360 Secure SkillHub
360 安全技能中心

onehome

该技能驱动 onehome-mcp 的 onehome_* 工具,在 CoreLogic OneHome 私人门户中按分组、已保存搜索或自由文本检索房源,拉取详情、照片、学校与步行评分,支持 2-8 套房源对比,并可在本地计算房贷月供与可负担性。认证依赖 ONEHOME_TOKEN 或 ONEHOME_MAGIC_LINK(或从已登录门户标签页捕获 token)。

chrischall
chrischall
热度 63.29
最近更新 2026-10-10 07:30:05
分类:搜索依赖:MCP 服务 / 第三方 API / 联网浏览搜索应用场景:研究 / 推荐

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add onehome
Metadata
name
onehome
description
Search OneHome (CoreLogic) portal listings, get property details, photos, schools, saved searches. Use when the user asks about real estate listings shared by their agent, OneHome links, portal.onehome.com properties, or specific addresses / MLS numbers they want to look up.

OneHome MCP

OneHome is a CoreLogic product. Buyers reach it through a private magic-link URL their real-estate agent emailed (https://portal.onehome.com/.../?token=...). It's an Angular SPA backed by a single GraphQL API at services.onehome.com/graphql.

This skill drives the onehome-mcp tools (onehome_*) — all GraphQL-bound except for the local PITI / affordability calculators.

When to use

  • The user mentions OneHome, a portal.onehome.com URL, or a CoreLogic magic-link they got from their agent.
  • They ask about listings in a specific "group" their agent set up (Lake Lure, downtown condos, etc.).
  • They want to look up a listing by MLS number or address and the other realty MCPs returned nothing.
  • They want photos, school info, or to compare two OneHome listings.

If the user asks generally about a property without naming OneHome, prefer one of the public-listing MCPs (zillow / redfin / homes / compass) first.

Auth modes

The MCP picks one at startup:

  1. ONEHOME_TOKEN (env): raw bearer JWT.
  2. ONEHOME_MAGIC_LINK (env): full URL with ?token=. The token IS the bearer.
  3. fetchproxy capture: capture from a signed-in portal.onehome.com tab.

Run onehome_healthcheck first to confirm auth is wired up — it returns the mode, token expiry (when known), and a one-line hint.

Workflow

  1. Pick a group. onehome_get_groups → the curated buckets the agent shared with the buyer. Most tools take a group_id.
  2. Search inside the group. onehome_search_properties { group_id, [saved_search_id] }. Pass a saved_search_id from onehome_get_saved_search to apply a saved filter; leave it off for the group default. onehome_get_saved_search_with_listings returns the saved search and its inflated listings in one call.
  3. Inflate a listing. onehome_get_property { group_id, listing_id } for the full record. The listing id is an OSK like EYxOzZSAbCdEf12345; you can pass a portal URL instead and the MCP will extract it.
  4. Photos / schools / walkability are separate calls. Use the lat/lng from onehome_get_property to drive onehome_get_schools and onehome_get_walk_score.
  5. Compare onehome_compare_properties { group_id, targets: [...], view? } — 2 to 8 listings, concurrent fetch, per-row error capture. Don't fan out manual get_property calls when comparing.
  6. Escape hatch. onehome_graphql { operation_name, query, variables?, view? } lets you send a raw read-only document (mutations and subscriptions are refused) when you need a field the structured tools don't expose. Common operation names: GetOneHomeUser, GetListings, GetPins, ListingById, MediaListingById, GetSavedSearches, ListingSuggestionsSearch. (LocalLogic schools / walk-score are REST endpoints, not GraphQL operations — use onehome_get_schools / onehome_get_walk_score instead.)

Free-text search

onehome_search_suggestions { query } bypasses the group / saved-search structure. Use it when the user knows what they're looking for (MLS number, address, partial street name) but doesn't care which group it lives in.

Local computation

onehome_calculate_mortgage and onehome_calculate_affordability are pure local math — no network, no token needed. Use them when the user asks "what would my payment be on this place" or "what can I afford"; identical math to the other realty MCPs.

Response shape (view)

Four tools take view: "compact" | "full" — onehome_get_by_address, onehome_compare_properties, onehome_get_user and onehome_graphql — and compact is the default. You get the slim shape without asking for it; an efficiency a caller has to request is one that is usually not requested.

What compact does here is strip image and avatar URLs, and nothing else. There is no field projection, and none is claimed: this server hands OneHome's payload back close to verbatim and holds no verified record of which of its fields matter, so nothing here could honestly name a keep-list. Stripping media is subtractive — it names what to remove, never what to keep — so it cannot lose a field nobody anticipated. Every non-media field, at every depth, survives compact untouched.

The surprise is onehome_graphql. Compact keeps every envelope key (data, errors, status, url), but the strip reaches inside data — and MediaListingById, one of the operation names this skill recommends, is all media. OneHome nests its variants under an Image key, so at the default rung each photo comes back as { Order, LongDescription } with Thumbnail, Medium and Large all gone: the call succeeds and returns nothing you asked for. Pass view: "full" for any media query — and for any call you are making because a payload was not what you expected, so a missing field is never this server's doing.

view: "full" returns what the tool built with nothing removed. There is deliberately no raw rung, for two different reasons: on onehome_graphql full already is the unprojected upstream envelope, so a third value would silently alias the second; and the other three assemble their record from a query result plus derived fields, so there is no single upstream payload left to hand back.

The other seventeen tools take no view at all. That is scope — a partial rollout, not an oversight — but for two of them it is also correct on the merits:

  • onehome_get_property_photos exists to return exactly the URLs compact strips. Its product is the image; stripping there would not shrink the response, it would empty it. Never media-strip a tool whose name is the test.
  • onehome_calculate_mortgage and onehome_calculate_affordability are local math returning a handful of numbers — already narrower than any projection.

Common pitfalls

  • Forgetting group_id on listing detail / photos. Every listing surface is scoped to a group. If you have a listing id from onehome_search_suggestions, you still need a group id to inflate it. Pick the group most likely to contain it (usually the user's primary one).
  • Stale token. OneHome JWTs expire. onehome_healthcheck shows seconds_until_expiry. Refresh the env var if it's close to zero.
  • fetchproxy capture mode + no interaction. In capture mode, the MCP can't read the bearer until the user's tab fires a GraphQL request. If a call hangs there, ask the user to scroll the map / click a listing / refresh the portal page.