eBay Browse

ebay-mcp

Give an AI agent real, live market prices — straight from the largest secondhand marketplace on the internet.

eBay is a continuously-updating ledger of what physical things actually cost right now. This wraps its Browse API as an MCP server with three tools, so an agent can search listings, pull a single item, and — the useful one — get an aggregated price landscape for anything: min, median, max, broken down by condition.

Python 3.12+ · MIT · MCP server · app-token auth, no user login

Setup is one free eBay app keyset — no user login, no OAuth consent screen to click through. Point an agent at it and ask "what does an RTX 5080 actually go for?" — one call back comes a grounded answer, split by condition, with the cheapest listings attached.

Contents

• The three tools
• Architecture: one client, three tools
• Authentication: client-credentials, cached
• ebay_price_check: how the landscape is built
• Install
• Configuration
• Sandbox vs. production
• Project layout
• License

The three tools

// ebay_price_check  ·  query: "RTX 5080", exclude: ["laptop", "notebook"]
{
  "count": 47, "currency": "USD",
  "min": 899, "median": 1099, "max": 2200,
  "by_condition": {
    "New":  { "count": 18, "min": 1099, "median": 1199, "max": 1634 },
    "Used": { "count": 21, "min": 899,  "median": 1050, "max": 1499 }
  },
  "cheapest": [ { "price": 899, "condition": "Used", "title": "…", "itemWebUrl": "…" } ]
}

Architecture: one client, three tools

Every tool flows through a single EbayBrowseClient, which owns the token and talks to eBay. There's one place credentials are read, one place a token is cached, one place HTTP happens — nothing to drift.

flowchart LR
    Agent(["AI agent / Claude"])
    subgraph server["ebay-mcp · stdio server"]
        Tools["ebay_search<br/>ebay_get_item<br/>ebay_price_check"]
        Client["EbayBrowseClient"]
        Cache[("OAuth token<br/>in-memory, auto-refresh")]
    end
    Cfg["~/.ebay-mcp.toml<br/>or env vars"]
    eBay["eBay Browse API"]

    Agent -->|"MCP tool call"| Tools
    Tools -->|"search / get_item"| Client
    Client <-->|"reuse or mint token"| Cache
    Client -->|"Bearer token + query"| eBay
    eBay -->|"listings JSON"| Client
    Cfg -.->|"keyset + active env"| Client

The server is async; the client is plain synchronous requests, run in a thread (asyncio.to_thread) so a slow eBay call never blocks the event loop. ebay_price_check is the one tool that does more than pass through — it runs a search and then aggregates the result (see below).

Authentication: client-credentials, cached

eBay's Browse API uses an application token (the OAuth client-credentials grant) — no user is involved. The client mints one on first use, caches it in memory, and silently refreshes when it's about to expire. You never think about it.

sequenceDiagram
    participant T as Tool call
    participant C as EbayBrowseClient
    participant O as eBay OAuth
    participant B as Browse API

    T->>C: search("RTX 5080")
    alt token missing or expired
        C->>O: POST /identity/v1/oauth2/token<br/>Basic(app_id:cert_id), grant=client_credentials
        O-->>C: access_token + expires_in
        Note over C: cache until (expires_in − 60s)
    end
    C->>B: GET /item_summary/search<br/>Authorization: Bearer …
    B-->>C: listings JSON
    C-->>T: parsed results

The 60-second buffer means a token is treated as expired slightly early, so a call never races a token that dies mid-flight. Tokens live ~2 hours; in practice one fetch covers a long session.

ebay_price_check: how the landscape is built

The other two tools are thin wrappers. This one is the reason the project exists: it turns a pile of raw listings into a number you can reason about.

flowchart LR
    Q["query<br/>+ exclude[]"] --> S["search<br/>(up to 50 listings)"]
    S --> F["drop excluded titles<br/>+ unpriced listings"]
    F --> G["group by condition"]
    G --> A["aggregate<br/>min · median · max"]
    G --> H["cheapest N<br/>(the tail)"]
    A --> R(["{ count, min, median, max,<br/>by_condition, cheapest }"])
    H --> R

exclude is what makes the number honest — a search for "RTX 5080" is full of laptops and prebuilt PCs, and exclude: ["laptop", "notebook", "prebuilt"] strips them so the median reflects the actual card. The by_condition split matters just as much: a "median" that blends new-in-box with used-and-abused is noise; split by condition and each tier tells the truth.

One honest limitation worth knowing: the Browse API returns active asking prices, not completed sales. Treat the floor as "best currently advertised," not "what it sold for."

Install

git clone https://github.com/cunicopia-dev/ebay-mcp
cd ebay-mcp
python3.12 -m venv .venv && source .venv/bin/activate
pip install -e .

You need a (free) eBay developer application keyset — see docs/SETUP.md for the five-minute walkthrough. Then wire it into your MCP client:

{
  "mcpServers": {
    "ebay": { "command": "/path/to/ebay-mcp/.venv/bin/ebay-mcp" }
  }
}

Configuration

Credentials come from environment variables (highest priority) or a ~/.ebay-mcp.toml file. The active env selects the keyset and the API base URL together — so production creds can never accidentally point at the sandbox, or vice versa.

flowchart TD
    Start(["load_config()"]) --> Env{"EBAY_ENV /<br/>EBAY_*_APP_ID<br/>in environment?"}
    Env -->|"set"| UseEnv["take keyset<br/>from env vars"]
    Env -->|"unset"| Toml{"~/.ebay-mcp.toml<br/>present?"}
    Toml -->|"yes"| UseToml["take keyset<br/>from TOML"]
    Toml -->|"no"| Default["default env = production<br/>(error if creds missing)"]
    UseEnv --> Pick["env → keyset + base URL<br/>(locked together)"]
    UseToml --> Pick
    Default --> Pick

# ~/.ebay-mcp.toml   (chmod 600)
env = "production"

[production]
app_id  = "YourApp-PRD-..."
cert_id = "PRD-..."

[sandbox]
app_id  = "YourApp-SBX-..."
cert_id = "SBX-..."

Check what's active any time — credentials are masked in the output:

ebay-mcp-config
# env:      production
# app_id:   Keit****87dd
# cert_id:  PRD-****0914
# api_base: https://api.ebay.com
# OK — configuration is valid.

Sandbox vs. production

Flip env between sandbox and production to switch environments — same code, different endpoints and keyset. The sandbox is good for proving the auth flow wires up; its inventory is sparse and seeded, so for real prices you want a production keyset.

Project layout

src/ebay_mcp/
  config.py    # env + TOML loader; ebay-mcp-config CLI
  browse.py    # EbayBrowseClient — OAuth cache + search / get_item
  server.py    # MCP server: list_tools / call_tool / main
tests/         # config precedence, aggregation, tool listing (no network)
docs/SETUP.md  # getting an eBay keyset

License

MIT