MCP server that gives Claude Code persistent, feature-scoped knowledge across sessions
kb-recall has 7 well-named tools with clear verb-noun patterns (list_, search_, load_, save_, update_, init_, report_). All tools have descriptions (avg ~60 chars, below the 194-char baseline but acceptable for domain). Input schemas are present with type declarations and parameter descriptions. However, output schemas are not documented in the source code, LLMs cannot predict what fields to expect from responses. Error handling is minimal; no recovery guidance or actionable error messages visible. The server lacks pagination support despite search_features potentially returning many results. No tool annotations (readOnlyHint/destructiveHint) despite clear risk stratification (4 WRITE tools, 3 READ_ONLY).
create a new feature KB directory
list all features across configured projects
load full context for a specific feature
record a context miss to memories.md
prepend an insight to a feature's memories.md
keyword search across all features' README + memories
replace or append content in a README section
Output schemas not documented. LLMs cannot predict response structure (fields, types, pagination). search_features and load_feature_context likely return complex objects but no schema is visible in source.
No tool annotations despite clear risk stratification. 4 WRITE tools (save_memory, update_readme, init_feature, report_miss) lack destructiveHint; 3 READ_ONLY tools lack readOnlyHint. Agents cannot distinguish safe from unsafe operations.
search_features lacks pagination. Code shows MAX_SEARCH_RESULTS=20 and SEARCH_RESULT_CHAR_LIMIT=20_000, but no limit/offset/page parameters in schema. Large result sets will blow context window.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
No error recovery guidance. Code references KB_CONTEXT_HARD_LIMIT_CHARS=40_000 and hints like '/recall:compact' and '/recall:tidy', but these are not exposed as tool outputs or error messages. LLMs cannot self-correct.
Parameter descriptions are terse (avg ~30 chars vs 72-char baseline). 'optional project slug to filter by' lacks context on what happens if omitted (falls back to session-detected project). Dependency hints missing.