Loading ToolBench
MCP server for Terminal Brain - a non-launching operator interface providing runtime status, snapshots, audits, and work-block management for a macOS app-based brain system.
Terminal Brain MCP server has 23 tools with consistent naming (verb_noun pattern: terminal_brain_*) and descriptions present for all tools. However, critical gaps severely limit production readiness: (1) ALL tools have empty input schemas (additionalProperties:false, no properties defined), this violates the schema requirement and prevents LLMs from understanding parameter constraints; (2) descriptions are verbose (200-300+ chars) but lack actionable structure, they read as marketing copy rather than LLM-optimized docstrings; (3) parameter descriptions exist for some tools (e.g., terminal_brain_use_now_markdown has 'limit', 'project', 'idea', 'title' with descriptions) but many tools have zero parameters documented; (4) no output schemas are visible, LLMs cannot predict what fields to expect; (5) no error handling guidance, tools fail silently or return unstructured markdown; (6) no security annotations or permission gates despite write operations (terminal_brain_use_now_markdown, terminal_brain_check_in_markdown). The server reads as a thin wrapper around shell scripts rather than a production-grade MCP tool set.
Run the non-launching Terminal Brain capability audit and return evidence for value, MCP, safety, readiness, and first commands as Markdown.
Read the non-launching Bubble Up brief: neglected ideas, delegated loops, repeated project pressure, and exact triage commands.
Run the non-launching guided clean-queue check-in. Without an idea, returns three plain prompts and exact commands; with an idea, captures it into Oracle Inbox memory.
Get a non-destructive cleanup plan for stale Terminal Brain MCP/kernel runtime noise. Prints candidates and manual review commands without killing anything.
Run the non-launching world-class completion audit: objective, prompt-to-artifact checklist, evidence, current state, and explicit uncertified visual-review gap.
Run a non-launching temporary Terminal Brain demo with seeded ideas, Review Queue, Bubble Up, Work Block, and real-use commands.
ALL 23 tools have empty input schemas (only additionalProperties:false, no properties defined). LLMs cannot infer parameter constraints, types, or valid values. This violates the schema requirement and prevents proper tool invocation.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 40 | 2026-07-28+ | v2 |
Run the non-launching Terminal Brain readiness doctor and return app install, MCP contract, agent config, process, launchctl, and API status as Markdown.
Get one non-launching first-minute artifact: what Terminal Brain is, what value is available, what to do first, and a working closed-app proof.
Get the safest next move as Markdown. Returns Start Here when the app is reachable, otherwise returns non-launching runtime status and the manual next step.
Get the structured app-backed Terminal Brain Now payload with bottom line, focus, do-this steps, process truth, and close loop. Requires the app API to be reachable.
Get the fastest non-launching Terminal Brain orientation: bottom line, next action, process truth, readiness, and outcome close loop.
Get the non-launching operator playbook: what command to run for common situations, first five-minute loop, daily cadence, agent cadence, and readiness.
Get a non-launching process map for Terminal Brain, Codex sessions, MCP children, brain-kernel children, brain-console helpers, Drafts, launchctl, and API reachability.
Read the non-launching Oracle Inbox review queue as Markdown without opening Terminal Brain.
Read non-launching runtime status for agents: repo, latest CI, local process, launchctl, and API reachability without requiring the app to be running.
Get a non-launching Markdown support bundle with What Now, Now, Doctor, Audit, Process Map, Cleanup Plan, and Git state.
Get the one-command non-launching Terminal Brain path for a new or overwhelmed operator: one executable move first, compact pull-forward context, ask, capture, delegate, and close the loop.
Run the non-launching value audit: success criteria, prompt-to-artifact checklist, evidence, current state, and remaining gaps.
Get the current Terminal Brain value read as Markdown. Returns the live Value Brief when reachable, otherwise explains the value path and safe next commands.
Run the non-launching Terminal Brain value proof: Oracle Brief, Agent Prompt, temporary accepted outcome note, and note preview.
Get the non-launching manual visual review plan for certifying the remaining native UX gate when the operator explicitly opens the app.
Get a concise non-launching situation read for humans and agents: what is running, repo/CI state, runtime noise, current blocker, and the next value command.
Get one non-launching work block: pull forward signals, review queue, and close-loop command shape.
No output schemas documented for any tool. LLMs cannot predict response structure, field names, or types. This forces LLMs to parse unstructured markdown responses, wasting tokens and increasing hallucination risk.
Write operations (terminal_brain_use_now_markdown, terminal_brain_check_in_markdown) lack error handling guidance, confirmation steps, or permission gates. No recovery instructions if capture fails. Agents cannot distinguish retryable from fatal errors.
Descriptions are verbose (200-300+ chars) and read as marketing copy rather than LLM-optimized docstrings. They lack actionable structure: WHAT does it do, WHEN to use it, WHAT does it return. Example: 'Get the fastest non-launching Terminal Brain orientation: bottom line, next action, process truth, readiness, and outcome close loop' is vague about what fields are returned.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) present. LLMs cannot distinguish safe read-only tools from destructive ones. This is especially critical for write operations like terminal_brain_use_now_markdown and terminal_brain_check_in_markdown.