Quality Screener

Quality Screener MCP server

A standalone Model Context Protocol (MCP) server that exposes the Quality Screener stock-screening engine as tools for AI agents (Claude, Cursor, and any other MCP client).

Once connected, an agent can screen and filter the scored universe, compute custom quality scores, inspect score history, manage saved scoring systems, and generate shareable screen links — acting as the signed-in user, against the same data they see in the web dashboard.

• No SDK dependency. The server is a thin HTTP layer over the public Quality Screener API. It has no dependency on the backend Python package — every tool just calls a REST endpoint and returns the JSON payload.
• Multi-tenant & credential-free. When deployed over HTTP the server stores no credentials of its own. Each request carries the caller's own access token, which the server forwards to the API, so a single public deployment can serve many users without ever acting on a shared account.

Table of contents

• How it works
• Quick start (remote)
• Running locally
• Configuration
• Authentication
• Tools
• Working with CustomScoreConfig
• Connecting an MCP client
• Deployment
• Deploy on Vercel
• Development
• Privacy Policy
• Support
• License

How it works

┌─────────────┐   MCP (stdio | streamable-HTTP)   ┌──────────────────┐   HTTPS   ┌──────────────────────┐│  AI agent   │ ────────────────────────────────► │  qscreener-mcp   │ ────────► │ Quality Screener API ││ (MCP client)│ ◄──────────────────────────────── │   (this server)  │ ◄──────── │   (FastAPI backend)  │└─────────────┘         tool calls / JSON          └──────────────────┘  REST     └──────────────────────┘

Each MCP tool maps to one Quality Screener REST endpoint. The server attaches the caller's bearer token to every outbound request (header X-Stobot-CLI-Token, Authorization: Bearer … also accepted) and returns the decoded JSON. There is no business logic in the server itself — it is a typed, authenticated façade over the API.

It runs in two transport modes:

Over HTTP the MCP endpoint is served at /mcp.

Quick start (remote)

The easiest way to use the server is to point your MCP client at the hosted deployment. No token to copy — the client triggers a browser sign-in on first connect:

{  "mcpServers": {    "qscreener": {      "type": "streamable-http",      "url": "https://mcp.qualityscreener.io/mcp"    }  }}

On first use your browser opens the Quality Screener sign-in page. Approve once, and the agent stays connected. You need a Quality Screener account; the agent inherits exactly your access.

Running locally

Requires uv.

# Install dependenciesuv sync
# stdio — for a local agent that launches this as a subprocessuv run qscreener-mcp
# streamable-HTTP — mirrors the remote deploymentQSCREENER_MCP_TRANSPORT=streamable-http QSCREENER_MCP_PORT=8080 \  QSCREENER_API_URL=http://localhost:8001 \  uv run qscreener-mcp# -> MCP endpoint at http://localhost:8080/mcp

With Docker:

docker build -t qscreener-mcp .docker run --rm -p 8080:8080 \  -e QSCREENER_API_URL=https://your-backend.example.com \  -e QSCREENER_MCP_PUBLIC_URL=http://localhost:8080 \  qscreener-mcp# -> MCP endpoint at http://localhost:8080/mcp

By default the container runs the streamable-http transport on port 8080.

Configuration

All configuration is via environment variables, resolved at startup.

Authentication

The server resolves a bearer token for each call with the following precedence:

1. HTTP request header — X-Stobot-CLI-Token, then Authorization: Bearer <token>.
2. $QSCREENER_TOKEN environment variable.
3. $QSCREENER_CONFIG_DIR/credentials.json — the token field.

Remote (OAuth 2.0)

For a streamable-http deployment, authentication is fully automated via the MCP OAuth flow:

1. The MCP client discovers the authorization server and opens the user's browser.
2. The browser lands on the Quality Screener web app, which exchanges the user's web session for a short-lived CLI token and redirects back to this server's /oauth/callback.
3. The server hands the token to the MCP client, which sends it as a bearer token on every subsequent request.

The token is validated on each request by calling the backend's /v1/cli/auth/whoami endpoint, so a revoked or expired token is rejected immediately. The server never persists user tokens. Access tokens are the CLI token, so they need no local store. Pending login state, authorization codes, and dynamically registered clients are HMAC-signed JWTs when QSCREENER_OAUTH_SIGNING_KEY is set, which is what a multi-instance host (Vercel) needs; a single process without the key keeps the original in-memory behaviour.

Local (stdio)

Mint a token through the browser login flow and store it locally, then run the server over stdio:

qscreener auth login                          # opens the web app, stores a tokencat ~/.config/qscreener/credentials.json      # the "token" field is your bearer token

Or set QSCREENER_TOKEN directly for CI / scripted use.

Tools

All tools require authentication. Filters use OR logic within a filter and AND logic across filters. Market caps are always in USD.

Account & status

Scores & screening

Sharing

Filters & tickers

Score history

Dates are YYYY-MM-DD. Pass scoring_system_id to compute history against a saved scoring system instead of the default quality score.

Saved scoring systems

A scoring system is a named, reusable CustomScoreConfig stored against your account.

Working with CustomScoreConfig

score_compute, screen_share, and the systems_* tools accept a CustomScoreConfig object describing how to weight financial metrics. Its shape mirrors the score builder in the web dashboard: weighted metric groups, each containing weighted metrics, plus scoring parameters and an optional nested filters block. A minimal example:

{  "name": "My quality screen",  "winsorizePercentile": 5,  "missingDataPercentile": 0.25,  "normalizeGroupZScores": false,  "includeDuplicatesInScoring": false,  "groups": [    {      "id": "returns",      "name": "Returns",      "weight": 0.5,      "metrics": [        { "id": "roe", "name": "ROE", "weight": 0.5 },        { "id": "roic", "name": "ROIC", "weight": 0.5 }      ]    },    {      "id": "profitability",      "name": "Profitability",      "weight": 0.5,      "metrics": [        { "id": "profit_margin", "name": "Profit Margin", "weight": 1.0 }      ]    }  ],  "scoringUniverseFilters": { "countries": ["Italy"], "min_market_cap": 1 },  "filters": { "min_score": 1.2 }}

Scoring parameters use camelCase: winsorizePercentile (1-10), missingDataPercentile (0.1-0.5), normalizeGroupZScores and includeDuplicatesInScoring (booleans). scoringUniverseFilters defines the peer group the scores are computed against; the nested filters block holds saved-screen state. Market caps are in billions USD inside both blocks (the tool arguments take USD). Loose inputs — snake_case keys, the legacy winsorize/zScore flags, or filter keys placed at the top level — are normalized to this shape automatically, but emitting it directly is preferred. Use filters_list to discover valid filter values, and build a config interactively in the dashboard if you want a starting point to copy.

Two stages: score against, then filter

Quality scores are relative — every company is winsorized and z-scored against a population — so who is in the peer group and which rows you look at are different questions, and score_compute takes them separately.

"Best European tech judged against European tech" and "best European tech judged against the world" are different lists, not the same list rescaled — narrowing the universe moves each metric's bounds, mean and σ by different amounts, so companies genuinely reorder:

// judged against European tech — the peer group is European techscore_compute(config, scoring_universe={"sectors": ["Technology"], "regions": ["Europe"]})
// judged against the world — the peer group is everyone, then Europe is shownscore_compute(config, sectors=["Technology"], regions=["Europe"])

Stage 1 accepts sectors, industries, regions, countries, currencies, exchanges, min_market_cap_usd and max_market_cap_usd. It rejects min_score, max_score, ticker and tickers with an error rather than ignoring them: the first two filter on the very scores being computed, the rest select rows.

Every response carries a scoring_universe field naming the peer group and its size. Scores computed against different peer groups are not comparable — do not mix them in one table.

Two edges worth knowing:

• min_market_cap_usd as a stage-2 argument also floors the scoring population. This is long-standing backend behaviour, kept for compatibility. Set min_market_cap_usd inside scoring_universe to control the peer group explicitly; it overrides the stage-2 floor. max_market_cap_usd filters rows only unless you set it in scoring_universe.
• A very small universe still scores. There is no minimum-population guard yet, so winsorizing at the 5th/95th percentile of a dozen companies returns confident nonsense.

What the nested filters block does

It is saved-screen state. screen_share and systems_create/systems_update persist it so a shared screen or saved scoring system restores its filter selections when reopened in the dashboard.

It does not define the peer group — scoringUniverseFilters does. Passing a saved config to score_compute applies its filters block as stage-2 filters (an explicit argument wins), matching what the dashboard does, so re-scoring a saved system keeps its view.

Connecting an MCP client

Remote (recommended)

Any streamable-http MCP client works. No token needed — OAuth handles login:

{  "mcpServers": {    "qscreener": {      "type": "streamable-http",      "url": "https://mcp.qualityscreener.io/mcp"    }  }}

If your client cannot perform the OAuth flow, send a minted token directly:

{  "mcpServers": {    "qscreener": {      "url": "https://mcp.qualityscreener.io/mcp",      "headers": { "X-Stobot-CLI-Token": "<your token>" }    }  }}

Local (stdio)

{  "mcpServers": {    "qscreener": {      "command": "uv",      "args": ["run", "--directory", "/path/to/quality-screener-mcp-server", "qscreener-mcp"],      "env": { "QSCREENER_API_URL": "https://your-backend.example.com" }    }  }}

Deployment

The server deploys as a single container. On Railway:

1. New service → Deploy from repo, pointing at this repository. The Dockerfile is self-contained, so the build context is the repo root.

2. Set environment variables:

   • QSCREENER_MCP_TRANSPORT=streamable-http
   • QSCREENER_API_URL=https://<your-backend-domain>
   • QSCREENER_WEBSITE_URL=https://<your-frontend-domain>
   • QSCREENER_MCP_PUBLIC_URL=https://<generated-mcp-domain>

   Railway injects PORT automatically; the server binds to it.

3. Networking → Generate Domain. The MCP endpoint is https://<generated-domain>/mcp.

   • Leave the HTTP healthcheck path unset (or use a TCP check): /mcp answers 406 Not Acceptable to a plain GET, so an HTTP healthcheck expecting 200 would mark the deploy unhealthy.

4. Connect your MCP client — the OAuth flow triggers automatically on first connection.

Deploy on Vercel

The same ASGI app that Docker/Railway serve via mcp.run() is exported from app.py as a FastAPI app so Vercel’s Python runtime can host it as one Function. The FastMCP app already listens on /mcp, /.well-known/*, /authorize, /token, /register, and /oauth/callback; Vercel routes every request to that Function with the original path, so those URLs resolve without remapping. vercel.json only sets maxDuration (60s) on app.py.

Do not rewrite /mcp (or the OAuth paths) onto / — FastMCP matches the literal paths.

Create a Vercel project pointed at this repository (Framework Preset: FastAPI / Other, Root Directory: repository root). Set these environment variables — names only; never commit values:

Vercel injects VERCEL_URL automatically (hostname, no scheme). It is only a fallback issuer when QSCREENER_MCP_PUBLIC_URL is unset. Do not use it as the production issuer if the MCP client will connect via a custom domain.

These are not required on Vercel (the app.py entrypoint is always HTTP, and the Function does not bind a port):

• QSCREENER_MCP_TRANSPORT
• QSCREENER_MCP_HOST
• QSCREENER_MCP_PORT / PORT
• QSCREENER_TOKEN
• QSCREENER_CONFIG_DIR

After deploy the MCP endpoint is https://<vercel-domain>/mcp. Point an MCP client at that URL the same way as the Railway deployment. Raise functions["app.py"].maxDuration in vercel.json if custom scoring hits the 60s cap.

Development

uv sync            # install dependencies (including dev)uv run pytest      # run the test suite

The codebase is small and self-contained:

Privacy Policy

The full privacy policy is published at PRIVACY.md (https://github.com/quality-screener/quality-screener-mcp-server/blob/main/PRIVACY.md).

In short:

• The only personal data retained is your email address, which identifies your account. Username and organization are optional profile fields.
• The MCP server is a stateless proxy — it holds no database and writes no personal data to storage of its own. Each request is forwarded to the Quality Screener API using your access token, never a shared account.
• Your data is not sold, not used for advertising, and not used to train machine-learning models.
• Operational logs reference accounts by a pseudonymous user ID, not by email.
• Account data is deleted when you delete your account.

See the policy for retention periods, third-party recipients, international transfers, and your GDPR rights.

Support

This README is the canonical documentation for the MCP server: https://github.com/quality-screener/quality-screener-mcp-server

Please report suspected security vulnerabilities privately by email rather than opening a public issue.

License

MIT © Quality Screener.