feat: kiwi MCP for corpus DB access #124

New issue

Open

opened 2026-04-27 11:31:55 -07:00 by pyr0ball · 0 comments

pyr0ball commented

2026-04-27 11:31:55 -07:00

Owner

Problem

Debugging browse/tag issues requires docker exec python3 -c "..." one-liners against the SQLite corpus. This is slow, error-prone, and the output disappears.

Proposed solution

Build a kiwi MCP server (similar to the existing Snipe MCP pattern) that exposes:

kiwi_query_corpus — run a read-only SQL query against the corpus DB, return results as JSON
kiwi_count_fts — run an FTS5 MATCH expression and return the row count
kiwi_sample_tags — return tag frequency distribution for inferred_tags, cuisine:X, meal:X, dietary:X etc.
kiwi_browse_preview — call the browse endpoint with given domain/category and return first-page results

DB path (inside container): /app/kiwi/data/kiwi.db
Browse counts cache: /app/kiwi/data/browse_counts.db

Benefits

Future tag/keyword audits take minutes instead of sessions
Can verify browse counts in-conversation without rebuilding containers
Documented access patterns become reusable tools

#122 Browse: meal_type categories near-empty
#123 Browse: audit and enrich domain keyword lists

Notes

Read-only only — no write access to corpus from MCP
Must work against both local dev DB and cloud per-user DBs

## Problem Debugging browse/tag issues requires `docker exec python3 -c "..."` one-liners against the SQLite corpus. This is slow, error-prone, and the output disappears. ## Proposed solution Build a kiwi MCP server (similar to the existing Snipe MCP pattern) that exposes: - `kiwi_query_corpus` — run a read-only SQL query against the corpus DB, return results as JSON - `kiwi_count_fts` — run an FTS5 MATCH expression and return the row count - `kiwi_sample_tags` — return tag frequency distribution for `inferred_tags`, `cuisine:X`, `meal:X`, `dietary:X` etc. - `kiwi_browse_preview` — call the browse endpoint with given domain/category and return first-page results DB path (inside container): `/app/kiwi/data/kiwi.db` Browse counts cache: `/app/kiwi/data/browse_counts.db` ## Benefits - Future tag/keyword audits take minutes instead of sessions - Can verify browse counts in-conversation without rebuilding containers - Documented access patterns become reusable tools ## Related - #122 Browse: meal_type categories near-empty - #123 Browse: audit and enrich domain keyword lists ## Notes - Read-only only — no write access to corpus from MCP - Must work against both local dev DB and cloud per-user DBs