AlphaDesk — dense, fast market-research terminal & MCP integration platform

A dense, fast market-research terminal and integration platform. Readers connect their own market-data and news providers; AlphaDesk is the workspace that connects, checks and presents what those providers — and the public record at SEC EDGAR and the US Treasury — actually say. Questions are asked in the reader's own AI agent (Claude, ChatGPT, Codex, Cursor, opencode), which reads the same records through 37 read-only agent tools (MCP).

AlphaDesk reads: quotes and charts, news, SEC filings, financial statements, ownership and insider activity, earnings and corporate calendars, options, crypto. It does not trade, route orders, hold positions, score ideas or write summaries.

Two ways to use it — the managed cloud or your own server. The code is open source under the GNU AGPL-3.0, with a commercial licence for anyone who cannot meet its terms. See Licence.

Contents

1. What AlphaDesk is — and is not

2. Two ways to run it

3. Design principles

4. Features

5. Data sources

6. Search

7. Agent access (MCP)

8. Accounts, security and privacy

9. Architecture

10. Repository layout

11. Development

12. Configuration reference

13. Production deployment and operations

14. Testing and quality

15. Known limitations

16. History

17. Licence

Related MCP server: MCP-Server-Financial-Analyzer

What AlphaDesk is — and is not

It is offered free at present (payments are designed but switched off; see Accounts).

Two ways to run it

Option A — Managed cloud

1. Open the hosted terminal and sign in with GitHub — the account is created on first sign-in. (Google sign-in admits invited accounts only until Google approves the app.)

2. On the Account page, connect the market-data and news providers you already use (Alpaca, FMP, Finnhub, Polygon, CoinGecko, Tiingo, …). SEC EDGAR and the US Treasury need no key.

3. It is free during early access — there is no plan to choose and no card to enter. Paid plans ($19 a month or $190 a year, each after a 14-day trial) will be announced to account holders before they start; nobody is charged without subscribing.

4. Optional: connect your own agent from the Account page — a token for Claude Code, Codex, Cursor or opencode, or add the server URL as a connector in Claude.ai or ChatGPT.

Option B — Self-hosted (AGPL-3.0)

You need a SEC User-Agent with real contact details (SEC blocks requests without one), Python 3.11+ or Docker, and — for anything beyond EDGAR and Treasury data — your own vendor keys.

1. Configure. Copy the template and fill in the two required values:

cp alphadesk/deploy/env.example .env

Generate the vault key with:

python -c "import os, base64; print(base64.b64encode(os.urandom(32)).decode())"

2a. Run with Python.

pip install -r requirements.txt
python -m alphadesk.main keys import-env
python -m alphadesk.main dashboard

The terminal is at http://127.0.0.1:8000. The first start downloads the embedding model into the Hugging Face cache.

2b. Or run with Docker. The image bakes the embedding model in, so a container downloads nothing at start:

docker build -t alphadesk .

docker run -d --name alphadesk -p 8000:8000 --env-file .env -e ALPHADESK_DATA=/data -v alphadesk-data:/data alphadesk

With ALPHADESK_AUTH left on and no sign-on provider configured, create a password account inside the container:

docker exec -it alphadesk python -m alphadesk.main user add you@example.com

3. Keep it private or publish your changes. Under the AGPL, if you let other people use a modified copy over a network, you must offer them the source of your version. Using it yourself, unmodified, or keeping your changes to yourself on your own machine carries no such obligation. Billing stays off (ALPHADESK_BILLING_ENFORCE is off by default), so every account on your instance has full access.

Design principles

These are the product. Much of what might look missing was built, measured and removed on purpose — DECISIONS.md says what and why. Bring evidence and open an issue.

1. Nothing is paraphrased. Every panel and agent tool returns what a vendor or EDGAR actually said — a filing in pages, a statement series, a story's own text. If a figure cannot be shown with its source, it is not shown.

2. Indicators hide when the data cannot support them. Bar coverage and gap size are measured per series; below the floor RSI and MACD are hidden rather than drawn over sparse prints that would render identically to a liquid name's chart.

3. Nothing is ranked for you. The screener window is alphabetical; calendars and agent tools order only by the figure on each row. Search results are newest first — a model may decide what is related, never what comes first.

4. An idle terminal spends nothing. Background loops fetch only public data and each reader's own news; everything else is fetched on request, on the requesting reader's keys.

5. Untrusted text stays untrusted. Headlines, articles, filings and transcripts are data. Nothing here acts on them, and the agent tools that return them say so.

6. Not an aggregator. Merging several news feeds is the reader's act, on their own credentials, for their own window.

7. The agent surface is read-only and per reader. Every agent call runs as exactly the reader whose token or grant it carries.

8. Vendor data is the reader's. No vendor key on the server; no keyless route to a commercial vendor or unofficial endpoint; every vendor cache keyed by reader; vendor data kept only as long as a feature reads it.

Features

Pages

Across the terminal

• The symbol strip scopes every page; picking a row adds and selects the symbol without moving the page.

• Composable boards: show, hide, reorder, size and place tiles; the layout lives in the URL, so a link restores the exact board.

• Live data where the reader's plan streams it (Alpaca stock, crypto and news sockets), polling elsewhere; a 428 prompt names the vendors and plans that would fill any panel the reader's keys cannot.

• Phones: every page is laid out for a 375-pixel screen as well as a desktop.

• Light and dark themes, a hand-rolled design system on a 4-pixel grid with six type roles, and full keyboard focus handling in every overlay.

Data sources

Market data — the reader's own keys

For each panel the reader's connected vendors are asked in a fixed, documented order; the first that carries the figure answers. Charts and option chains are pinned to one vendor, never stitched across tapes.

News — the reader's own feeds

Alpaca (Benzinga), Polygon, Finnhub, Benzinga, Tiingo, Alpha Vantage, Marketaux and FMP. Several feeds merge into one window per reader, de-duplicated by URL. Alpaca's news streams live; the rest poll every five minutes.

Public data — no key

• SEC EDGAR: filings and their text, XBRL financial statements, Form 4 insider trades, 8-K Item 2.02 and 6-K results releases (release day read from the filing itself), ticker and company lists.

• US Treasury: the daily par yield curve.

Every source, how it is collected and its terms are listed in docs/data-sources.md. Vendor terms are separate from the code licence; see Known limitations.

Search

News search works two ways at once, everywhere a reader searches — the News filter box, "Search all" and the agent's news_search tool:

• By words — one rule shared by the server and the browser: whole words in order, plurals matched to singulars, a capitalised ticker matched exactly, and a query that names a company also finds stories tagged with it ("robinhood" finds stories tagged HOOD).

• By meaning — stories whose headline is close in meaning to the query are added and marked related, so "AI data center spending" finds "Equinix plans $5B–$7B annual data center buildout" without a shared phrase. This uses Qwen3-Embedding-0.6B (Apache 2.0), self-hosted inside the server: no text leaves AlphaDesk and there is no model key. Headlines are embedded once as they arrive; the threshold (cosine 0.50) was calibrated so that only on-topic stories are added. Results remain newest first.

  The embedding work never competes with the site: the model runs on a single CPU thread, the background worker runs at the lowest operating system priority, and it embeds only while no request is being served. Searches take about a third of a second.

A coin's news panel reads all crypto and what moves it — any coin's stories, crypto stocks, stories naming crypto, and Fed and rates headlines — each marked with the reason it is there.

Agent access (MCP)

AlphaDesk exposes the same records the interface shows as 37 read-only tools over the Model Context Protocol, at /api/agent/tools/mcp. Every call runs as the reader, on their keys, rate-limited to 120 requests a minute per token.

Connecting

• Claude.ai, ChatGPT — add a custom connector with the server address and sign in when asked (OAuth 2.1 with PKCE; one live grant per registered app).

• Claude Code, Codex, Cursor, opencode — create a token on the Account page under Agent access (shown once, stored hashed, revocable at once); the page gives each client's exact setup.

Tools

The tools are written for an agent that cannot see the screen: find_symbol resolves a name to a ticker from the SEC list rather than letting an agent guess; price_chart returns a thinned series carrying each point's RSI and MACD; filing_text and transcript_text return whole documents in pages; news_search marks each story as a word or meaning match. Tools that return publisher or filer text say that it is untrusted input.

Accounts, security and privacy

• Sign-in: single sign-on is the only door and is also sign-up — Google and GitHub are live, Microsoft is a configuration slot. The email address is the account key across providers; each sign-in's method is recorded.

• Sessions: HMAC-signed cookie, 14-day lifetime, sign-out everywhere.

• Vendor keys: sealed per reader with AES-256-GCM under a master key held outside the database; never logged, never shown after entry.

• Agent credentials: tokens and OAuth codes stored as SHA-256 hashes; OAuth clients sealed; consent page signed, same-site and unframeable.

• Owners (configured by email) reach the Admin page; everyone else is a reader. Access is free: a trial is recorded but the access gate is off and no payment processor is configured.

• Retention — vendor data is kept only as long as a feature reads it: stories 7 days, full article text 72 hours, meaning vectors with their stories, company announcements 30 days past the report, the forecast log 120 days, release habits 3 days, press-release checks 24 hours. Removing a key deletes what that key fetched.

• Deletion: a reader can delete their account, confirmed by typing its email; every row keyed to it is removed in one transaction. The table list is checked against the schema by the test suite.

• No analytics, advertising or tracking.

The full account, session, key-vault and agent-credential design is in docs/hosted-mode.md.

Architecture

                       ┌──────────────────────────────────────────────┐
  Browser (React SPA) ─┤  FastAPI · one process · one port            │
  Reader's agent (MCP) ┤                                              │
                       │  /api/*  ── panels, composed per request     │
                       │  /api/agent/tools/mcp ── 37 read-only tools  │
                       │  OAuth 2.1 at the root (/authorize, /token…) │
                       │                                              │
                       │  Per-reader DataRouter ──► reader's vendors  │──► Alpaca · FMP · Polygon
                       │    (ordered per panel; 428 when none carry)  │    Finnhub · Alpha Vantage
                       │                                              │    CoinGecko
                       │  Background:                                 │
                       │    news poll (5 min) + held news sockets     │──► each reader's feeds
                       │    EDGAR results sweep (15 min, weekdays)    │──► SEC EDGAR
                       │    daily earnings forecast capture           │──► Treasury
                       │    embedding worker (search by meaning)      │
                       └──────────────┬───────────────────────────────┘
                                      │
                         Postgres (Cloud SQL) in production
                         SQLite (WAL) in development

Backend — Python 3.11+. FastAPI with synchronous handlers on a 40-worker thread pool and explicit socket deadlines for every upstream. Providers are structural Protocols (NewsProvider, PriceProvider, TranscriptProvider), built per reader from their sealed keys and discoverable through entry points. The request's reader identity travels in a context variable; background threads deliberately do not inherit it. The store speaks SQLite locally and Postgres through the pure-Python pg8000 driver in production. The embedding model runs in process on CPU via sentence-transformers and PyTorch's CPU build.

Frontend — TypeScript, React 19 and Vite, built to static files the Python process serves. Tailwind CSS v4 carries the design tokens (14-pixel root, 4-pixel spacing grid, six type roles, light and dark). TanStack Query shares one query per endpoint. There is no component library and no chart library: primitives are hand-rolled, and the chart is AlphaDesk's own SVG renderer, with candles batched into four paths so the node count is constant in bar count.

Repository layout

alphadesk/
  main.py            entry point: web server + background loops, and the CLI
  app/               FastAPI app, auth, admin, agent access (tokens, OAuth, MCP mount)
  providers/         the plugin seams, vendor implementations, catalogue, per-reader router
  ingest/            EDGAR, news polling, calendars, movers, prices and indicator math
  desk/              screener window, filings, transcripts, market-today
  ledger/            store (SQLite / Postgres), database adapter, key vault
  mcp_server.py      the agent tools
  semantic.py        search by meaning (self-hosted embedding model)
  cryptonews.py      a coin's news selection
  newsquery.py       the word-search rule
  billing.py         owners, trial, the (off) access gate and payment seam
  config.py          settings, retention, curated baskets
  ui/                React 19 + Vite frontend (built into app/static)
  deploy/            deployment guide and the configuration template
docs/                data sources, providers, widgets, hosted mode
scripts/deploy.sh    build on Cloud Build and roll the Cloud Run service
tests/               pytest suite

Development

Prerequisites

• Python 3.11+, Node with pnpm

• A SEC User-Agent string with real contact details (SEC requires one)

• Your own vendor keys for anything beyond EDGAR and Treasury

• About 1.5 GB of disk for the embedding model, downloaded on first use (or set ALPHADESK_SEMANTIC_SEARCH=off)

Set up and run

pip install -r requirements.txt
cp alphadesk/deploy/env.example .env        # then edit: vault key, SEC user agent
python -m alphadesk.main dashboard          # API + built SPA on http://127.0.0.1:8000

Generate a vault key (32 random bytes, base64):

python -c "import os, base64; print(base64.b64encode(os.urandom(32)).decode())"

For local work without sign-in, set ALPHADESK_AUTH=off (one local account), put your own vendor keys in .env, and seal them into that account:

python -m alphadesk.main keys import-env

Frontend with hot reload (proxies /api to the running server):

cd alphadesk/ui && pnpm install && pnpm dev

Command line

Extending

• Providers: implement a Protocol and register it through a local module (ALPHADESK_PLUGINS) or the alphadesk.providers entry point — docs/providers.md.

• Dashboard tiles: register a widget in ui/src/widgets/, or serve tiles from an external JSON backend (ALPHADESK_WIDGET_BACKENDS) — docs/widgets.md.

• Conventions and checks: CONTRIBUTING.md; the rules and what was tried and undone: DECISIONS.md; reporting a vulnerability: SECURITY.md.

Configuration reference

Environment variables (also read from .env). Only the first two are required.

The full annotated template is alphadesk/deploy/env.example; every setting's default lives in alphadesk/config.py.

Production deployment and operations

The reference service runs on Google Cloud Run (alphadesk, us-east4) with Cloud SQL for Postgres.

Deploying — continuous integration runs the tests, lint and build on every pull request; deploying stays a maintainer's step. After a merge to main:

scripts/deploy.sh

The script refuses unless the checkout is a clean main matching the remote, builds the image on Cloud Build, rolls the service with the new image only (environment, database mount and scaling untouched), and checks that the live page serves the new build. The frontend bundle is committed under alphadesk/app/static, so the image needs no Node build step.

Operations

• Logs: Cloud Logging for the alphadesk service; the ingest loops, pruning and the embedding worker log their progress there.

• Always-on is required as built; approximate cost at this size is $110–120 a month for the service (plus Cloud SQL). The lever for cost is a smaller embedding model, not scaling to zero.

• Graceful shutdown is bounded: a revision stops within seconds.

• Kill switch for search by meaning. If the service ever slows or refuses requests, turn it off without a rebuild — search falls back to words alone:

  gcloud run services update alphadesk --region us-east4 --update-env-vars ALPHADESK_SEMANTIC_SEARCH=off

  Always use --update-env-vars (adds or changes one variable), never --set-env-vars (replaces every variable). The worker's start-up log line reports the cores the machine claims and the thread the model uses (expected: one).

• Background work is shipped switched off, then enabled and watched. A 2-vCPU container reports more cores than it has, so behaviour on a many-core laptop does not predict production: sizing threads from the reported core count once starved the web server of a live instance.

Testing and quality

Known limitations

• Vendor terms. Market data is fetched on your own key, under your own agreement with each vendor, and those terms govern how you may use it. Plans differ — some are personal, some restrict display to others or storage — so check your plan before using AlphaDesk in a hosted or shared setting. Each vendor's terms are linked in docs/data-sources.md.

• Untested paths. Some paid-plan surfaces (Alpha Vantage, Finnhub premium, Polygon paid) were built from documentation and have not been exercised against a live key.

• Search by meaning reads headlines only (summaries would cost several CPU-hours a day per reader), and a new deployment embeds the stored backlog before older stories can match by meaning.

• Legal pages are drafts awaiting counsel.

History

Earlier versions of this repository traded. Two autonomous engines were built, measured against the S&P 500 and deleted in August 2026 (−0.072% mean alpha over 503 backtested trades; −1.123% over 44 live exits); the manual booking and grading layer followed when the product became a consumption terminal. Screener ranking, operator-held data and unofficial sources, the in-app agent and the in-app language model were each removed in turn, each for a measured or stated reason. The one model that remains is the self-hosted embedding model used for search.

Licence

AlphaDesk is dual-licensed. Copyright © 2026 Vignesh Murugan.

Open source — GNU AGPL-3.0 (LICENSE). You may use, study, modify and redistribute AlphaDesk. The AGPL is a strong copyleft licence with one clause beyond the GPL: if you run a modified copy and let others use it over a network, you must offer those users the complete source of your version under the same licence. Distributing copies, modified or not, likewise carries the source with it. Running it for yourself imposes nothing.

Commercial licence. For organisations that want to embed AlphaDesk in a proprietary product, run a modified hosted service without publishing their changes, or need an enterprise exception, a commercial licence is available — contact muruganvignesh0810@gmail.com. The managed cloud is offered under its own terms of service; subscribers take on no AGPL obligations.

Contributions are accepted under a contributor licence agreement, so the project can continue to offer both licences; contributors keep the copyright in their work.

Dependencies are all under permissive licences (MIT, ISC, BSD, Apache 2.0), and no copyleft dependency may be added — it would prevent the commercial licence. The web interface ships its third-party notices at /third-party-notices.txt, written by every build from the packages the bundle actually contains (the fonts are under the SIL Open Font License). The embedding model, Qwen3-Embedding-0.6B, is Apache 2.0.

Data is not code. The code licence grants nothing over market data. Each vendor's terms govern the data fetched on your key; see docs/data-sources.md.