Open-source MCP hub fronts 9 research repos with 14 tools and zero dependencies
Emergency-Tutor-7 · reddit · 2026-11-11
The author released scientific-toolkit-mcp 0.4.1 (Apache-2.0, Python 3.11+, zero runtime dependencies): a hub that exposes nine separate research repos (modeling engine, numerical stacks, literature/evidence server, security-reasoning gate, plan verifier, skill packages) as one MCP surface with 14 tools, while each repo keeps its own version history and install path.
Three defended design choices:
- Fail-closed discovery: missing upstreams or version mismatches fail with explicit errors (NOTINSTALLED, BADCONFIGURATION) rather than degrading to stubs; shadowed installs fail too, with an env pin for exact versions
- Per-tool local/network declaration: literature search needs network; quote verification and NomosGuard decisions run locally
- Authenticated HTTP gateway: anonymous calls get 401; the executable audit tool is never exposed over HTTP and requires explicit env opt-in even on stdio
The author is clear it's not a scientific validator: stats adapters stay independent, thresholds are uncalibrated defaults, and the gateway never invents missing validation outcomes. CI runs eight real adapter invocations plus a fail-closed smoke test.
More from coding & agent
- Cursor Projects adds multi-user collaboration with team-wide agent communication — aigclink · 2026-11-11
- Fable 5.2 lands and users already built 5 playable 3D games with it — FinanceYF5 · 2026-11-11
- Dev vows never to run coding agents locally, sparking workflow debate — verrsane · 2026-11-11
- Using Codex with Capwise CLI to Auto-Generate a Product Intro Video — vista8 · 2026-11-11
- Amazon blocks Scale AI's Muse browser agent, citing Conditions of Use — mohamedmansour · 2026-11-11
- Asking how much code agents wrote is like asking how much assembly the compiler generated — hargup13 · 2026-11-11