Model Context Protocol server exposing the curator API as native MCP tools for cognitive memory layer (typed graph + bi-temporal substrate)
kaeru-mcp exposes 10 tools with consistent naming (verb_noun pattern: awake, overview, pin, unpin, surface) and detailed descriptions (avg 150+ chars). All tools have input schemas with typed parameters and descriptions. However, output schemas are not documented in the visible code, responses are inferred from descriptions only. Parameter descriptions are strong (e.g., 'time window (e.g. '30m', '3h', '2d', or seconds)' for `since`), but some lack constraints (e.g., `layers` in `surface` accepts free-form strings with no enum or format validation). Error handling and recovery guidance are absent from tool definitions. No tool annotations (readOnlyHint/destructiveHint) despite clear risk classifications in metadata.
Restore session context: pinned set, recent episodes (24h), open reviews, plus the read-back of unfinished work — open tasks (overdue first), claims awaiting a verdict, and the saved reasoning trails. Run this when re-entering a project.
List the clouds this daemon can reach, with their endpoints and which one is default. Ask this before any cloud verb in an unfamiliar setup: with more than one cloud configured, `share` / `pull` / `cloud_recall` and the initiative verbs require `cloud` named explicitly, and nothing is routed to a default you did not choose.
Show resolved configuration: vault path, the configured clouds and which is default, and every cap (initiative not relevant).
Read FIRST before bulk-importing knowledge. Returns the import playbook: scope by initiative, pick the verb by epistemic status, stamp the memory layer at creation by importance, and link after capturing.
List initiatives that have at least one node attached. Use this first when re-entering, then pick one for subsequent calls.
Output schemas not documented. Tool descriptions explain what is returned (e.g., 'counts by tier/type, provenance forests'), but formal response schemas are absent from visible code. LLMs cannot plan downstream tool calls or extract structured fields without knowing the response shape.
No tool annotations (readOnlyHint/destructiveHint/idempotentHint) despite clear risk metadata. `pin` and `unpin` are marked WRITE but lack destructiveHint. READ_ONLY tools lack readOnlyHint. Annotations enable agents to reason about safety and retry logic.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 69 | <=2025-11-25 | v2 |
Print a terminal-readable map of the substrate: counts by tier/type, provenance forests, open questions, edge stats.
Pin a node to the active window. Accepts either a name or a UUIDv7 id.
List episodes whose latest assertion is within the time window (defaults 24h). Use `since` like `30m`, `3h`, `2d`, or raw seconds.
Explicit layered recall — surface nodes from specific memory layers (default `cold,frozen`) that `awake` does not load. Use when you deliberately need archived/not-surfaced material. `layers` is a comma/space list; scoped to `initiative` when given.
Unpin a node. Accepts name or id.
`surface` tool accepts `layers` as free-form string ('comma/space-separated list'). No enum constraint or validation. LLMs may pass invalid layer names. Should define valid layers as an enum or provide discovery tool.
No error handling or recovery guidance in tool definitions. If `pin` fails (node not found, invalid UUID), the description does not explain what to do next. Should include 'If node not found, call initiatives() or recent() to discover valid nodes.'
`recent` tool accepts `since` as free-form string with examples ('30m', '3h', '2d'). No regex pattern or formal constraint in schema. LLMs may pass invalid formats like '30 minutes' or '3hours'. Should add format constraint or enum.