (retryable vs fatal), or actionable error messages with constraints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 66 | 2026-07-28+ | v2 |
Execute a shell command on the server. Use for system tasks, checking status, running scripts, or using installed tools like Playwright. Commands run in a sandboxed environment with a 60-second timeout. Output is capped at 50KB.
Search for a text pattern across workspace files. Returns matching lines with filename:lineNumber format. Case-insensitive. Results capped at 50 matches.
List all workspace files and their sizes. Also lists daily memory logs.
Read the contents of a workspace file (SOUL.md, IDENTITY.md, AGENTS.md, USER.md, TOOLS.md, MEMORY.md, HEARTBEAT.md, BOOT.md) or a daily memory log (memory/YYYY-MM-DD.md).
Save an important fact or preference to long-term memory. This appends to the MEMORY.md file and saves to the database.
Search through stored memories in the database for relevant facts about the user or past conversations.
Deep semantic search — dumps ALL stored memories, MEMORY.md, and recent daily logs so you can reason about relevance. Use this when search_memories returns nothing or when the user asks about something vague that keyword search can't match. Costs more tokens but finds things word search misses.
Fetch the contents of a URL. Returns the response body as text (HTML, JSON, etc). Useful for checking APIs, fetching web pages, or downloading data.
Write or overwrite the contents of a workspace file. Use this to update your personality (SOUL.md), identity (IDENTITY.md), user notes (USER.md), long-term memory (MEMORY.md), tools (TOOLS.md), or agent rules (AGENTS.md).
Parameter constraints missing. exec timeout_ms lacks min/max bounds; web_fetch method enum is present but other tools lack similar constraints. No regex patterns or length limits documented.
browser tool's solve_captcha action references external services (2Captcha, Capsolver) but no API key injection pattern documented. Credentials may leak into logs if passed as parameters.