Loading ToolBench

Reading the current corpus…

puck MCP server · C · ToolBench

← Browse

puck

github.com/Laboratoriet/puck

Shared persistent-memory MCP server for Rocky + Claude clients

2026-07-28+STREAMABLE_HTTP4 tools

66/100

Definition Quality83

Protocol Readiness82

Supportability28

Spec posture

Inferred effective spec 2026-07-28+·current baseline 2026-07-28

Static source inference · medium confidence · evidence: stateless requests, Streamable HTTP

Current-spec patterns detected

Summary

Four tools with clear, well-structured definitions using Zod schemas. All tools have descriptions (194 - 250 chars, within baseline 34 - 392 range) and documented parameters. Naming follows verb_noun convention (recall, remember, forget, list_memories). Input schemas are properly typed with enums for scope. However, output schemas are not formally documented, responses are text-formatted strings rather than structured objects. Error handling is minimal (no recovery guidance or actionable error messages). Missing tool annotations (readOnlyHint, destructiveHint) despite clear risk profiles.

Tools (4)

forgetdestructivesource verified82/100

Delete a stored memory by its id (obtained from a prior recall) when it is wrong or outdated. To update a fact, forget the old one and remember the corrected one.

list_memoriesread onlysource verified80/100

List stored memories, optionally filtered by scope. Useful for review or when you want the full set rather than a search.

recallread onlysource verified85/100

Retrieve stored facts relevant to the user question. Call this whenever the answer may depend on remembered personal or household facts, preferences, people, places, routines, devices, or prior context. Returns matching memories with their ids.

rememberwritesource verified85/100

Store a durable, normalized fact about the household, a person, a place, a preference, a routine, a device, or a project. Use when the user asks you to remember something, or states a stable fact/preference worth recalling later. Store a concise fact, never a whole transcript. To UPDATE a fact, forget the old one then remember the new one.

Issues (5)

HIGHresponse-shaper ↗

Output schemas not documented. Tools return text-formatted strings (fmt() helper) instead of structured JSON objects. LLMs cannot reliably parse or chain results.

recallrememberforgetlist_memories

HIGHtool-annotations ↗

No tool annotations (readOnlyHint, destructiveHint, idempotentHint). 'forget' is destructive and 'remember' is write-only, but MCP spec annotations are absent.

forgetremember

MEDIUMrecovery-guide ↗

Error handling lacks recovery guidance. 'No memory with that id' is bare; should suggest 'Call recall() first to find valid memory IDs.'

forget

MEDIUMconstrained-input ↗

remember

Recommendations

Convert text-formatted responses to structured JSON. E.g., recall() should return {memories: [{id, scope, category, subject, fact}], total: N} instead of a formatted string.
Add tool annotations: recall/list_memories with readOnlyHint=true; remember with writeHint=true; forget with destructiveHint=true.
Formalize 'category' as an enum in remember() schema: z.enum(['preference', 'person', 'place', 'routine', 'device', 'project', 'misc']).
Add pagination to list_memories(): accept optional limit (1 - 100, default 20) and offset parameters; return {memories: [...], total: N, hasMore: boolean}.
Enhance error messages with recovery hints. E.g., forget() should return 'No memory with id "xyz". Call recall() to find valid IDs.' instead of bare 'No memory with that id.'

Score history

Overall score trend

First recorded score · v2 rubric

66/100

Scored	Grade	Overall	Spec posture	Rubric
2026-09-23	C	66	2026-07-28+	v2