Local, observable computers for any compatible AI agent. Provides MCP tools for desktop control, file system access, terminal execution, browser automation, and system management.
Qubicl demonstrates solid tool definition quality with comprehensive schemas and clear descriptions. All 18 tools have explicit input schemas with proper JSON Schema structure, type definitions, and parameter constraints. Descriptions are generally actionable (avg ~120 chars), explaining WHAT the tool does and WHEN to use it. Lease-based ownership model is well-documented. However, output schemas are not visible in source code, parameter descriptions are sparse or missing for many fields, and error handling guidance is absent. Tool naming follows verb_noun convention consistently. No security-sensitive parameters exposed as tool inputs (lease tokens are structured objects, not raw secrets). Composition is sound, tools are single-responsibility and chainable via lease/path references.
Acquire attributable agent ownership. Optionally wait in FIFO order when another client owns it; observations never need ownership.
Perform an action in the browser: click, type, scroll, navigate, or wait.
Capture the current browser viewport as a PNG screenshot with element references.
Delete a file or directory.
Perform an action on the desktop: click, type, keypress, scroll, drag, move, or wait.
Capture the current desktop as a PNG screenshot.
Edit a file by replacing exact text matches.
Output schemas not documented in source code. LLMs cannot plan downstream tool calls or extract required fields without knowing response structure.
Parameter descriptions missing or minimal for many fields. E.g., 'detail' enum values lack explanation of what 'compact' vs 'full' returns; lease object properties undocumented.
No error handling guidance in tool descriptions. LLMs cannot determine if failures are retryable, user-fixable, or fatal. E.g., exec_command does not document what happens on timeout or permission denial.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 69 | 2026-07-28+ | v2 |
Run a managed command with bounded output. Combined output is default. Tasks survive disconnect and human takeover. Explicit services also restart with the computer until stopped. Session commands are fenced with the lease.
Explain whether a tool is supported, enabled, authorized for this client, and temporarily blocked by ownership.
Return status and effective cgroup limits; full detail adds contracts and diagnostics.
List files and directories in a directory.
Read a file from the computer's filesystem.
Release agent control and stop its managed processes.
Renew an active lease explicitly; tool activity also refreshes it.
Invoke a skill with optional parameters.
Create, update, delete, enable, or disable a custom skill.
List enabled skills and their descriptions.
Write or create a file in the computer's filesystem.
Destructive operations (delete_file, skill_mutate with action='delete') lack confirmation or dry-run support. Agents can permanently delete files without safeguards.
Tool annotations (readOnlyHint, destructiveHint, idempotentHint) not visible in schema definitions. Risk classification is inferred from tool names and descriptions, not explicit metadata.