An institutional-grade, dual-market order-flow visualizer — real-time liquidity heatmap, DOM ladder, time & sales, and order-flow overlays for crypto and US equities, in one renderer.
Live binance-spot:BTCUSDT — WebGL2 liquidity heatmap, DOM ladder, and tick tape, ~15 minutes
into a session. Columns are rasterized once on the GPU, so pan/zoom cost never grows with history.
FlowMap is a ground-up rebuild of an earlier PyQt6 desktop app that re-rasterized the entire visible history on the CPU every pan/zoom frame, so scrolling back through history collapsed to ~1 fps. FlowMap puts history in a WebGL2 texture and makes pan/zoom a pure view transform — interaction cost is independent of history depth. It renders one market-data engine — Crocodile, which covers crypto and US equities — behind one market-agnostic view.
- 60 fps pan/zoom at any history depth. A column, once uploaded to the GPU, is never re-rasterized; pan and both zoom axes only change a uniform. Measured: draw cost is ~0.2 ms whether 200 or 10 000 columns are resident (history-independent — the old 1-fps bug is structurally gone).
- Professional order flow: liquidity heatmap (inferno ramp — indigo → red → gold → white, so relative size reads by hue and not just brightness — with correct SUM-mip zoom-out so walls don't dilute), DOM ladder, time & sales tape, prominent trade bubbles, a bright last-price line that persists as far back as the book, BBO, VWAP, a CVD (cumulative volume delta) lower pane locked to the chart's time axis, volume profile, event markers, crosshair with exact liquidity readout, deep scroll-back, replay transport. Every overlay toggles individually.
- Chart controls that behave like the tools you already use. The price gutter is a control surface: wheel scales price at the cursor, a vertical drag scales the axis, double-click re-fits. The two axes follow independently — scroll back through time and price keeps auto-scaling; zoom price and the right edge stays pinned to now, with your zoom preserved and the window recentring only when price leaves a central deadband. Re-arming price-follow keeps your zoom (it never re-frames the book out from under you); an in-chart Go Live / Track Price pill appears the moment you scroll off the live edge. Price and time both zoom out far past the grid, so nothing hits an artificial wall. A Tolerance black point (calibrated with a sensible non-zero default) hides sub-threshold density so only liquidity worth reading paints.
- The whole crypto market, not a shortlist. 104 crypto venues are subscribable (plus
equities and the sim feed — 106 rows from
/api/venues). Five have hand-written connectors that stream true incremental book diffs (L2) — binance, bybit, coinbase, deribit, okx — and every other ccxt venue id is served by the universal connector, which re-reads the book whole each tick (L2-snapshot). That difference is a badge in the UI, never a silent downgrade. Symbols are enumerated live from the venue itself, in the venue's own spelling, and a symbol typed in the other spelling (ETH/BTCvsETHBTC) is translated rather than rejected. - A command palette for the whole market. Press ⌘K / Ctrl-K (or
/) for a centre-screen search over every crypto + equity symbol the engine reaches — fuzzy-ranked, with the day's top movers, live price, % change and a mini sparkline, and honest capability chips per row. On first launch you choose from the settings how much history to pull straight onto the chart. - Configurable price coverage, including one that does not compromise. On a linear grid range
and resolution are the same knob:
nativekeeps the finest rows,±50%and−100%/+1000%trade resolution for reach, and the widest of those is a range scan mode.Deepbreaks the tie with a piecewise scale — a linear core at the instrument's native tick, wrapped in logarithmic wings. On BTC at $60k that is ±0.853% at $0.50/row — the exact ladder the narrow grid gives — plus coverage to −99%/+1000% at ~0.34%/row.Deepis now the default, so price zooms out to the full range out of the box;native(finest rows, narrowest coverage) is one click away in the drawer. - Two markets, one renderer, honest tiers: crypto shows full L2 depth + tick tape; US equities
show what their free data actually supports — a keyless two-sided volume-at-price SYNTH depth
(Yahoo 1 m bars, bid below / ask above a reference price that tracks the market) that upgrades to
real Alpaca IEX L1 top-of-book with zero code change when
ALPACA_API_KEY/SECRETare set. Synthetic depth renders in a distinct amber ramp and carries aSYNTHbadge — capability badges (L2/L1/SYNTH,TAPE TICK/TAPE POLL,SIDE EXCHANGE/SIDE NA) are always honest, no fabricated depth.
 equity:AAPL on the keyless SYNTH tier, replaying a recorded market-hours session — distinct amber ramp, honest SYNTH / TAPE POLL badges, market-closed banner, synthetic book in the DOM. |
 sim:SIM-DEMO deterministic feed — liquidity walls, buy/sell trade bubbles, VWAP line, and LIQ liquidation event markers. |
 Crosshair with the exact per-cell readout: source tier, column timestamp, price, and resting bid/ask liquidity. |
 The display pipeline is live: contrast (gamma), colormap, tolerance (black point), normalization percentile, tick grouping, bubble threshold, price range, both follow axes, and per-overlay toggles. |
client/ TypeScript + React + WebGL2 renderer (Vite)
heatmap tile-array + SUM-mips + camera + overlays + DOM/tape + UI shell
server/ Python 3.13 asyncio gateway (FastAPI, binary WebSocket, loopback-only)
time-weighted density grid + sessions + parquet recording/replay
feeds/ crypto (104 venues) · equity · deterministic sim — one router
The client is a pure renderer of a canonical binary stream (docs/superpowers/specs/); the server
normalizes every market into that stream + a capability descriptor. See
docs/superpowers/plans/m1-verification.md, m2-verification.md, m3-verification.md for the
verification record (live Binance + live equity evidence, perf gates, parity matrix).
Two supported paths, and the first one asks you to trust nothing:
- Run from source — below. Python 3.13 +
uv, Node 22 + npm, one script. No binary involved. - Download an installer — self-contained, and verifiable: every installer is signed with a build-provenance attestation that ties it to this repo and the commit it was built from. Check it in one command before you run it.
Every installer on the latest release
is fully self-contained — it bundles the WebGL2 client and a relocatable Python 3.13 running
the server, so there is nothing else to install (no Python, no Node, no uv). Every one of them is
built on its own CPU — there is no emulated build in the list, Windows on ARM included.
Click your platform — the download starts immediately (v1.3.1):
| Platform | Download | Install |
|---|---|---|
| 🍎 macOS — Apple Silicon (M1/M2/M3/M4) | ⬇️ FlowMap_1.3.1_aarch64.dmg · ~184 MB | open the dmg, drag FlowMap into Applications |
| 🍎 macOS — Intel | ⬇️ FlowMap_1.3.1_x86_64.dmg · ~184 MB | same as above |
| 🪟 Windows 10/11 — Intel / AMD (x64) | ⬇️ FlowMap_1.3.1_x64-setup.exe · ~121 MB | run it — installs per-user, no admin needed |
| 🪟 Windows 11 — ARM (Snapdragon X, ARM Surface) | ⬇️ FlowMap_1.3.1_arm64-setup.exe · ~90 MB | run it — native ARM64, not emulated |
| 🪟 Windows x64 — MSI, managed deployment | ⬇️ FlowMap_1.3.1_x64_en-US.msi · ~197 MB | standard MSI installer |
| 🪟 Windows ARM — MSI, managed deployment | ⬇️ FlowMap_1.3.1_arm64_en-US.msi · ~158 MB | standard MSI installer |
| 🐧 Linux x64 — Debian/Ubuntu/Mint | ⬇️ FlowMap_1.3.1_amd64.deb · ~479 MB | sudo apt install ./FlowMap_1.3.1_amd64.deb |
| 🐧 Linux x64 — any distro, portable | ⬇️ FlowMap_1.3.1_amd64.AppImage · ~354 MB | chmod +x it, then run it |
Which Mac do I have? → Apple menu → About This Mac: "Chip: Apple M…" = Apple Silicon, "Processor: Intel…" = Intel. Which Windows do I have? → Settings → System → About → System type: "ARM-based processor" = ARM, anything else = x64.
An installer from a stranger's GitHub deserves suspicion, and "the source is open" does not answer it — the source is not what you downloaded. So every installer is signed at build time with a SLSA build-provenance attestation: a public, Sigstore-backed statement binding that exact file's digest to this repository, this workflow, and the commit it was built from.
gh attestation verify ~/Downloads/<the-file-you-downloaded> -R nazmiefearmutcu/flowmapA pass prints the workflow run and commit that produced it. A failure means the file did not come
from this repository — don't run it. Each release also ships a SHA256SUMS asset, for a check
that needs no tooling: shasum -a 256 -c SHA256SUMS.
Attestations start with the next tagged release; v1.3.0's assets predate them. What the app does on your machine — loopback-only server, no telemetry, no wallet or trading keys, no auto-updater, and the full list of endpoints it talks to — is written down in SECURITY.md.
First launch — unsigned app warnings. The apps are ad-hoc / unsigned and not notarized (code-signing certificates for a paid Apple Developer ID / Windows publisher are not available for this project), so the OS will warn on first open:
- macOS: right-click → Open (confirm once), or run
xattr -cr /Applications/FlowMap.app.- Windows: SmartScreen → More info → Run anyway.
After the first confirmation the app launches normally. Do this after the attestation check above — the warning is about a missing certificate, and the attestation is what replaces the assurance that certificate would have given you.
Installers are produced per OS+arch on CI (native Python wheels can't be cross-built), and each one
boots its bundled server and answers /api/health on that OS before the release accepts it.
Platforms without a prebuilt installer — Linux on ARM and 32-bit systems — can run from source
(below).
On Windows ARM the ARM64 build ships one deliberate difference, for a reason outside this project's
control: pyarrow publishes no win_arm64 wheel in any release, so that build omits it and
cryptography is pinned to the last version that has one. Neither is on a path FlowMap uses —
recording is polars, and the omitted pyarrow only serves an Arrow export FlowMap never calls — and
the release gate imports the full lazy closure on the machine that built it, so the difference is
measured rather than assumed. app/README.md has the table.
Prereqs: Python 3.13 + uv, Node 22 + npm.
./scripts/dev.sh # boots the server (:8720) + the client dev server (:5173)
# then open http://localhost:5173Or manually:
# terminal 1 — server
cd server && uv sync && FLOWMAP_PORT=8720 uv run python -m flowmap_server
# terminal 2 — client
cd client && npm install && npm run devIn the top-bar symbol search: pick SIM-DEMO (deterministic demo feed), a crypto pair
(BTCUSDT → live Binance), or a US ticker (AAPL → keyless two-sided SYNTH depth; real Alpaca L1
top-of-book + live tick during market hours with Alpaca keys).
Optional live tiers (auto-detected from the environment):
ALPACA_API_KEY + ALPACA_API_SECRET → equity L1 tick tape + quotes; FINNHUB_API_KEY → equity
tick tape. Without keys, equities run the honest keyless SYNTH tier.
cd server && uv run pytest -q # gateway: grid, protocol, sessions, feeds, recording
cd client && npm test && npm run e2e # renderer units + Playwright (heatmap, perf gate, parity)The Playwright suite includes the §10 performance gate (history-independent frame cost) and the two-market parity matrix.
What FlowMap connects to, what it writes, what it never asks for, and how to verify a download — SECURITY.md. Vulnerabilities go to a private advisory, not a public issue.
Apache-2.0 — see LICENSE.