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 53 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.

_{Figures in every screenshot are blurred on purpose: they came from a reader's own vendor keys, and vendors restrict public display of their data. The layout is the real thing.}

Chart — your own SVG engine, session bands, drawings	Earnings week — dated by the company's own release and its SEC filing
News — your feeds in one three-day window, searched by word or meaning	Options — chains by expiry, calls and puts either side of the price

Quick start — run it yourself, for one person

You need Python 3.11+ (or Docker; where a command below says python and your shell only knows python3, use that, or the virtual environment's own python), and your own market-data and news keys for anything beyond SEC EDGAR and Treasury data. No account, no sign-in, no database to set up.

pip install alphadesk

python -m alphadesk.main init --email you@example.com

python -m alphadesk.main dashboard

Open http://127.0.0.1:8000, then connect your vendors on the Account page and your agent on Agent access. init runs once: it makes the settings file ~/.alphadesk/.env (owner-only) holding a vault key that seals your stored vendor keys, the contact address the SEC asks for, and single-person mode. Keep a copy of that file — without the vault key the stored vendor keys cannot be opened. The server listens on this machine only.

To run exactly what the cloud runs — Postgres and a login of your own, instead of SQLite and no sign-in — set it up with --like-cloud. It asks for a password (type it in a terminal; it is never printed), makes the settings file with a password hash, and creates the Postgres database if the server is running:

brew install postgresql@16 && brew services start postgresql@16

python -m alphadesk.main init --email you@example.com --like-cloud

Then python -m alphadesk.main dashboard as above, and sign in with that email and password. Use --login for a different login email, --database-url for another Postgres, and --password-hash (from python -m alphadesk.main hash-password) to skip the prompt. For a setup with no typing at all, for example from a script:

python -m alphadesk.main init --email you@example.com --like-cloud --password-hash "$HASH" --database-url postgresql://localhost:5432/alphadesk

What to know about the Postgres path:

• Version. It is built and run against Postgres 16. The driver is a pure-Python one that comes with the install, so there is nothing to compile.
• The database. init --like-cloud creates the database if the server is running and the user may create databases. If it cannot, it still writes the settings and prints what to do, usually createdb alphadesk (or start Postgres first). Tables are created at the first start, and starting again is safe.
• The word-search index. The first start also tries to create the pg_trgm extension and the indexes that make word search fast. If the database user may not create extensions (some managed services), the log says so once and search still works, only without the index. Ask the service to allow pg_trgm, or run CREATE EXTENSION pg_trgm as an administrator, and restart.
• Moving from SQLite. The two stores are separate: switching an existing setup to Postgres starts with an empty ledger. Your vendor keys sealed in the old store are not carried over, so keep the settings file and add the keys again on the Account page, or import them with python -m alphadesk.main keys import-file.
• Backups. The ledger is one database: back it up with pg_dump. Keep the vault key in the settings file as well, because stored vendor keys cannot be opened without it.

With Docker instead:

ALPHADESK_CONTACT_EMAIL=you@example.com docker compose up -d

The first start does the same setup inside a data volume and keeps it there. It also starts Postgres beside the app, the same database the live server runs, so what you test locally is what runs in the cloud; the word search uses a Postgres index there that SQLite cannot have.

What to know about the Docker path:

• The first build is large and slow. The image holds Python, a CPU-only build of the machine-learning library and the 1.2 GB embedding model that search by meaning uses, baked in so a starting container fetches nothing. Expect several gigabytes and a long first build. ALPHADESK_SEMANTIC_SEARCH=off stops the model being loaded when the app runs; it does not shrink the image.
• Two volumes hold everything. alphadesk-data has the settings file (with the vault key) and alphadesk-pg has the database. docker compose down keeps both; docker compose down -v deletes them and with them your history and stored keys.
• Back up the settings file. docker compose cp alphadesk:/data/.env ./alphadesk.env copies it out. Without the vault key the stored vendor keys cannot be opened.
• Update. git pull, then docker compose up -d --build. Settings and data stay. Watch it with docker compose logs -f alphadesk.
• No sign-in inside the container. It runs in single-person mode on 127.0.0.1:8000, as above. For a login of your own use the host setup with --like-cloud, or put an access token in front as described below.
• Agents. Create the token on the Agent access page as usual; the tools answer on http://127.0.0.1:8000/api/agent/tools/mcp. To reach them from another machine, set ALPHADESK_BASE_URL to the address that machine uses and ALPHADESK_ALLOWED_HOSTS for any other name, and keep an access token in front. Reachable from other machines? Sign-in is off in this mode, so anyone who can reach the port can use it and read your connected vendors. Keep it on 127.0.0.1, or set an access token — the browser then asks for it once, and agents keep using the tokens you make on Agent access:

python -c "import secrets; print(secrets.token_urlsafe(24))"

Put the result in ~/.alphadesk/.env as ALPHADESK_ACCESS_TOKEN=... (with Docker, pass it as an environment variable) and restart. Over the internet, put HTTPS in front of it.

Other ways to run it, serving other people, and every setting are under Two ways to run it and Configuration reference.

---

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

---

What AlphaDesk is — and is not

AlphaDesk is	AlphaDesk is not
A consumption terminal: it fetches, checks and presents market information	A trading system — there is no order routing, position state or broker integration
An integration platform: every vendor figure runs on the reader's own key	An aggregator — the server holds no vendor keys and never shares one reader's data with another
A presenter of records: filings, statements, stories and prices, shown whole with their source	A summariser — no model writes, paraphrases or scores anything here
Pluggable: news feeds, market data, transcripts and dashboard tiles are swappable providers	A black box — every list is chronological, alphabetical or ordered by the figure it shows

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

---

Two ways to run it

	Option A — Managed cloud	Option B — Self-hosted
Price	Free during early access. Planned: $19 a month or $190 a year, announced before it starts	Free
Licence	Commercial terms of service — no AGPL obligations for you	GNU AGPL-3.0
You run	Nothing: sign in with the login its operator set	One container or one Python process, and a database
Search by meaning	Included (the embedding model runs on our servers)	Runs on your CPU (~1.2 GB model, 2 vCPU / 4 GiB recommended) or switched off
Vendor keys	Your own, connected on the Account page	Your own, in .env or the Account page
Updates	Continuous	git pull and restart

Option A — Managed cloud

1. Open the hosted terminal and sign in with the email and password its operator set. There is one user per server and no sign-up.
2. On the Account page, connect the market-data and news providers you already use (Alpaca, FMP, Finnhub, Polygon, …). SEC EDGAR and the US Treasury need no key.
3. Optional: connect your own agent from the Agent access page (in the sidebar, beside Account) — 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

Setting in .env	Required	What to put
ALPHADESK_VAULT_KEY	yes	32 random bytes, base64 — seals stored vendor keys; keep a copy, losing it makes them unreadable
SEC_USER_AGENT	yes	AlphaDesk (you@example.com) — a name and a real email
ALPHADESK_AUTH	for one person	off: a single local account with no sign-in
ALPHADESK_ACCESS_TOKEN	if reachable beyond your machine	with sign-in off, one shared secret (16+ characters) the browser asks for once; see Quick start
ALPHADESK_DATABASE_URL	no	a postgresql:// URL (recommended: the Docker setup and init --like-cloud use one, and it enables the word-search index, which needs the pg_trgm extension); unset, SQLite in ALPHADESK_DATA (~/.alphadesk)
ALPACA_API_KEY, ALPACA_SECRET_KEY, FMP_API_KEY, FINNHUB_API_KEY, POLYGON_API_KEY, ALPHAVANTAGE_API_KEY	no	your own keys. Found at start, they are sealed into your account — the local one with sign-in off, or the login's — so the settings file is the record: they add or update keys, never delete them, and win over a key changed on the Account page at the next start. You can still connect keys on the Account page instead
ALPHADESK_SEMANTIC_SEARCH	no	off skips the ~1.2 GB embedding model; search then matches words only. Without the search extra installed it is off anyway
ALPHADESK_LOGIN_EMAIL + ALPHADESK_LOGIN_PASSWORD_HASH	no	with sign-in on: your own sign-in, and the only user the server accepts. With sign-in on it is required; the server will not start without it. At start the account is made, or an existing account with that email gets the password and keeps its data. While set, the password form is the only door. Make the hash with python -m alphadesk.main hash-password; ALPHADESK_LOGIN_PASSWORD (12+ characters) also works but is readable to anyone who can read the settings. The identifier is an email address

Generate the vault key with:

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

2a. Install from PyPI — the shortest route; the built interface ships with the package:

pip install alphadesk

python -m alphadesk.main dashboard

The terminal is at http://127.0.0.1:8000. This install is small and quick and searches news by words. For search by meaning too, install the extra (pip install "alphadesk[search]", about a gigabyte of PyTorch); the first start then downloads a 1.2 GB model into the Hugging Face cache.

2b. Or run from a clone, which is what you want if you intend to change anything:

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

2c. 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, set your login in .env (an email and a password hash from python -m alphadesk.main hash-password); the server will not start with sign-in on and no login set.

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.

---

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

Page	What it shows
Landing (/)	Product overview with blurred screenshots; sign-in
Markets	A composable board: chart, equity overview, funds built on the stock, stock/ETF/crypto/currency/option movers, Treasury yields, heatmap, news. Stock, ETF, crypto and option movers step back to a past session: stocks and ETFs from the whole market that day against the session before it, coins from each day's close against the day before (UTC days, Alpaca's venue), and options from the list recorded after each close — no vendor keeps a past day's busiest contracts, so option history starts on the first evening the server recorded one. Every finished session's close is recorded in your own store (Alpaca or Polygon, whichever your keys reach), the last 15 sessions are filled in at start and each new close is added after the market shuts, and past days are read from the store first. Each finished day's finished list is kept too (big moves checked, volatility, liquidity and names filled, per set of filters; the five days the stepper offers are built in the background), so opening a recorded day asks no vendor anything
Chart	Full workspace: candles, line, area, step and other styles; 1-minute to multi-year intervals; indicators and templates; drawing tools (desktop, per visit); multi-chart layouts; overnight, pre-market, after-hours and weekend session shading; a crowded view draws one candle per pixel column so zooming stays smooth, a range you press shows a loading state, and the percent scale measures from one bar (the first the chart opened with, or any bar you pick from the right-click menu)
Analysis	One name end to end: chart, filings, price performance, key statistics, earnings history and consensus, analysts, rating changes, financials as filed, splits, dividends, institutional and insider ownership, news. Funds add holdings and breakdown
Profile	Who a company is: EDGAR registrant facts, the latest 10-K/20-F business and properties sections verbatim, locations, officers
News	Each reader's merged feeds, three days deep, newest first; filter by words, source or board; search by words and by meaning; an in-page reader with full text where the feed carries it
Filings	What companies just filed with the SEC, newest first, as a reader: a list down the left (time, ticker, company, form, grouped by day) and the chosen filing in full on the right, with every item the company declared in its own wording and the SEC's number, its tickers, and a link to EDGAR. Narrow it by item; move with the arrow keys (or j and k). Nothing is ranked or coloured by importance
Earnings	The week's reporters, dated by the company's own release and joined to its SEC results filing; sessions predicted from history; estimates, actuals, surprise, market cap, volatility, liquidity
Calendars	Economic releases, dividends, corroborated splits and IPOs
Options	Chains with implied volatility, calls green and puts red; options flow seen live
Sectors	Sector funds by weight and dollars traded, with breadth
Baskets	36 curated baskets grouped by the news that moves them (rates, oil, tariffs, chip export rules, bitcoin, …), plus the reader's own
Portfolio / custom views	The reader's saved boards
Account	Coverage matrix of connected vendors and feeds (with the plain-text Export keys download) and sessions
Agent access	Its own page: connect an AI assistant (Claude.ai, ChatGPT) by address, make tokens for programs — each optionally tied to the addresses it may be used from — and ready-to-paste setup for Claude Code, Codex, Cursor, opencode or your own program over the data API
Terms, Privacy, Disclaimer	Public drafts pending legal review

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

Vendor	Carries (plan-dependent)
Alpaca	Consolidated (SIP) and IEX bars, overnight (Blue Ocean) session, quotes, movers, crypto, option chains with IV, options flow, corporate actions
Financial Modeling Prep	Calendars (earnings, economic, dividends, splits, IPOs), key statistics, profiles, analysts, market caps, currencies, press releases, fund data
Polygon (Massive)	Bars, quotes, movers, currencies, options
Finnhub	Company metrics, profiles, earnings calendar and sessions
Alpha Vantage	Company overview, bars

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, 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.

Sources with no key — read, not licensed

Three sources need no key because there is nothing to buy: the reader turns each on with a button on the Account page.

Source	Carries
Nasdaq	Earnings, dividend, split and listing calendars — and today's trading halts, which no vendor in the catalogue carries at all
Yahoo	Charts, quotes and daily history from the public chart endpoint
Social	One public account's posts, read from a public copy of the account. Only text posts are served, and every post seen is kept in an archive, so a search reaches back past the copy's newest hundred

Two rules keep them honest. A keyed vendor is always asked first — every licensed vendor is ordered ahead of every scraped one, so a scraped source answers only where none does. And provenance travels with the answer: the catalogue marks the source unofficial, each payload carries it, the tile subtitle reads "scraped", and the data_sources agent tool resolves any vendor name back to licensed-or-read. The marker protects the reader's reasoning; it is not consent from the site.

No ticker is ever read out of a social post: a ticker inside a post is the author's claim, and attaching it would route an unverified assertion into that symbol's context.

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. The 5-minute news poll also asks the feed for stories tagged with the main coins by name, because crypto is a small share of the newest stories across everything. A feed writes a coin as BTCUSD; those tags are shown as BTC-USD, the way the board writes them, for the coins the account trades and the large ones in the fixed list — on the page and in the agent's news tools alike.

---

Crypto

Crypto comes from Alpaca alone: trading is on Alpaca, so its coins and its venue's numbers are the ones that count. There is no worldwide market-cap order or volume.

• Search follows the account. Every crypto pair the connected Alpaca account can trade is searchable and chartable, including the pairs against a stablecoin or bitcoin (BAT-USDC, ETH-BTC); a coin the account does not carry is not offered. The funds the SEC ticker file lacks (QQQ, TLT, VOO and many more) are added from the account's own asset list, loaded in the background once an hour, so the first search after a start does not have them yet.
• The crypto movers list the dollar pairs only (the stablecoin and bitcoin pairs are the same coins priced another way). Active ranks by dollars traded on Alpaca's venue (price times coins); the stablecoins (USDT, USDC, USDG) appear under All but not under Active, Gainers or Losers. A coin's price is the live bid/ask midpoint when its last trade is stale, and a 24-hour change with no starting price in the six hours before the 24-hour mark is shown as a dash. Volume and liquidity are Alpaca's own venue's, which is thin; the column says so.
• Checking it. Signed in, /api/crypto/tradable lists the coins and pairs the account can trade, and /api/crypto/check tests every pair for search, price bars and a place in the movers.
• A coin chart uses real trades for its live edge (the bid/ask midpoint moves many times a second and made the chart shake), and fetches one spare day of bars, not the three a stock needs for weekends.

---

Agent access (MCP)

AlphaDesk exposes the same records the interface shows as 53 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 Agent access page, in the sidebar beside Account (shown once, stored hashed, revocable at once); the page gives each client's exact setup, and a tab for your own program using the data API below.

Tools

For	Tools
Today	market_today, index_board, movers (stocks, ETFs, indices, crypto, currencies, options, bonds), sector_performance, sector_breadth
The reader's names	my_board, quotes, baskets, find_symbol
One company	quote (with the order book), key_stats, company_profile (including a share count, with its basis), fund_profile, analyst_view, financial_statements, filed_report, earnings_history, symbol_events, ownership, insider_activity, peers, compare_metrics, related_funds
Why it moved, and what next	what_moved (the big-move days with the stories and filings leading into each), move_state (has the move held, faded or turned), priced_in (its past reactions to reports, the run into them, analyst targets), candidates (dated reasons to move in the next trading session — Friday evening and weekends answer for Monday), movers_in_context (gainers or losers with their own news, filings and shape), news_scan (the news window grouped by name), related_assets (what a company's own filings tie it to, such as a token it holds), entry_facts (spread, liquidity, tradability, levels and risks before entering a position — read-only, no order is ever placed)
Prices	price_history, price_chart
News	symbol_news, news_search, news_story, news_scan
Filings and calls	list_filings, filing_text, transcripts, transcript_text
Calendars	earnings_calendar (upcoming, and reported with days_back), economic_calendar, corporate_calendar
Options	option_expirations, option_chain, options_flow; option_chain(on=…) reads a past day's chain as it stood at that day's close (the board's stocks, saved after each close from 2026-10-07; without an expiry it lists the expiries saved that day)
What just happened	catalysts (filings, halts, government action and social posts on one tape), filing_feed, trading_halts, government_actions, social_posts
A past session	movers(session=…) for stocks, ETFs, crypto and options, with market_sessions(category=…) for the days that exist; read from the recorded closes first. A big move explained by a corporate event is marked: a spin-off from Alpaca's corporate-actions feed carries the holder's move (the parent plus the new shares handed out), and a move with no listed event but the pattern of one (volume ten times its average, the next close within 15 percent) is flagged possible_corporate_event
Provenance	data_sources — whether a figure came from a licensed vendor or a scraped page

A quote that is not a market. A quote whose symbol has no daily bar in the last ten days — a retired ticker or a long halt — carries stale: true: its price is the last trade there was, as_of is that trade's date, and there is no change. The company may trade under a new ticker; find_symbol finds it.

Which address. The tools answer only on the names the server is set to answer to: the public address in ALPHADESK_BASE_URL, plus any names in ALPHADESK_ALLOWED_HOSTS (a comma-separated list). A host that is not on that list is refused. On Cloud Run, which serves one service under two addresses, put the second in ALPHADESK_ALLOWED_HOSTS.

Measures, not verdicts. The tools return figures and plain-fact flags — a move's size, how much of a swing was given back, whether a story is about the name or only lists it — and never a score, a rating or a recommendation. Each reply says what it could not read (unavailable, reliable), and a refusal says what happened: no key, a plan that excludes the data, or connected vendors that simply have nothing for that company (small companies often have no analyst coverage). The reader's price plan is stated in data_freshness: a free plan's volume is one exchange's alone and understates the market.

A smaller tool list for a coin agent. The full list is 53 tools and about 66,000 characters of descriptions on every connect. Send the request header X-Agent-Toolset: crypto and only the 17 tools that apply to coins are listed (about 28,000 characters); a tool left out of the list still answers if called by name. A coin is written BTC-USD; BTC/USD and BTCUSD are accepted too, but a bare BTC means a listed fund. The answers that carry a price or a tradable flag (quote, quotes, movers, price_history, price_chart, entry_facts) include data_venue, the broker those figures came from (today alpaca; unknown when none can be named), so an agent that trades on one broker can see whose market it is reading. A token may make 120 calls a minute; set ALPHADESK_AGENT_RATE_PER_MIN to change that.

Watching how agents use it. Every tool call is logged: the tool, its arguments, how long it took and how it ended (answered, empty, incomplete or failed, and why). A caller can add the request headers X-Agent-Task (one id for all the calls of one question) and X-Agent-Intent (the question in a line), and can say whether a result helped by sending a small JSON message to the same address with /feedback in place of /mcp, using the same token: {"tool": "what_moved", "useful": "yes" | "partly" | "no", "note": "…", "missing": "…"}. That is the only write on the agent door, and it writes to the usage log and nothing else. Read the report with python -m alphadesk.main agent-usage --days 30, or /api/agent/usage?days=30 while signed in: which tools are never called, which come back empty or fail, which are slow (and the twenty slowest single calls, each with its time, tool, arguments and duration), which chains of calls repeat (a tool that should exist), and what agents said they were missing. The log never holds a token, a key, or the agent's reasoning or final answer.

Trying the tools yourself. scripts/agent_probe.py signs in the way an agent does (with a token saved in ~/.alphadesk-agent-token) and prints what the main tools return for a symbol: python scripts/agent_probe.py NVDA. Point it at another server with ALPHADESK_PROBE_URL. It only reads.

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. Every tool carries readOnlyHint, so a client can call them without asking permission for each read.

The same tools as a plain HTTP API. For a program that is not an agent — a trading bot, an importer — the tools are also GET endpoints under /api/v1, with the same token, the same per-reader keys and the same limits, plus every bar of price history (the agent tools thin it) paged by time, and an optional list of addresses a token may be used from. It is read-only, and polling rather than streaming: see docs/rest-api.md.

---

Accounts, security and privacy

• Sign-in: one user per server. The only door is the email and password the operator sets in the settings (ALPHADESK_LOGIN_EMAIL with a password hash); there is no sign-up and no third-party sign-in (Google, GitHub and Microsoft sign-in were removed on 2026-10-03). Any other account is refused.
• 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. The one way a key leaves is the reader's own Export keys download on the Account page: a plain-text file (no passphrase, by the owner's choice on 2026-10-02, so anyone holding the file holds the keys), available only to their own browser session — never to an agent token or an OAuth grant — signed in within the last ten minutes, limited to five an hour and logged by count, never contents. A passphrase sent to the route still seals the file (scrypt and AES-256-GCM, at least 12 characters).
• Agent credentials: tokens and OAuth codes stored as SHA-256 hashes; OAuth clients sealed; consent page signed, same-site and unframeable.
• Retention — the defaults, all lengthened by ALPHADESK_KEEP_DATA: 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, and the kept copies of slow-changing vendor answers (history, fundamentals, profiles, ratings, calendars) 14 days. Those copies make a restart warm, spare your vendor rate limit, and answer for a vendor that is down; live answers (quotes, movers, the tape, the intraday chart tail) are never kept. Recorded session closes — each finished session's whole-market daily bars, about 200 KB a session compressed — are kept for good whatever the retention setting (only ALPHADESK_PURGE_ON_KEY_REMOVAL or deleting the account removes them): a finished session never changes, and the record is what keeps past-session movers working. Removing a key keeps what it fetched unless ALPHADESK_PURGE_ON_KEY_REMOVAL is set.
• 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) ┤                                              │
  Your own bot (HTTP) ─┤                                              │
                       │  /api/*  ── panels, composed per request     │
                       │  /api/agent/tools/mcp ── 53 read-only tools  │
                       │  /api/v1 ── the same tools, GET only, HTTP   │
                       │  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
                       │  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), the plain-HTTP data API (rest_data.py)
  providers/         the plugin seams, vendor implementations, catalogue, per-reader router
  ingest/            EDGAR, news polling, calendars, movers, recorded session closes, prices and indicator math
  desk/              screener window, filings, transcripts, market-today
  ledger/            store (SQLite / Postgres), database adapter, key vault and the sealed key export
  mcp_server.py      the agent tools
  semantic.py        search by meaning (self-hosted embedding model)
  cryptonews.py      a coin's news selection, and the feed's coin tags written as the board writes them
  cryptocheck.py     the check that every crypto pair is searchable, has bars and reaches the movers
  newsquery.py       the word-search rule
  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, only with the search extra, 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

Command	Purpose
python -m alphadesk.main dashboard	The web server and background loops
python -m alphadesk.main keys import-env	Development: seal .env vendor keys into the local account
python -m alphadesk.main keys decrypt FILE	Print the keys in a file exported from the Account page (asks for its passphrase)
python -m alphadesk.main keys import-file FILE	Seal the keys in an exported file into the local account, under this instance's own vault key
python -m alphadesk.main earnings	Stamp today's EDGAR results releases and list the last three days
python -m alphadesk.main backfill --hours 72	Backfill EDGAR results releases
python -m alphadesk.main calendar-accuracy --days 30	Score calendar vendors against EDGAR release days
python -m alphadesk.main mcp [--http]	The agent tools standalone (EDGAR only — no reader identity)
python -m alphadesk.main init --email you@example.com [--like-cloud]	First-time setup: vault key, SEC contact and settings file. --like-cloud writes Postgres plus a login (asks for a password) instead of no sign-in
python -m alphadesk.main hash-password	Make the password hash for ALPHADESK_LOGIN_PASSWORD_HASH (asks for the password)

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.
• Is this the right tool for you, and how it compares: docs/comparison.md. Why the data runs on your own keys: docs/your-own-keys.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.

Variable	Purpose
ALPHADESK_VAULT_KEY	Required. 32 bytes, base64: seals every reader's keys. Losing it makes them unreadable
SEC_USER_AGENT	Required. Descriptive User-Agent with contact details, as SEC asks
ALPHADESK_DATABASE_URL	Postgres connection string; unset uses SQLite in ALPHADESK_DATA
ALPHADESK_DATA	Data directory for SQLite (default ~/.alphadesk)
ALPHADESK_AUTH	off for a single local account without sign-in
ALPHADESK_ACCESS_TOKEN	with sign-in off, a shared secret (16+ characters) that guards the browser; ignored where accounts gate
ALPHADESK_LOCAL_USER_EMAIL	with sign-in off, act as this existing account instead of a fresh local one — for a server that began with accounts and became one person's own; nothing is moved
ALPHADESK_OAUTH_REDIRECT_HOSTS	optional: only these addresses (and their subdomains), comma-separated, may receive a Claude.ai or ChatGPT connector grant; any other is refused at the consent page. Unset, any address may, and the page leads with the address and warns on one it does not recognise
ALPHADESK_ALLOWED_HOSTS	extra names the server answers to. With sign-in off and no access token it answers only to localhost, 127.0.0.1 and [::1] (so a web page cannot reach it by rebinding its own name); list any other name you reach it by, comma-separated. ALPHADESK_BASE_URL's host is allowed too
ALPHADESK_AGENT_RATE_PER_MIN	calls per minute one agent token may make (default 120)
ALPHADESK_AGENT_HEAVY_CONCURRENCY	how many of the heavy agent tools (what_moved, related_assets, news_search, movers_in_context, candidates, earnings_calendar, entry_facts, options_flow, priced_in, analyst_view) run at once; more wait their turn, never refused (default 2)
ALPHADESK_KEEP_DATA	forever or a number of days: keeps the records (stories and their text, announcements, the forecast log, scraped pages) that long instead of the defaults of 3 to 120 days, and reaches further when the store has less (a key save refills 30 days, "Load older" looks 30 days back, a symbol's own ask a year). It only lengthens a default. A vendor's own terms about storing its data still apply to you; this setting does not change them
NEWS_BACKFILL_DAYS	how many days a key save refills (default: the retention window, 30 with ALPHADESK_KEEP_DATA, never over 365)
ALPHADESK_SKIP_KEY_CHECK	1 saves a news key without trying it at the vendor first (offline)
ALPHADESK_PURGE_OTHER_ACCOUNTS	maintenance: count logs how many accounts besides the login exist; delete removes them, whole, at the next start (unset it afterwards)
ALPHADESK_PURGE_ON_KEY_REMOVAL	1 deletes a vendor's stored data when its key is removed (for an instance serving other people whose vendor terms require it); off by default
ALPHADESK_BASE_URL	Public URL; OAuth redirects and the agent host allowlist depend on it
ALPHADESK_SECRET	Session signing secret
ALPHADESK_COOKIE_SECURE	Secure cookies (on behind HTTPS)
ALPHADESK_SEMANTIC_SEARCH	off disables search by meaning
ALPHADESK_EMBED_MODEL / ALPHADESK_SEMANTIC_THRESHOLD	Embedding model (default Qwen/Qwen3-Embedding-0.6B) and similarity cutoff (0.50)
ALPHADESK_PLUGINS / ALPHADESK_WIDGET_BACKENDS	Provider plugins; external tile backends
NEWS_REFRESH_MINUTES / NEWS_LOOKBACK_HOURS / NEWS_KEEP_DAYS	News poll interval (5), window (72 h), retention (7 days)
CHART_MIN_COVERAGE / CHART_MAX_MEDIAN_GAP_MIN	The indicator coverage gate
DASHBOARD_HOST / DASHBOARD_PORT	Web server bind (Cloud Run injects PORT)
MCP_HOST / MCP_PORT	Standalone agent server bind (default port 8010)

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.

Setting	Value	Why
Size	2 vCPU, 4 GiB	The embedding model and the web server share the instance
Instances	exactly 1 (min = max = 1)	One writer for the background loops and live sockets
CPU	always allocated, startup boost	Background loops run between requests
Image	Python 3.12 slim, PyTorch CPU build, model baked in (~1.2 GB)	Nothing is downloaded at start; model loads in seconds

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.

• If requests stall. Measure from the request logs and ignore each version's first five minutes: every deploy is a cold start. The stalls seen so far came after 15 to 25 minutes of use as memory climbed from about 35 percent to over 80 of the 4 GiB. Three things kept them away: the live stream subscribes in one message per group, the background panel refresh runs 15 panels at a time with a pause (at full pace only in the first five minutes), and expired price answers leave the cache on each insert. If memory climbs above about 75 percent again, raise it with gcloud run services update alphadesk --region us-east4 --memory 8Gi (this changes only the memory).

• 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

Check	Command	Scope
Backend tests	python -m pytest -q	~720 tests: providers, calendars, EDGAR parsing, news rules, search, agent tools, auth, accounts, retention
Frontend tests	cd alphadesk/ui && pnpm test	Pure logic under src/lib/__tests__ (chart scales, sessions, news matching, layouts)
Type-check and build	cd alphadesk/ui && pnpm build	tsc -b (project references — tsc --noEmit checks nothing here) then Vite
Lint	python -m ruff check alphadesk	Python

---

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

AlphaDesk became a consumption terminal by subtraction. Screener ranking, operator-held data and unofficial sources, the in-app agent and the in-app language model were each removed in turn, every one for a measured or stated reason — see DECISIONS.md. 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.