Verdict

Documentation

Verdict v0.1. Sibling product to Preflight and Design Check.

Overview

Verdict is a review layer for design work. Each tool takes something you can paste (a description, a screenshot URL, some strings, a brief) and returns a structured review made by a language model working from a fixed method. It does not connect to Figma, Storybook, analytics or research tools; it reviews what you give it.

Findings keep three things apart: what is visible in your material, what the reviewer inferred, and what it could not assess. Treat the output as a second opinion. You decide what to do with it.

REST API

One route per tool. The body is the tool's arguments as JSON.

curl -X POST https://verdict.allthepossibles.com/api/find_assumptions \
  -H "content-type: application/json" \
  -H "Authorization: Bearer vd_..." \
  -d '{"prd_text": "Users want a weekly digest. We will add it to settings."}'

GET /api/tools returns every tool with its description and input schema. Errors come back as {"error": "..."} with a status of 400 (bad input), 401 (inactive key), 429 (rate limit or allowance), 502 (the review model failed) or 503 (free tier closed for the day).

MCP server

{
  "mcpServers": {
    "verdict": { "url": "https://verdict.allthepossibles.com/mcp" }
  }
}

Streamable HTTP, POST /mcp. Send the key as an Authorization: Bearer header to use a paid tier. Input and allowance problems come back as tool results with isError set, so an agent can read them and adjust.

Install, app by app

Works with Claude, Cursor, VS Code, Codex, ChatGPT, Gemini and Copilot. Every app needs the same address, https://verdict.allthepossibles.com/mcp; what differs is the file it lives in and the name of the field. Gemini CLI, Copilot CLI and VS Code were tried on 2026-10-29 with this server; the rest follow each vendor's current reference. Each review is a real model call, so the free tier allows two a day; a check with an empty input returns an error without using one.

Free needs no key. Pro: get a key and send it as Authorization: Bearer vd_YOUR_KEY; the server also reads x-api-key and, for apps that cannot send headers, ?key= on the URL (a key in a URL can land in client logs, so prefer the header).

Claude Desktop and claude.ai

  1. Settings, then Connectors, then Add custom connector.
  2. Name: Verdict. URL: https://verdict.allthepossibles.com/mcp. Save.
  3. In a new chat, open the tools menu and switch Verdict on.

Custom connectors need a paid Claude plan. These apps cannot send headers, so a Pro key goes on the URL: https://verdict.allthepossibles.com/mcp?key=vd_YOUR_KEY.

Check it: say Call the find_assumptions tool from the Verdict server. Give it one paragraph of a PRD and it lists the assumptions by risk.

Claude Code

claude mcp add --transport http verdict https://verdict.allthepossibles.com/mcp

With a key:

claude mcp add --transport http verdict https://verdict.allthepossibles.com/mcp --header "Authorization: Bearer vd_YOUR_KEY"

The entry lands in ~/.claude.json (user scope) or .mcp.json in the project with --scope project. Run /mcp inside Claude Code to see the server and its tools.

Cursor

Settings, then MCP, then Add new MCP server; this opens ~/.cursor/mcp.json (or .cursor/mcp.json in a project). Add:

{
  "mcpServers": {
    "verdict": {
      "url": "https://verdict.allthepossibles.com/mcp",
      "headers": {
        "Authorization": "Bearer vd_YOUR_KEY"
      }
    }
  }
}

Delete the headers block on the Free tier. Restart Cursor fully.

Codex

Add to ~/.codex/config.toml and restart Codex:

[mcp_servers.verdict]
url = "https://verdict.allthepossibles.com/mcp"
# Pro: bearer_token_env_var = "VERDICT_KEY"   (set it in your shell)

ChatGPT

Settings, then Connectors, then Advanced, turn on Developer mode, then Create a connector. Name it Verdict and paste https://verdict.allthepossibles.com/mcp. Developer mode exists on some paid plans only. A Pro key goes on the URL as ?key= because connectors cannot send headers.

Gemini CLI

Config file: ~/.gemini/settings.json for every project, or .gemini/settings.json inside one project. The quickest route is the command, which writes the file for you:

gemini mcp add --transport http verdict https://verdict.allthepossibles.com/mcp

With a key:

gemini mcp add --transport http --header "Authorization: Bearer vd_YOUR_KEY" verdict https://verdict.allthepossibles.com/mcp

If you edit the file by hand, the field is httpUrl:

{
  "mcpServers": {
    "verdict": {
      "httpUrl": "https://verdict.allthepossibles.com/mcp",
      "headers": {
        "Authorization": "Bearer vd_YOUR_KEY"
      },
      "timeout": 600000
    }
  }
}

Why the field name matters: a bare url with no type is read as the old SSE transport and the connection fails. httpUrl means Streamable HTTP. The gemini mcp add --transport http command writes url with "type": "http", which also works. timeout is in milliseconds (default ten minutes). Add "trust": true to skip the confirmation before each tool call.

Check it: gemini mcp list prints verdict: https://verdict.allthepossibles.com/mcp (http) - Connected. Then in a session: Call the find_assumptions tool from the verdict server. Gemini shows a prompt before the first call unless the server is trusted.

Verified 2026-10-29 with Gemini CLI 0.66: all four studio servers connect and list their tools through gemini mcp list; the tool call itself needs a Google sign-in or GEMINI_API_KEY. Google moved unpaid accounts to Antigravity CLI in June 2026; the Antigravity section below covers that config.

Antigravity

Config file: ~/.gemini/config/mcp_config.json for every workspace, or <workspace>/.agents/mcp_config.json. In the agent panel: MCP Servers, Manage, View raw config. The field is serverUrl:

{
  "mcpServers": {
    "verdict": {
      "serverUrl": "https://verdict.allthepossibles.com/mcp",
      "headers": {
        "Authorization": "Bearer vd_YOUR_KEY"
      },
      "timeoutSeconds": 600
    }
  }
}

Why the field name matters: Antigravity does not read httpUrl or the old SSE form; serverUrl (or url) is Streamable HTTP. Tool calls time out at 180 seconds by default; timeoutSeconds raises that. Our longest call (the 15-page accessibility crawl on Designers) ran in under two seconds on 2026-10-29, so the default is fine unless the site you crawl is slow. Tools run in Ask mode until you allow them. This section is the config shape from Google's reference; it was not run on a machine here.

Gemini Code Assist in VS Code

Config file: the same ~/.gemini/settings.json as Gemini CLI, with httpUrl:

{
  "mcpServers": {
    "verdict": {
      "httpUrl": "https://verdict.allthepossibles.com/mcp",
      "headers": {
        "Authorization": "Bearer vd_YOUR_KEY"
      }
    }
  }
}

Then run Developer: Reload Window from the Command Palette and open agent mode in the Gemini Code Assist chat. JetBrains keeps a separate mcp.json in the IDE config directory; Google's example there is local only, so headers on a remote entry are unverified. This section is the config shape from Google's reference; it was not run on a machine here.

GitHub Copilot in VS Code

Config file: .vscode/mcp.json in the workspace, or the user file from the Command Palette command MCP: Open User Configuration (on macOS it is ~/Library/Application Support/Code/User/mcp.json). VS Code's own format uses servers, not mcpServers, and the key goes through an inputs entry so it never sits in the file:

{
  "inputs": [
    {
      "type": "promptString",
      "id": "verdict-key",
      "description": "Verdict key (leave empty for Free)",
      "password": true
    }
  ],
  "servers": {
    "verdict": {
      "type": "http",
      "url": "https://verdict.allthepossibles.com/mcp",
      "headers": {
        "Authorization": "Bearer ${input:verdict-key}"
      }
    }
  }
}

Why the field name matters: "type": "http" is Streamable HTTP; without it VS Code may try the old SSE transport. VS Code asks for the input value the first time the server starts and keeps it in its secret store. From a terminal the same entry can be added to the user file in one line:

code --add-mcp '{"name":"verdict","type":"http","url":"https://verdict.allthepossibles.com/mcp"}'

Open Copilot Chat in agent mode, confirm the trust prompt, and start the server from the inline Start link if it is stopped. Verified 2026-10-29 with VS Code 1.108 and Copilot Chat 0.36: both files parse and the server appears in the MCP list; the tool call itself needs a GitHub sign-in in Copilot Chat.

Check it: say Call the find_assumptions tool from the Verdict server. Give it one paragraph of a PRD and it lists the assumptions by risk.

GitHub Copilot CLI

copilot mcp add --transport http verdict https://verdict.allthepossibles.com/mcp

With a key:

copilot mcp add --transport http --header "Authorization: Bearer vd_YOUR_KEY" verdict https://verdict.allthepossibles.com/mcp

Config file: ~/.copilot/mcp-config.json (user), or .mcp.json or .github/mcp.json in a project. The shape it writes:

{
  "mcpServers": {
    "verdict": {
      "type": "http",
      "url": "https://verdict.allthepossibles.com/mcp",
      "headers": {
        "Authorization": "Bearer vd_YOUR_KEY"
      },
      "tools": [
        "*"
      ]
    }
  }
}

Why the field name matters: "type": "http" is Streamable HTTP; sse is the old transport and will not connect. The key sits in plain text in that file, so keep the file private. Query-string keys are not supported here; use the header.

Check it: copilot mcp list shows the server, then in a session: Use the find_assumptions tool on the verdict server. Verified end to end on 2026-10-29 with Copilot CLI: add, list, a free tool call and a keyed tool call all worked on all four studio servers.

GitHub Copilot coding agent (github.com)

Repository Settings, then Copilot under Code, planning, and automation, then MCP servers. Paste this and click Save MCP configuration:

{
  "mcpServers": {
    "verdict": {
      "type": "http",
      "url": "https://verdict.allthepossibles.com/mcp",
      "headers": {
        "Authorization": "Bearer $COPILOT_MCP_VERDICT_KEY"
      },
      "tools": [
        "*"
      ]
    }
  }
}

The key is a repository secret in the copilot environment, named with the COPILOT_MCP_ prefix (here COPILOT_MCP_VERDICT_KEY); delete the headers block on the Free tier. Only tools are available to the agent (no prompts or resources), they run without a confirmation step, and query-string keys are not supported. Every tool on this server is read-only, so ["*"] is safe. This section is the config shape from GitHub's reference; it was not run on a repository here.

Microsoft Copilot Studio and Gemini Enterprise (admin paths)

Both are set up by a tenant administrator, not by an end user. Copilot Studio: in an agent, Tools, Add a tool, Model Context Protocol, then give the name, a description, the URL https://verdict.allthepossibles.com/mcp and authentication None (Free) or API key in a header named Authorization with the value Bearer vd_YOUR_KEY (Pro). Streamable HTTP only; the agent needs generative orchestration on; your tenant's data policies apply. Gemini Enterprise: an admin registers the server as a custom MCP data source (preview since April 2026, Streamable HTTP only, off by default) and assigns it to agents. Neither was run from here; the owner can request either on your behalf at support@allthepossibles.com.

The six tools

ToolInputOutput
critique_experienceartifact ({type: description or screenshot_url, value}), optional prd, contextfindings (category ux, visual or product; severity critical, warning or note; finding, evidence, suggested_resolution) and a summary
generate_state_matrixfeature_description, optional known_dimensionsdimensions (name, states) and up to five likely_gaps, each marked stated or inferred
map_journeyobjective, optional user_maturity (first_use, occasional, habitual, expert)stages, entry_points_missing, exit_paths_missing, interruption_handling
audit_copystrings: up to 60 of {text, context: button, error, empty_state, confirmation or notification}findings (text, issue, fix); strings that work get no entry
find_assumptionsprd_textassumptions (assumption, risk low, medium or high, research_question), highest risk first
check_terminologysources: two to twelve of {label, text}concept_clusters (concept, terms_used by source, drift), drifting clusters first

Text inputs are limited to 30,000 characters per field. Screenshot URLs must be public and use https; the service passes the URL to the model and never fetches it itself.

Limits and allowance

Every call runs a real model request, so every tier has a ceiling on total use as well as on speed.

TierRate limitAllowance
Free, no key20 calls a minute2 calls a day per caller, plus a small shared daily limit across all free callers that resets at 00:00 UTC. A demo allowance, not a plan.
Your own Anthropic key20 calls a minuteUnlimited. Send the key in an x-anthropic-key header and the call runs on your account. We never store or log it.
Basic, $5 a month150 a minuteA monthly allowance that resets on the 1st (UTC)
Pro, $19 a month500 a minuteFive times Basic's monthly allowance

The allowance is measured in real model cost, not call count, so a long brief uses more of it than a short string. /usage shows how much of yours is used.

Known limitations

A review is only as good as the material. A description cannot show motion, hover states or real data, and a screenshot cannot show behavior over time; the reviewer is told to say so and not to pretend otherwise. Contrast and size figures read from an image are estimates.

The reviewer is a language model. It can be wrong, and it can miss a problem. Use it to widen what you look at, then verify what matters. find_assumptions never simulates user answers, because imagined participants are not evidence.

Output is validated into the documented shape, but wording varies between calls on the same input.

What happens to your input

Your input is sent to Anthropic's API to produce the review and is not written to Verdict's database. Usage is logged as event counts (which tool, which tier, time, rough cost) with no input text and no tracking cookies. Do not send material you are not allowed to share with a third-party model provider.

Support

Billing, invoices, canceling: /billing (a form that takes your key). Anything else: support@allthepossibles.com.