Passthrough MCP server that materializes tool results into a DuckDB scratch database
Wellknown found it in public sources; nobody has proven control of it yet. Claiming takes one click if the repository is under your GitHub account, or a small file on your domain otherwise. Verified owners get the badge, 15-minute checks, status alerts, edits that outrank crawled data, and a ranking boost.
Agents can do it too: POST https://wellknown.network/api/v1/claims with {"agent":"mcp-sluice","method":"well_known_file"} — machine-readable steps at claim.json, guide at /docs/claim.
Everything here was measured by our prober or read from a registry. Nothing is self-reported.
Attributed to the source that supplied each field. Treated as claims, not facts.
# Sluice A passthrough MCP server. It sits between an MCP client and **exactly one** downstream MCP server, mounts the downstream tools, and proxies calls through. Every eligible result is materialized into an in-memory DuckDB scratch database before being returned, so the agent gets a short preview plus a table handle instead of the payload, and runs SQL over it with Sluice's own `query` tool. Two things this buys: aggregation becomes exact rather than a model reading rows out of a payload, and an eligible payload is replaced by a bounded handle, so it is not re-sent on later turns of the agent loop. Error, binary, and over-ceiling results deliberately pass through unchanged. ## Status The current release, [v0.2.1](https://pypi.org/project/mcp-sluice/0.2.1/), is available from PyPI. Do not use 0.2.0 with DuckDB 2.0: its query gate can be bypassed through a CTE (see `CHANGELOG.md`). Proxying, envelope/handle recording, flattening and type inference, the read-only `query` gate, aggregate-correctness properties, bounded runtime retention, final adversarial review, and bounded pagination for approved read-only tools (spec 002) are implemented and covered by the test suite. The reproducible live-model demo recorded one baseline miss (71.5) and one Sluice-backed exact answer (72.5); that sampled run is evidence, not a deterministic guarantee. CI runs the full test, lint, formatting, strict type-check, and package-build gates. The end-to-end canary against a real configured GitHub MCP server passed; its evidence is in `review/canary-v0.1.0.md`. The project is distributed as `mcp-sluice`; the Python import package and console command remain `sluice`. Read `spec/001-scratch-db.md` before changing behavior; `intent/` records why the non-goals are non-goals; `plan/001-notes-m0.md` holds the measured third-party behavior the design rests on. **Behavioral baseline verified against:** MCP protocol `2026-07-28`, SDK `mcp` 2.1.1, DuckDB 2.0.0 and 1.5.5, CPython 3.14. Behav…
Mapped onto the structured taxonomy from declared text and observed tool names. Confidence shown for derived entries.
Every source is kept verbatim. Field changes are logged as events.