Loading ToolBench
MCP server over your second brain — so Claude Desktop (or any MCP client) can search, read, and add to the brain over a local stdio connection. Everything stays on your machine. Exposes search, fetch, wiki, topics, recent, by_series, overview, source_status, list_agents tools, plus write tools (ingest_note, save_chat) and agent playbooks as MCP prompts.
Second Brain exposes 15 tools with mostly clear naming (verb-noun pattern: search, fetch, wiki, topics, recent, by_series, overview, source_status, list_agents, ingest_note, save_chat) and reasonable descriptions (50-150 chars typical). Input schemas are visible and typed (string, integer parameters with descriptions). However, several tools lack output schema documentation, and some descriptions are generic. Prompts (research_brief, interview_prep, caption_pack, weekly_review) are present but lack visible input schemas. Write tools (ingest_note, save_chat) have good descriptions but no confirmation/dry-run pattern. Error handling is not visible in the code sample.
Get items from a named content series.
Agent playbook prompt: caption generation recipe the client model executes with the read tools.
Read one search result in full after search.
Save an idea to the brain. Always ask the user first: never write silently.
Agent playbook prompt: interview preparation recipe the client model executes with the read tools.
Auto-detected roster of everything built on the brain, with per-surface invocation hints. Check it BEFORE telling the user a capability doesn't exist.
Prompt tools (research_brief, interview_prep, caption_pack, weekly_review) lack visible input schemas. Prompts should declare expected parameters and output format so clients can invoke them correctly.
Write tools (ingest_note, save_chat) lack confirmation/dry-run pattern. Agents should be able to preview changes before committing irreversible writes. Current descriptions say 'Always ask the user first' but no tool-level confirmation mechanism is visible.
Output schemas not documented in code sample. Tools like search, fetch, wiki, recent, by_series, overview, source_status, list_agents return results but the response structure (fields, types, pagination) is not visible. LLMs cannot plan downstream calls without knowing what fields to expect.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
Get stats about the brain itself — not for content questions.
Get what's new/latest in the brain.
Agent playbook prompt: research recipe the client model executes with the read tools.
Archive this conversation to the brain. Always ask the user first: never write silently.
Search the brain by meaning — the default door for any 'what have I said/made/thought about X' question.
Observability tool: when did my sources last sync / is anything stale.
List all synthesized wiki topics that exist.
Agent playbook prompt: weekly review recipe the client model executes with the read tools.
Read the user's synthesized take on a broad topic (synthesized knowledge pages).
Error handling and recovery guidance not visible in code sample. Tools should return actionable error messages (e.g., 'Topic not found. Available topics: [list]') rather than generic failures.
Pagination parameters (cursor, k) present in search and recent, but no documentation of max limits or total count return. search accepts k up to 50 (mentioned in description) but recent and by_series lack explicit bounds.