Loading ToolBench
A free open-source thinking partner for coding agents — mediates dialogue between Claude Code, Codex CLI, Cursor, Cline, and more via MCP. Exposes three thinking-partner skills (consult, perspective, assumptions) as MCP tools so any MCP-aware coding agent can use them.
Quill MCP has 4 tools with clear, domain-specific purposes (reframing, perspective-taking, assumption translation, multi-voice decomposition). All tools have descriptions (100-250 chars, within baseline 194 avg). Input schemas are present and typed. However, parameter descriptions are minimal (single-line, under 50 chars), output schemas are not documented, and error handling guidance is absent. Tools lack actionable recovery hints and field-level documentation. The server targets a specialized use case (thinking partner for developers) rather than general-purpose operations, which affects composition patterns.
Translate the technical assumptions you've been making into plain language. Use when the developer might be making decisions they don't understand because of technical jargon. Pass an enumeration of the load-bearing technical assumptions baked into your current approach. Returns a 3-5 item plain-language yes/no checklist the developer can actually answer.
Reframe a stuck or frustrated moment from a humanistic perspective. Use when the developer seems stuck, frustrated, or when the conversation feels like it's spinning. Pass your framing of what's going on (2-4 sentences). Quill's advisor returns a 2-4 sentence reframing — surface it to the developer alongside your own view.
Produce a mosaic response to a multi-aspect task — two heads, not one. Decomposes the task into 2-4 voice-assigned slices, executes them in parallel with independent priors preserved (no agent sees the others' WIP), cross-reviews for consistency without homogenizing voice, and returns a structured response that *surfaces* rather than smooths the seams between agents. Best for multi-texture work where different aspects benefit from different voices: code + docs + UX + tests, or backend + frontend + migration plan, or methodology + findings + caveats. NOT for single-aspect tasks — the overhead isn't worth it. Cost: ~4-6x a single quill_consult call; wall-clock 60-180s. Requires both `codex` and `claude` CLIs installed and logged in (free with ChatGPT Plus + Claude Pro subscriptions).
Output schemas not documented. LLMs cannot plan downstream operations or extract structured data from responses. Quill tools return advisor text, but field names, types, and structure are not declared.
Parameter descriptions are minimal (single sentence, <50 chars). Rubric baseline is 72 chars avg. 'Your framing of what's going on' lacks guidance on tone, length constraints, or what constitutes a valid framing.
No error handling or recovery guidance. If the advisor backend fails, times out, or returns malformed output, the tool provides no actionable next steps for the LLM.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 54 | <=2025-11-25 | v2 |
Layer in another vantage point alongside your current thinking. Use when the developer is curious or exploring (NOT stuck) and would benefit from a perspective you haven't taken. Pass your framing of your current angle. Returns a 2-4 sentence perspective that's additive, not corrective.
quill_mosaic has high latency (60-180s wall-clock) and cost (~4-6x single call) but no timeout, cancellation, or progress reporting. Agents cannot monitor or interrupt long-running operations.
Tool names use underscores (quill_consult, quill_perspective) but lack action verbs. Names like 'consult', 'perspective', 'assumptions' are nouns, not verbs. Better: 'reframe_thinking', 'add_perspective', 'translate_assumptions'.