Loading ToolBench
AIOS compatibility MCP server for browser automation with profile and CDP support
Mixed quality across 10 tools. Memory and skill tools (aios_memory_*) have clear, action-oriented names and good descriptions (80-150 chars), with well-typed schemas. Browser auth tools (browser_login, browser_check_auth, browser_configure_credentials) have solid descriptions and enum constraints on 'site' parameter. However, aios_judge has an exceptionally complex schema with nested object definitions that may confuse LLMs; aios_codemap_search lacks specificity on what 'code references' means. Most tools lack explicit output schema documentation. Error handling guidance is minimal across all tools.
Search codebase structure and code references (read-only codemap) before editing. Use to locate files, symbols, and callers.
Ask the opt-in System One (Jev) judgment gate a structured question and get back a typed verdict. This is a GATE, not a generator: it can only narrow an action (act / confirm / abort), never author content. Disabled by default. The tool only appears when the operator ran `aios judgment enable <vendor>` AND the vendor credential is present in the environment. When disabled, ask nothing here and fall back to your own declared judgment. Every answer is a proposal with a confidence and a vendor request id; never treat an answer as a verified fact about the repository.
Pin a milestone checkpoint to AIOS workspace memory so future sessions recall it. Use at milestones and before claiming work complete.
Search AIOS project memory before continuing prior work. Use when the user says continue/resume or the task builds on earlier conclusions.
Mark recalled AIOS memo event(s) as adopted so future recall ranks them higher.
aios_judge has overly complex nested schema with 'questions' as additionalProperties object. LLMs struggle with deeply nested, dynamically-keyed structures. Simplify to a flat array of question objects with explicit type/instructions/options fields.
No tool documents its output schema. LLMs cannot plan downstream calls or extract required fields without knowing what each tool returns. Add explicit 'Returns' sections to all tool descriptions.
aios_codemap_search description is vague: 'code references' is undefined. Does it return file paths, function signatures, call graphs, or all three? Clarify what structure is returned and when to use it vs aios_skill_search.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 62 | 2026-07-28+ | v2 |
Persist a verified conclusion, fix, or preference to AIOS project memory. Only confirmed facts, never guesses.
Search AIOS skills and client guidance for the current task before building something new.
Check whether the browser has an active authenticated session for a site.
Guide for setting up credentials for a site. Opens the login page for manual authentication. After the user completes login manually, the session is preserved in the browser profile.
Log into a supported site using credentials stored in macOS Keychain. Credentials are NEVER visible to the LLM — they are resolved and injected entirely server-side.
aios_memory_useful and aios_memory_checkpoint lack guidance on when to call them. Should aios_memory_useful be called after every successful recall? Should checkpoint be called only at major milestones? Add dependency hints.
No error handling guidance across any tool. If aios_memory_recall returns no results, should the LLM try a broader query or proceed without memory? If browser_login fails, is it retryable? Add recovery hints to descriptions.