Voice agent for Windows. You talk, it does the thing. Provides MCP server interface to Wilco's tools over stdio.
Wilco provides 10 tools with reasonable naming (verb_noun pattern) and detailed descriptions (100-300 chars, well above the 34-char minimum). However, critical gaps undermine production readiness: input schemas are present but lack explicit type definitions for most parameters (only 'type': 'string' or 'type': 'boolean' visible; no enums, patterns, or constraints). Parameter descriptions exist but are often vague (e.g., 'Application name to launch' lacks guidance on format or discovery). No output schemas are documented, LLMs cannot predict what fields to expect. Error handling is absent from visible code; no recovery guidance or categorization. Security concerns: run_powershell and run_bash expose command execution with minimal guardrails beyond regex-based heuristics, and no audit logging visible. Tool composition is reasonable (single responsibility per tool), but critical tools like run_powershell lack confirmation patterns despite destructive capability.
Close a whole running application and every window it owns. Asked politely, so anything with unsaved work still shows its own save prompt. Leave name empty to close whatever is in front. This closes the ENTIRE app — to close one browser tab use close_tab, which is almost always what 'close this' means when a browser is involved.
Close ONE browser tab with Ctrl+W, leaving the rest of the browser open. This is what 'close this tab' means. app: which browser, if they named one — otherwise the frontmost browser is used. Never use close_app for a tab; that would shut the whole browser.
Close a single window with Alt+F4, leaving the app's other windows open. app: part of the window title, or empty for the window in front.
Bring a window to the front so the next typing or keypress goes to it. Give any part of its title — 'chrome', 'word', 'youtube'. ALWAYS call this before type_text when the user names where the text should go.
List installed apps, optionally filtered by a word. Use when the user asks what is installed, or when you need the exact name of something before opening it.
No output schemas documented. LLMs cannot predict return structure, forcing them to guess field names and types. This breaks tool chaining and increases hallucination risk.
Input schemas lack enums, patterns, and constraints. Parameters like 'browser' in open_browser_window accept free-form strings instead of constrained enums (edge|chrome|firefox|brave|opera|vivaldi). This invites hallucinated values.
run_powershell and run_bash rely on regex-based heuristics (READ_ONLY, CHANGES, POWERSHELL_CRITICAL) to gate destructive commands. This is fragile: regex can be bypassed, and no explicit confirmation pattern exists. Destructive operations should support dry-run or explicit confirmation.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 48 | 2026-07-28+ | v2 |
List the windows currently open on screen. Use this to find out what the user is looking at before typing into something, or when they say 'this window' or 'the browser'.
Launch an installed application by its spoken name — 'notepad', 'chrome', 'vs code'. Works for desktop and Store apps. If several match, they are listed instead of guessing: ask the user which one, then call this again with the fuller name.
Open a NEW browser window, private or ordinary. THIS is the tool for "incognito", "InPrivate", "private window", "private browsing", "secret tab" — never go at it through the browser's menu. Works whatever has focus.
Search the file system with find, grep, ls, wc, du, sort and pipes. This is how you look for files by name, pattern or CONTENT, across any file type.
Run PowerShell on this machine — the escape hatch for services, processes, network config, registry, scheduled tasks, installed packages and hardware. Routine commands run immediately; destructive, security-sensitive, install, power and credential-revealing commands are parked only while confirmations are on; all-access mode (the default) runs them. Prefer a purpose-built tool when one fits. For screen brightness or speaker volume always use set_screen_brightness / change_volume — never WMI classes recalled from memory, which usually don't exist on the machine.
No error handling guidance visible. Tools do not document what errors are retryable, user-fixable, or fatal. No recovery suggestions (e.g., 'App not found. Try list_installed_apps() to see available options.').
Parameter descriptions lack actionable constraints. 'Application name to launch' does not specify format, case sensitivity, or whether partial matches work. 'part of the window title' is vague, does it match substring, regex, or exact?