Skip to content
Phase logo

Phase MCP server

Which MCP servers are Phase's own, from the official registry and from asking its own domain, whether they answer, and what they offer. A program checked each of these, nobody copied them from a product page.

read 2026-10-28

← Back to Phase

Whose server it is
  • Registry namexyz.phase.ai/gatewayverified in the MCP registry

The registry name sits under phase.xyz. The registry only lets a publisher use that namespace after proving control of the domain through DNS or HTTP.

When we asked itchecked 2026-10-28
  • Stateanswerschecked 2026-10-28
  • Tools7
  • Server calls itselfphase-ai
  • Server version0.3.0
  • Protocol version2025-06-18

On 2026-10-28 the server answered initialize and tools/list without a sign-in. First answer in 492 ms.

Should an agent connect itNot proven either way yet
  • Whether its tools do the product's job has not been judged yet.
  • Checked on 6 of the 7 days it takes to judge reliability.

Said by rules from the measurements above. The likely first call is a model's pick from the tool list; read its description before relying on it.

Measured dailydaily checks: 6, the first on 2026-10-28
  • INITE MCP scoreafter a weeka score needs 7 daily checks: 6 so far
  • Answeredmeasuring6 of the 7 daily checks it takes to print a share
  • Median first answer466 ms
  • Space in an agent's context≈ 1,130 tokensthe definitions of all its tools (7), which an agent carries on every turn
  • Tool definitionsA · 98 of 100
  • Typed results7 of 7 toolsdeclared outputSchema: results come typed, read once
  • Parameters without a description: in 1 of 7

Worth a look

None of the five warning signs in these definitions.

A program checks the server once a day: initialize and tools/list, no model. Tokens are the definitions' length at about four characters a token. The grade averages six checks of every tool's name, description, schema and annotations; it says how well the tools are described, not whether they work. The score weighs what was measured and nothing else: answering (25%), doing the product's job (25%), definitions (20%), context (10%), use (10%) and how readily a client connects (10%). A part that was not measured is left out, and a hidden instruction caps the score at 40. The state of vendors' MCP servers: method and numbers →

Tools it offers7 tools · read 2026-10-28
  • chat

    Run a chat completion on the Phase AI gateway, paid from the Phase API key or the wallet this server uses (free models cost nothing). Returns the answer, actual price, task id for the receipt and usage.

  • embed

    Compute embeddings on the Phase AI gateway (paid from the Phase API key or the wallet).

  • key

    Balance, daily cap and spend of the Phase API key this server pays with (key mode).

  • list_models

    List models of the Phase AI gateway with prices (USD per 1M tokens, fee per call) and layer (flagship | open | class | free).

  • quote

    Get the price ceiling for a chat request without paying (the gateway's 402 quote).

  • receipt

    Fetch the public receipt of a paid request by task id (model, provider, tokens, price, settlement tx, request/response hashes, signatures) and its Merkle proof once the weekly epoch is closed.

  • session

    Saved chat sessions on the Phase AI gateway: create one (returns an id and a shareable /s/<id> link), read its thread, list the wallet's sessions (wallet mode only: signed with the wallet key, no transaction), or delete one. Pass the id…

How to connect

Add the server to your MCP client config. Claude Desktop, Cursor and most other clients read this shape:

{
  "mcpServers": {
    "gateway": {
      "type": "http",
      "url": "https://mcp.fermat.ai.phase.xyz/mcp"
    }
  }
}

Add it

Claude Code, in a terminal

claude mcp add --transport http phase https://mcp.fermat.ai.phase.xyz/mcp

Add to CursorAdd to VS Code

claude.ai, ChatGPT and other clients: add https://mcp.fermat.ai.phase.xyz/mcp as a custom connector in the settings.

Each line is something a program found when it went looking: the registry entry, the server's own answer, a spec that parses. A missing line means we did not find it, not that it does not exist.