MCP server, private beta

    Connect your AI client to Picari.

    Picari runs a remote MCP server at https://mcp.picari.ai/v1. Point Claude, Cursor or VS Code at it and your assistant can run the same deterministic shortlist engine the app runs, read your own briefs and bake-offs, and search the security vendor catalogue on structured facts. It reads your data, not everyone's: every call is authorised as you and filtered by row-level security in Postgres.

    Request beta accessYour connections

    What you get

    What the connection gives you

    Two personas, two data sets. You pick one when you authorise the client, and the connection can never see the other.

    For buyers

    • Run the deterministic ranking engine against a brief and get the same shortlist the app produces, with the reason each product matched.
    • Read your own briefs, shortlists and bake-offs, including weighted criteria and vendor answers.
    • Search the catalogue on structured facts: category, deployment, compliance, integrations, pricing model.
    • Draft a brief, create a bake-off and set criteria, then finish the step that needs a human in Picari.

    For vendors

    • Read your company and product records exactly as a buyer sees them, so you can tell what a shortlist is scoring.
    • Search and compare the product catalogue on the same structured facts a buyer sees: category, deployment, compliance, integrations, pricing model. Those facts are already on the public directory; the connection just puts them in your assistant.
    • Read your claim status, your bake-off invitations and the requests waiting on you.
    • Draft answers to a buyer's criteria; you submit them yourself in the Vendor Portal.
    • Nothing about another vendor's evaluation is ever readable. Postgres row-level security decides, not the tool.

    Ranking is deterministic. The engine that produces a shortlist over MCP is the same scored, weighted engine the app uses, with the weights read from Postgres. No model reorders the result, so the same brief gives the same ranking on every run.

    How to connect

    One endpoint, five clients

    Authorisation is OAuth 2.1 with PKCE. Your client registers itself, you approve the connection in Picari, and no API key is ever copied anywhere.

    Claude Code
    Supported
    claude mcp add --transport http picari https://mcp.picari.ai/v1

    Then run /mcp in Claude Code and approve. The callback is a loopback URL on an ephemeral port; pass --callback-port to pin it behind a strict firewall.

    Claude Desktop and claude.ai
    Supported
    Settings, Connectors, Add custom connector, URL https://mcp.picari.ai/v1

    The callback is https://claude.ai/api/mcp/auth_callback. Nothing to configure on your side.

    Cursor
    Supported
    Settings, MCP, add an HTTP server at https://mcp.picari.ai/v1

    Cursor uses its own web callback and a fixed loopback port, both already registered.

    VS Code
    Supported
    mcp.json: servers.picari = { "type": "http", "url": "https://mcp.picari.ai/v1" }

    The callback is a fixed loopback port or the vscode.dev redirect, depending on where you are running it.

    ChatGPT
    Supported
    Connectors, developer mode

    Add Picari as a connector in developer mode. ChatGPT cannot send custom headers, so the connection is the OAuth flow only.

    The rules

    Access expires, and the limits are published

    1,000 requests per hour

    Per connection, sliding window. Over it, the client is told to retry and when.

    25 rows per page

    Every list is paged. A cursor comes back with the page, so a client can walk the rest.

    2,000 rows per day

    Per connection, resetting at 00:00 UTC. The usage tool reports what is left.

    You re-authorise every 30 days. A connection is granted for 30 days and then stops, with no silent renewal. We email you 7 days and 1 day before it lapses, and you can revoke it at any moment from your Profile or the Vendor Portal.

    No email is ever sent from a connection. An assistant can draft a bake-off invitation or an answer, but sending it is a step you take in Picari. Nothing leaves the platform because a model decided to send it. The only mail the connection causes is the expiry warning above, and the app sends that, not the client.

    Structured facts only. Vendor free text, AI summaries and internal notes never cross the boundary. What a client reads is the same column set the public directory renders.

    Not on a Picari staff account. An account holding the Picari administrator role cannot authorise a client, and any token it holds is refused on every call. That role reads every customer's evaluation data, so the blast radius of one leaked token would be the platform rather than one account.

    Private beta

    Ask for access

    The server is live and the allowlist is short while we watch the first connections. Tell us which client you use and which account to enrol.

    Request beta access