lurq MCP Server

What moved in an MCP server's tool contract between two versions: tools removed, parameters that became required, types narrowed, and annotation flips. Two findings here have no npm equivalent and are why this exists. SILENT DRIFT is a schema that changed while its description stayed byte-identical, invisible to anyone reading a changelog. PRIVILEGE WIDENING is a tool that stopped being read-only or started being destructive, which does not break anything and is worse than a break. Use before upgrading a server an agent depends on.

Check whether a set of MCP servers can be wired into one agent together. The npm question does not apply — servers are separate processes with nothing to resolve between them. They clash in the single flat TOOL NAMESPACE the agent assembles from all of them: two servers exposing the same tool name leave the agent unable to express which it means, and nothing errors, one simply shadows the other. Also reports the standing context cost, since every tool's schema rides in every request. Pass `tools` for a server when you already hold its tool list (any server: remote, PyPI, Docker, private) and it is analysed as-is; otherwise the npm server's probed surface is used, and one that has not been probed makes the answer UNKNOWN, never clean.

What an MCP server ACTUALLY exposes: every tool, its required and optional parameters, and its behaviour annotations, read from a live `tools/list` handshake in a sandbox rather than from a README or the model's memory. Call before wiring an agent to a server, or when a tool call is failing for reasons the error does not explain. Also returns `requires` — the API keys and settings the server declares it needs — and `configRequest`, a ready-made line to put in front of your user when something is missing, so 'it needs a token' never presents as 'it is broken'. A miss returns UNKNOWN and queues a probe; UNKNOWN never means the server has no tools.

Assess an entire project's dependencies in ONE call: which packages are outdated, deprecated or carry advisories for the exact installed version, and which configured MCP servers have drifted, need credentials, or cannot be observed at all. Send the inventory you read locally (names and versions only — never source). Returns a per-item verdict plus an explicit coverage count: how many were answered, how many are queued because the index has not seen them, and how many were skipped and why. An item lurq could not assess is reported as unassessed, never as clean.

Look up which lurq tool answers a situation, and what to run next. Call when you're unsure whether lurq covers something (an upgrade, a licence rule, a version's exact exports, publishing a package) instead of guessing or skipping it. Returns matching capabilities with the tool or command to use.

Side-by-side comparison of 2–5 npm packages you are choosing between, ranked by health score. Untracked names are fetched on demand; one still being scored comes back under `pending` (retry shortly), and a name not on npm under `notFound`.

Check whether a set of packages forms a coherent stack. Resolves the whole set the way npm would and returns a definitive verdict: compatible, conflict (with the exact clashing constraints), or unknown. Checks peer-dependency and engine ranges from declared metadata, plus any recorded sandbox conflicts. Read-only: never installs or executes package code. Not always instant: a set checked before answers immediately, but a new set is resolved live from registry metadata, which can take up to ~25 seconds. Call once with the whole stack before committing to it.

Before wiring an MCP server into a client, find out whether it will work there and exactly what it takes. Accepts an endpoint URL, an official-registry server name, or an npm package name. Answers per client (Claude Code, Claude.ai, ChatGPT, Cursor, VS Code, Codex, Gemini CLI and more): WORKS; NEEDS_SETUP with the steps (a key to send as a header, an OAuth client to pre-register and the redirect URIs to allow, a URL placeholder to fill); BLOCKED with the reason and which side causes it; or UNKNOWN when a decisive fact is not established. Built from a credential-free probe of the endpoint (whether it answers, how it authenticates, which OAuth registration methods it offers, spec deviations strict clients refuse, tool names and schemas) and from primary-sourced client constraints. Returns ready-to-paste config in each client's own format, with placeholder values. Tell the user what it found, including which client it was checked for. UNKNOWN never means it will not work.

Emit a reference-architecture Mermaid diagram for a stack you have already chosen (package names). A labeled starting point keyed by layer, not a validated architecture, and not an architecture designer.

What changed in a package's runtime surface between two versions: symbols removed, added, and arity changes, plus renames the package itself proves (a removed name that shared one declaration with a name the new version still exports). Removals break `node`; type-only removals are returned separately because they break `tsc` instead. Answers "when did this stop working" from static comparison, with no install required. Use before an upgrade, and to explain a break after one.

Full evidence read for one npm package: health and quality scores and the signals behind them, advisories, the shared safety verdict, a summary and a usage guide. Use it when a choice needs more than verify's install gate. Enforces the account's dependency policy. A package lurq has never tracked is fetched and scored on demand, waiting up to ~4 seconds; if scoring takes longer the result is `tracked: false` with a note to retry in a few seconds, which is not an answer about the package.

The rules this account's selection policy enforces on which packages you may add: denied packages (with the reason), license allowlist, confidence, advisory, adoption, staleness and bundle-size floors. Read it once before choosing dependencies so you pick an allowed package first; evaluate already enforces it. Read-only.

Opt-in feedback after you act on lurq's evidence about a package (verify, evaluate, compare, compat): whether you went with it and whether it built. No source code, only the coarse decision + a build signal. Helps lurq learn which packages agents actually succeed with; safe to skip.

What a package version ACTUALLY exports at runtime, extracted from its shipped JavaScript rather than from documentation or the model's memory: names and arity, not type signatures (use usage for those). Call before writing code against a package whose API may have moved. Runtime existence is what decides whether an import throws; a removed type breaks tsc, a removed runtime symbol breaks the program. A miss returns UNKNOWN and queues extraction, UNKNOWN never means the symbol is absent.

A package version's TYPED API: exported symbols and their signatures from its shipped .d.ts (or DefinitelyTyped), exact to the version, none of it in the model's training data. Use before writing code against a package whose API may have moved. For whether a name exists at RUNTIME (what decides if an import throws) use resolve_surface; for is-it-safe-to-install use verify. Pass knownVersion (e.g. the version you were trained on) to get the precise delta: added, removed, renamed, changed. Also returns the version's declared engines (Node/runtime floor). Large surfaces are paged, 80 symbols per call: `totalSymbols` is the size, `query` filters by name, `offset` pages. `shallow: true` means the API lives on an interface's members that are not listed, and the note says where to read them. For framework file/convention changes (not exported symbols), consult the official migration guide / Context7 instead.

Confirm an npm package is real, healthy, and not risky before installing: guards against hallucinated or typosquatted names, deprecations and known advisories, and returns a safety verdict with its reasons and what was not checked. Checks the live registry. The cheap gate to call before every install; use evaluate when you need the full scores.