Cardog MCP Server

Decode a 17-character VIN into its full Cardog identity: canonical entity refs, the market grains (nano/squish), spec highlights, and links to adjacent resources. VIN ONLY — this tool never fuzzy-matches. If you hold free text ("2021 Civic", a make or model name), do NOT call this: call resolve_entity — free text enters the platform in exactly one tool. A non-VIN input returns a redirect hint, not a decode. A ref is `{domain}:{key}`, lowercase, with `/` separating composite key segments: "make:tesla", "model:mini/hardtop", "model-year:honda/cr-v/2026", "fuel-type:electric". (Exception: nano/squish keys are uppercase VIN charset — machine-derived, never typed from text.) The result's `refs` block (make/model/modelYear/fuelType/…) contains the join keys for every other tool; a null ref means "not derivable for this VIN", never "unknown ref". `squish` (WMI+VDS+year) is always derivable and is a valid market_quote instrument; `nano` is the fungible build grain for dedup/comparables. `specHighlights` is a best-effort skim of the spec sheet (horsepower, economy, range, seating…), each value with its unit; `specHighlightsTrimDependent` names the highlights that differ between trims of the model year. The full sheet lives at GET /v2/specs/{refs.modelYear}. Errors are instructions: every failure returns {code, message, hint, suggestions} — follow `hint` for the next call; `suggestions` lists nearest valid refs for a bad ref. Unknown-but-well-formed refs are a 400 naming the ref, NEVER a silent fuzzy match. Next: check_recalls({ vin }) — outstanding recalls; market_quote({ ref: refs.modelYear ?? squish }); search_inventory({ models: [refs.model] }).

Search live Canadian vehicle listings — ref-native. One call returns listings + facets + the total count. Filters take entity REFS from resolve_entity / identify_vehicle, never free-text names: makes: ["make:mini"], models: ["model:mini/hardtop"], fuelTypes: ["fuel-type:electric"] — plus year/price/odometer ranges and canonical spec filters, e.g. spec: {"fuelEconomyCombined": {"min": 35}, "heatedSeatsFront": ["standard"]} (numeric attrs take {min,max}; equipment attrs take ["standard"|"optional"|"unavailable"]). A ref is `{domain}:{key}`, lowercase, with `/` separating composite key segments: "make:tesla", "model:mini/hardtop", "model-year:honda/cr-v/2026", "fuel-type:electric". (Exception: nano/squish keys are uppercase VIN charset — machine-derived, never typed from text.) Errors are instructions: every failure returns {code, message, hint, suggestions} — follow `hint` for the next call; `suggestions` lists nearest valid refs for a bad ref. Unknown-but-well-formed refs are a 400 naming the ref, NEVER a silent fuzzy match. A typo'd or unknown ref 400s with code "unknown_entity_refs" naming it, with nearest-ref suggestions — correct the ref (usually via resolve_entity) and retry. Facets in the result are (ref, name, count) buckets over the MATCHING set — they double as the valid filter vocabulary for your next, narrower call. Every listing row carries its refs (makeRef/modelRef/nano). Next: market_quote({ ref: "model-year:…" }) for pricing context; check_recalls({ vin }) per listing; GET /v2/listings/vin/{vin} for the full canonical spec.

The live market card for one instrument: quote (live listing count, best/p25/median/p75 price, average days-on-market, 30-day price cuts), a daily-bar history summary, and a bounded sample of the live listings behind the numbers. `ref` must be an INSTRUMENT ref — one of two grains: - "model-year:{make}/{model}/{year}" (e.g. "model-year:honda/cr-v/2026") — lowercase, /-separated; get it from resolve_entity (domain "model-year") or identify_vehicle's refs.modelYear. - "squish:{9 uppercase VIN chars}" (e.g. "squish:5TDGSKFCS") — the exact-config grain; get it from identify_vehicle. (squish/nano keys are the ONLY uppercase refs; every other domain is lowercase.) No other ref domain quotes. Errors are instructions: every failure returns {code, message, hint, suggestions} — follow `hint` for the next call; `suggestions` lists nearest valid refs for a bad ref. Unknown-but-well-formed refs are a 400 naming the ref, NEVER a silent fuzzy match. Optional `window` picks the history span: 1w, 1m, 3m, 6m, ytd, 1y, 3y, 5y, 10y, all. Next: search_inventory with the model's refs to walk the full book; check_recalls({ ref }) on a model-year ref; GET /v2/tape/history/{ref} for every daily bar.

The authoritative "is this vehicle under recall?" check — Transport Canada + NHTSA recall campaigns, fused and ref-keyed. Compliance guide: https://cardog.app/docs/compliance. Pass EXACTLY ONE of: - `vin` (17 characters) — the per-vehicle recall check. In the result, `resolved: false` means the VIN is not bridged into the graph yet — distinct from "no recalls" (`resolved: true, total: 0`). - `ref` — an entity ref scoping campaigns: "make:honda", "model:honda/cr-v", or "model-year:honda/cr-v/2026". A ref is `{domain}:{key}`, lowercase, with `/` separating composite key segments: "make:tesla", "model:mini/hardtop", "model-year:honda/cr-v/2026", "fuel-type:electric". (Exception: nano/squish keys are uppercase VIN charset — machine-derived, never typed from text.) Get refs from resolve_entity or identify_vehicle — never construct them from guessed names. Each campaign carries: authority (tc/nhtsa) + campaign number, component, defect/consequence summaries, the corrective action, recall date, units affected, and `affects` — the affected model-years as refs. `asOf` (VIN checks) is when the recall data was last updated, citable. Errors are instructions: every failure returns {code, message, hint, suggestions} — follow `hint` for the next call; `suggestions` lists nearest valid refs for a bad ref. Unknown-but-well-formed refs are a 400 naming the ref, NEVER a silent fuzzy match. Next: identify_vehicle({ vin }) for the vehicle's full identity; market_quote({ ref: "model-year:…" }); GET /v2/recalls/{recall-ref} for one campaign; GET /v2/recalls/feed for the newest campaigns.

Turn free text into canonical Cardog entity refs — THE text entry point for every other tool. A ref is `{domain}:{key}`, lowercase, with `/` separating composite key segments: "make:tesla", "model:mini/hardtop", "model-year:honda/cr-v/2026", "fuel-type:electric". (Exception: nano/squish keys are uppercase VIN charset — machine-derived, never typed from text.) Every other tool takes refs, never names. Call this FIRST whenever you hold text — "Civic", "2024 Model Y", a misspelling like "teslla" — then reuse the refs for the rest of the session. Returns candidates with confidence, best-first. `best` is the top candidate ONLY when it clears the confidence floor; otherwise it is null and YOU choose from `candidates` (or ask the user) — the API never guesses. Pass `domain` to constrain the search (use domain "model-year" when you need a market_quote instrument). Errors are instructions: every failure returns {code, message, hint, suggestions} — follow `hint` for the next call; `suggestions` lists nearest valid refs for a bad ref. Unknown-but-well-formed refs are a 400 naming the ref, NEVER a silent fuzzy match. Next steps (also echoed in each result's `next` block): search_inventory with make/model refs; market_quote with a model-year: ref; check_recalls with any make/model/model-year ref; dereference a ref (parents, children, counts) at GET /v2/entities/{ref}.