briskapi

English | 日本語 | Documentation

An unofficial Python API and brisk command line for BRiSK auction data, with Databento-style Live and Historical clients. Consume a live feed, query recordings at any point in time, keep them in your own private archive, and use BRiSK at your broker (SBI Securities, Tachibana e-shiten's BRiSK Next, Matsui, Monex or SMBC Nikko) with your own account, signing in with a passkey if you like. This is an independent project, not affiliated with or endorsed by BRiSK, Tachibana, SBI, TSE or JPX.

Data sources:

• No account needed: the public BRiSK Next demo of 27 September 2021 (one pre-open snapshot and the first three minutes after the open), replayed at its recorded pace. This is not live market data.
• With your broker's BRiSK (SBI Securities, Tachibana e-shiten's BRiSK Next, Matsui, Monex or SMBC Nikko): BRiSK market data (candles, margin, alerts, schedule, watchlist) and a live feed.

Which API should I use?

Recording isn't on this list: brisk live, brisk record and their Python forms record as they run, and what they record is read through the calls above (Recording).

Ticker.history() and .trades() never go to the network. They read the loaded recording or the live feed:

• A recording gives exactly the span it covers.
• A live feed needs history=True. It then keeps every update since it connected, and trades back to the day's start, as far as BRiSK's snapshot restores them.

Ticker.candles() is the call that asks your broker's BRiSK. 5m covers today only; how far back 1d, 1w and 1mo go is up to BRiSK and not yet known. For bars from your own recordings, across days, use Historical (ohlcv-1s to ohlcv-1d).

The local calls and the cloud API name the same data differently (indicative_price locally, ind_match_price in the cloud, for example). docs/data-format.md maps one to the other. briskapi.Live uses the cloud API's names and records.

Install

Free-threaded Python 3.14 (python3.14t) and Node 22+. The feed records the whole market, whose opening auction changes thousands of books within seconds, by reading, parsing and writing on several cores at once, which Python does only without the GIL (a Python with it works, but warns that it can fall behind at the open). BRiSK's own decoder is a WebAssembly module that runs under Node.

pip install 'briskapi[pandas]'      # `import briskapi` and the `brisk` command

BRiSK's decoder and demo data are downloaded at runtime; the package doesn't include them. Optional prebuilt Rust tools (brisk_quote_ingest, brisk_recording) for Linux, macOS and Windows are attached to each GitHub release. To work from source, clone the repository and run pip install -e '.[pandas]'.

On Windows, run from PowerShell with Windows' own Python and Node:

py install 3.14t
py -3.14t -m venv .venv
.\.venv\Scripts\python.exe -m pip install -e '.[pandas]'
.\.venv\Scripts\brisk.exe live --site demo --codes 7203 --speed 0 --limit-frames 3 --no-upload --quotes

Recording

Recording isn't a third API. The live calls record as they run, and brisk live and brisk record are their command-line form, with the same code, options and upload choice.

What you record is then read through the other two APIs:

• On your computer: briskapi.load("days/sbi_20261013_085800-113000"), then Ticker and Market, exactly as recorded; brisk export for NautilusTrader or JSONL; or the Parquet tables themselves (data format). brisk pull (briskapi.pull()) brings an uploaded recording back to disk.
• In the cloud: briskapi.Historical() or brisk query XTKS.BRISK reads every broker day you uploaded, merged across days, with OHLCV bars. Demo recordings aren't in XTKS.BRISK; everyone can query the same demo as XTKS.BRISKDEMO.

While it runs, brisk live also serves its feed to briskapi.Live() on this computer (Live records): the records as they arrive, not a recording.

While a day runs, it is written to disk as it arrives: Parquet files in a spool folder (~/.local/share/brisk/spool), about as large as the finished recording, so a crash loses at most its last half-minute. A broker's day normally becomes two recordings as it runs: the morning session's, kept and uploaded once the lunch break begins, and the afternoon's at the day's end, starting from a snapshot of the book where the morning's ended. A busy session is split again when a completed part brings its recording to 80 million quote rows. The demo's replay is one recording.

• Without --output, each recording goes to your archive and its local copy is removed. With uploads off as well (--no-upload, or --no-upload-market-data for a broker's market recording), that day isn't kept anywhere, so add --output DIR to keep it.
• A lost connection doesn't lose the day. brisk live --site sbi (or another broker's site) reconnects by itself, after 5 s, then 10 s, 30 s, a minute and every 2 minutes. What it recorded before the drop is a complete recording of its own (..._085800-101500), kept and uploaded like a day, and the rest of the day is recorded from a fresh snapshot (..._102000-153512). Only the gap is missing; the cloud API merges the two. It stops for good when another connection can't help: an expired session, another BRiSK window open on your account, or a site that changed.
• Stopping (Ctrl-C, kill, shutting the computer down, feed.close()) ends the day where it is, as the close would: the day so far is a complete recording (..._085800-121500), kept and uploaded like any day. Start again later and the rest of the day is another recording; the cloud API merges them.
• A crash or a forced kill can't end the day, so it is repaired: the day so far (all but its last half-minute, still in the spool) gets the end marker its host never wrote, and is then kept and uploaded as that session's settings said. After a crash this happens when you next start a feed; for a host that doesn't end within 30 seconds of being asked to stop, right away. The repaired end counts the day's quote updates exactly; its frame count and duration, which died with the host, are rebuilt from the recording.

Live feed

import briskapi

feed = briskapi.connect(site="demo", codes=["7203", "6758"])   # returns once initial state is in
toyota = briskapi.Ticker("7203")
toyota.quote()        # current quote, updated as each frame arrives
toyota.auction()      # indicative price/volume and market-order imbalance

feed.on_quote(lambda q: print(q["code"], q["indicative_price"]), codes="6758")
for q in feed.quotes("7203"):       # one item per update; ends with the session
    if q["last_price"]:
        print("opened at", q["last_price"], q["time"])
        break

briskapi.Market().imbalances(top=10).to_pandas()
feed.wait()           # or feed.close(); `with briskapi.connect(...) as feed:` also works

briskapi.connect() takes a site, as brisk live --site does:

• site="demo": the public BRiSK Next demo. It isn't a live connection: the demo site serves a recording of 27 September 2021 as files (BRiSK's decoder, the master, a snapshot and the stream), which are downloaded, checked against pinned hashes and replayed at their recorded pace (speed: 1 real time, 0 as fast as possible; limit_frames stops early). cache=DIR replays a copy of those files downloaded before instead.
• site="sbi", "e-shiten", "matsui", "monex" or "smbcnikko": your broker's live BRiSK (below).

Other options: codes, history=True (keep updates for Ticker.history()), upload, output, on_day and gateway. Callbacks and iterators first receive each security's current quote, then every update in order. A slow consumer slows the feed instead of losing updates.

Live records (Databento-style)

brisk live also serves its feed to other programs on this computer, as the records the cloud API returns. briskapi.Live is shaped like Databento's live client and needs no key or sign-in: only the feed signs in to your broker.

brisk live --site sbi   # records, uploads and serves XTKS.BRISK until you stop it (--site demo: the demo, XTKS.BRISKDEMO)

import briskapi

live = briskapi.Live()          # finds the feed running on this computer
live.subscribe(dataset="XTKS.BRISK", schema="imbalance", symbols=["7203", "6758"])
live.subscribe(dataset="XTKS.BRISK", schema="trades", symbols=["7203"], start=0)   # the day so far, then live
live.subscribe(dataset="XTKS.BRISK", schema="ohlcv-1m", symbols=["7203"])          # each minute's bar as it ends
for record in live:             # or live.add_callback(fn); live.start(); live.block_for_close()
    print(record.schema, record.symbol, record.ts_event)

• Schemas: trades, mbp-10, imbalance, definition, brisk-quotes and the OHLCV bars ohlcv-1s to ohlcv-1d, with the same columns and values as Historical returns for the same session. Live builds its bars from the trades it receives, by the cloud's rules.
• Bars come as soon as their interval ends, by the feed's clock: when a later batch is stamped past the end, or when this computer's clock, less the delay measured from BRiSK to this session (the largest of the last 10 seconds), passes it. A trade that arrives after its own bar has ended goes into the next bar still open (the cloud places it by its own time), so every trade counts and bars come in order. A bar is given only if the session saw its whole span within the trading sessions: with start, from the feed's start; without, the bars that begin after it subscribed.
• Where a session starts: with the day's definitions, then each update as it arrives. start (a time, a date, UTC nanoseconds, or 0 for everything the feed has) first replays the feed's trading day from then: the feed has the day from when it started, plus the day's trades its snapshot restored. snapshot=True begins mbp-10, imbalance or brisk-quotes with each security's current quote (flags 32).
• Records are dicts, also readable as attributes (record.price), with record.schema. They come in the order the feed receives its batches; within a batch, in Historical's order. add_callback(fn, exception_callback) and add_stream(path) (JSON lines; Databento writes DBN) take them too. block_for_close(timeout) waits for the session's end, and stop() ends it. Subscribe before or after start(): a later subscription gets its own replay or snapshot, and its symbols from then on. A session's subscriptions share one dataset. symbology_map maps instrument_id to the code.
• The fast lane: a broker feed records the whole market, which at the opening auction can fall seconds behind the stream on a slower computer. A session that subscribes to its own symbols, without start, gets them on the fast lane: read as each frame arrives, ahead of the recording, with the same records. The lane takes about 500 symbols per worker it has (a quarter of the feed's, so 500 to 1,000 by default). ALL_SYMBOLS, a start replay, or symbols beyond what it takes come at the recording's pace instead (a warning says so and how many it would take); a session whose later subscription no longer fits moves there without missing or repeating a record. For a system of a thousand symbols or more, start the feed with brisk live --site sbi --fast-symbols 1500 (briskapi.connect(fast_symbols=1500)), which gives the lane workers for them: each 500 takes one from the recording, so on a computer with fewer cores also set BRISK_WORKERS.
• The latest quotes of many symbols, at once: a program that decides at a moment (a cut a few seconds before an auction) wants each symbol's freshest quote then, not every update on the way. live.subscribe("XTKS.BRISK", "mbp-10", symbols=codes, stream=False) has the feed keep those symbols fresh on its fast lane and send nothing; live.latest() then returns, in one request, a pyarrow table of their last quotes (mbp-10, imbalance or brisk-quotes records, ts_recv the moment the feed received each, so its age is now - ts_recv): some 18 ms for 1,500 symbols, however busy the stream. A session either streams or keeps quotes for latest(). tools/live_age.py prints how old a stream's records are, split into BRiSK's delay and briskapi's own.
• Only this computer: the feed listens on 127.0.0.1 and admits a session only with the key it keeps, with its port, in an owner-only file (~/.config/brisk/gateway/XTKS.BRISK.json), which Live() reads. Nothing it serves leaves your computer. A session that falls more than 1 GB behind is dropped rather than slowing the feed or its recording.
• Serving: brisk live serves unless you pass --no-gateway; one feed per dataset at a time. In Python, pass gateway=True to briskapi.connect().

Recordings and your archive

recordings = briskapi.recordings(source="historical_mock")   # yours only; no account needed
if recordings:
    briskapi.pull(recordings[0]["prefix"])      # verify, decode and cache; becomes the default
else:
    briskapi.record("recordings/my-session")   # record the demo (uploaded unless you opt out)
# Or briskapi.load("recordings/my-session") for a local recording folder (or an older events.jsonl[.gz]).

briskapi.Ticker("7203").quote(at="08:59:59.9999")             # available pre-open quote, in JST
briskapi.Ticker("7203").history(start="09:00", end="09:01")   # every update in a window
briskapi.Market().snapshot(at="09:00:00").to_pandas()

A recording is a folder of Parquet tables laid out like a NautilusTrader v2 catalog: data/order_book_depths.parquet (every quote update: v2's OrderBookDepth, then BRiSK's other quote fields), data/trades.parquet (every execution, as v2 TradeTick), data/equity.parquet (the securities master, as v2 Equity) and data/custom/BriskBatch.parquet (per-frame timing). The complete demo replay takes about 29 MB. pandas, Polars and DuckDB read the tables directly; brisk export writes a NautilusTrader catalog or JSONL. docs/data-format.md describes every column, with examples for pandas, Polars and DuckDB.

Time queries use each quote's timestamp. Toyota's first demo quote is at 08:59:59.993551 JST; an earlier query raises NotFoundError. Market().summary() returns both start and end, including for local recordings and live feeds. Your archive holds only what you uploaded, under your identity: a random key the tool makes on your computer and never registers (see Your private archive). Listings skip invalid or unsupported recordings with a warning. One sample stays a public download for everyone: a raw wire capture (2026-03-11, from 12:12:40 JST), in the legacy brisk-sbi-wire-jsonl-v1 format the tool no longer reads.

Your broker's BRiSK

The same client serves SBI Securities' BRiSK (sbi.brisk.jp, the default site) and those of Tachibana e-shiten, Matsui, Monex and SMBC Nikko (subsections follow); you need a BRiSK subscription with the broker. For SBI, log in on sbi.brisk.jp in your browser, then pass its session cookies: copy them from DevTools, or use pycookiecheat's chrome_cookies("https://sbi.brisk.jp"). Or let briskapi sign in with your passkey (below).

briskapi.login(cookies={"session_bfaf77a2": "v2.local..."})   # remember=True saves them (owner-only file)
toyota = briskapi.Ticker("7203")
toyota.candles("5m").to_pandas()   # price bars: 5m (today), 1d, 1w or 1mo
toyota.margin(days=30)             # margin balances and stock-lending fees
market = briskapi.Market()
market.turnover()     # turnover and shares outstanding, all stocks
market.lists()        # NK225, recent IPOs, …
market.events()       # basket orders, limit up/down, volume surges
market.schedule()     # trading date, status and session times
market.watchlist()    # your saved codes

feed = briskapi.connect(site="sbi", codes=["7203"])   # live; each session uploaded unless you opt out
toyota.quote()                              # same calls as any feed
# upload_market_data=False uploads no market recordings, upload=False nothing; output="days" keeps each day

A broker feed records continuously and keeps running across days. It normally packages and uploads each day in two recordings, the morning session's once the lunch break begins and the afternoon's (starting from a snapshot of the book where the morning's ended) at the day's end; a busy session is split again at 80 million quote rows, at the next completed spool part. In the cloud API they read as one day. When the site's session is done or it starts serving the next trading date (checked every 10 minutes), or the server closes the stream after the market, the day ends: its last recording is packaged and uploaded, with the day's timing summary (unless you opt out), while the feed waits for the next session and records it from a fresh snapshot. Started when no session is running (after the close, on a holiday), the feed waits for the next trading day's: connect() returns once that day has begun (feed.status is "waiting" meanwhile). Brokers end their sessions at 05:00 each morning, so a feed that waits overnight keeps waiting once its session has expired (it says the session has expired; waiting for new cookies) and takes up the cookies of a brisk login you make after, with no restart. If a passkey is saved for that broker (brisk enroll), the feed signs in again itself, as brisk login does (Chrome opens): up to four tries over the next hour and forty minutes, then it leaves it to you. A saved passkey is one broker's, so a feed for another broker, or with no passkey, says how to sign in. output=DIR (briskapi.connect(output=...), brisk live --output DIR) also keeps each recording in DIR, in a folder named SITE_YYYYMMDD_FROM-TO (its broker's site and its first and last source clock, JST HHMMSS), such as sbi_20261013_085800-113000; every broker's recordings are the same source, brisk_live, as all carry BRiSK's same market data, and the folder names say which broker each came through, so two recording the same session (to compare them) never meet. feed.recordings lists them (trading_date, name, path, uploaded, timing_uploaded, the timing summary's on a day's last, repaired, and error: what failed, packaging or upload), and on_day=callback is called for each. Without output, a recording is kept only in your archive, and with uploads off, nowhere. Stopping the feed (feed.close(), Ctrl-C) ends the day there, so what is left of it is kept and uploaded like any recording; a lost connection reconnects by itself (Recording). SBI and e-shiten allow one BRiSK session per user (expect the same of the others): close BRiSK in your browser before you start the feed, or the two cut each other off within seconds.

Results use the conventions below. Errors are briskapi.brisk.SessionExpiredError (log in again), briskapi.NotFoundError, briskapi.brisk.RateLimitError and briskapi.brisk.APIError. Requests are limited to one per second.

The live feed runs BRiSK's own decoder under Node, downloaded with your session; no browser is involved. It follows the vendor client's connect sequence: it feeds the decoder from the first frame, catches the snapshot up to the stream, answers the server's pings through the decoder and watches the server heartbeat. It hasn't yet been validated against a live SBI session. The stream is expected to be a plain binary WebSocket, as in the Next client. Should the server open with Socket.IO instead, the host joins its /v2/user namespace with unverified parameters that briskapi.connect(site="sbi", profile={...}) can set. See what the server answers with trace_protocol=True (every token redacted); anything unexpected stops with an explicit error rather than a guess. Please report what you see. Live sessions now get through sign-in, the data endpoints, the decoder download, the stock master and the snapshot (loaded with the session schedule). The WebSocket upgrade to api.brisk.jp was then answered with HTTP 502; those sessions didn't ask for the flex subprotocol the vendor client asks for, which this version does, and whether that is enough is not known yet. Your SBI cookies go only to sbi.brisk.jp.

Tachibana e-shiten BRiSK Next

e-shiten's BRiSK Next (next.e-shiten.brisk.jp) is the same app, so the same client serves it. Log in there in your browser and pass its cookies with site="e-shiten", or let briskapi sign in with your e-shiten passkey (below): brisk enroll --site e-shiten once, then brisk login before brisk live --site e-shiten.

briskapi.login("e-shiten", cookies={...})   # or BRISK_ESHITEN_COOKIES; remember=True keeps them apart from SBI's
feed = briskapi.connect(site="e-shiten", codes=["7203"])   # live

Everything above applies, recordings and uploads included (the market recordings as they are made and each day's timing summary); its cookies go only to next.e-shiten.brisk.jp. e-shiten allows one BRiSK session at a time: opening it in a second tab or browser has stopped the first (message 10011), so the feed and BRiSK in your browser can end each other. Before you record or upload, read what Tachibana's BRiSK Next terms say about storing and reusing its data.

Matsui, Monex and SMBC Nikko

BRiSK for Matsui Securities (matsui.brisk.jp), Monex Securities (monex.brisk.jp) and SMBC Nikko Securities (smbcnikko.brisk.jp) is the same vendor's app on hosts of their own, so the same client serves them: site="matsui", "monex" or "smbcnikko", with cookies from your browser, the site's variable (BRISK_MATSUI_COOKIES, BRISK_MONEX_COOKIES, BRISK_SMBCNIKKO_COOKIES; JSON) or a passkey sign-in (brisk enroll --site matsui once, then brisk login).

briskapi.login("matsui", cookies={...})   # remember=True keeps them apart from the other brokers'
feed = briskapi.connect(site="matsui", codes=["7203"])   # live

Everything above applies, recordings and uploads included (source brisk_live, as every broker's, in folders named matsui_…, monex_…, smbcnikko_…), and each site's cookies go only to its own host. These three have not been tried with an account yet: should a site's app differ from SBI's, the feed stops with an error that says what it did not recognise (--trace-protocol shows the connection) rather than record anything doubtful. Expect one BRiSK session per user, as at SBI and e-shiten, and close BRiSK in your browser first. Before you record or upload, read what your broker's BRiSK terms say about storing and reusing its data.

Signing in with a passkey (experimental)

Many brokers now have you sign in with a passkey. If yours does, briskapi can do the sign-in for you and keep the BRiSK session cookies, instead of you copying them from a browser. It works with the BRiSK sites of SBI, Matsui, Monex, SMBC Nikko and Tachibana e-shiten, or any broker you add. It needs Chrome, Node 22+ and a safe place for the passkey: the macOS Keychain or Windows Credential Manager (on Linux, an owner-only file you opt into). It runs on macOS and Windows (run it from PowerShell or cmd so it uses Windows' own Chrome and Node, not WSL's) and is also tested on Linux.

How it works

A passkey is a pair of keys. The broker keeps the public half. The private half normally lives in your phone, laptop or password manager, where a fingerprint or PIN lets it sign a challenge the broker sends, which proves it is you. briskapi borrows that arrangement. It starts its own Chrome, which has a virtual authenticator: software that plays the part of your phone and signs without asking. The broker's site cannot tell the difference.

1. Once, brisk enroll. You pick your broker. A separate Chrome window opens on its login page (a temporary profile, not your everyday Chrome). Sign in as you normally would, then register a new passkey in the broker's security settings, just as you would for a new phone. The virtual authenticator receives it. When the site says it is registered, press Enter in the terminal: briskapi saves the passkey and your broker choice in the macOS Keychain or Windows Credential Manager, and Chrome closes. At e-shiten, enroll first asks for your user ID, because its passkey form wants the ID before the passkey; it is saved with the passkey. The window opens at e-shiten's usual sign-in (user ID, PIN and phone-number authentication), and the passkey is registered on the screen shown after signing in or under お客様情報 › 設定情報 › パスキー認証.
2. Each session, brisk login. The same Chrome opens with the saved passkey loaded, goes to the login page and presses the passkey button, and the virtual authenticator signs. Open BRiSK from the broker's site in that window (or pass --launch-url to have it opened for you). briskapi reads the BRiSK session cookies, saves them and closes Chrome.
3. Then use it. For each of the five brokers, brisk live --site sbi (or its own site) and the briskapi.brisk API use the saved cookies. For a broker you add yourself briskapi has no data client, so the cookies are only saved (under ~/.config/brisk/cookies/) for your own use.

Your existing passkeys on your phone or in a password manager are not touched. The broker simply lists one more passkey, which you can delete in its security settings.

What to know

• The saved passkey is a credential. Whoever has it can sign in to your broker account with no further check. It is kept only in the Keychain or Credential Manager (BRISK_PASSKEY_STORE=file keeps it in an owner-only file instead, which is weaker; on Windows that file's privacy comes from your user profile folder, which only you and administrators can read), never leaves your machine and is never shown in a log. For e-shiten your user ID is kept with it, in the same place. brisk forget deletes it here; remove it at the broker too.
• It is experimental. Each broker's login page, passkey button text and BRiSK launch come from its public pages, and none has been tried against a live account. Override them with --login-url, --passkey-button and --launch-url, or in sites.json below, and please report what fails.
• A broker may refuse it. Chrome tells the site it is being automated. Check that your broker's terms allow this before relying on it.

Commands

brisk sites                    # the brokers you can pick, and where their BRiSK lives
brisk enroll [--site matsui]   # once; you are asked to choose a broker if you don't say
brisk login                    # each session; the broker is the one you enrolled with
brisk forget                   # delete the saved passkey

From Python: passkey.enroll(site="matsui") (for e-shiten, passkey.enroll(site="e-shiten", user_id="...")) and signin = passkey.login(), where signin.cookies are the BRiSK cookies and signin.client is the data client for sites that have one.

To add a broker, or correct a built-in one, edit ~/.config/brisk/sites.json:

{"sites": [{"id": "mybroker", "name": "My Broker", "login_url": "https://broker.example/login",
            "cookie_host": "mybroker.brisk.jp", "passkey_button": "Sign in with a passkey"}]}

Other keys: cookie_prefix, launch_url, enroll_url (where enroll starts, if the first sign-in is on another page than the passkey one) and user_field (a CSS selector for a user ID box that must be filled before the passkey control works; enroll then asks for the ID).

How it works inside: ARCHITECTURE.md.

API reference

Prices are yen floats, with None for the vendor's zero "unavailable" value. Times are JST datetimes on the trading date. Quantities are shares; side, flag and status codes are raw vendor values. raw=True returns vendor fields (*_price10 in tenths of a yen, *_us in microseconds since JST midnight). Tabular results are lists of dicts with .to_pandas(). Errors are briskapi.BriskError and briskapi.NotFoundError. The first whole-market query reads a recording's tables into memory (about two seconds for the complete demo); one security's history is read on its own.

Command line

brisk live --site demo --codes 7203,6758 --quotes   # one JSON object per quote update (--raw for vendor fields)
brisk sites | enroll | login | forget       # sign in to your broker's BRiSK with a passkey through Chrome (experimental)
brisk live --site sbi --codes 7203          # BRiSK at SBI Securities; cookies from BRISK_SBI_COOKIES (JSON) or `brisk login`; --trace-protocol shows the handshake, wire details in BRISK_PROFILE
brisk live --site e-shiten --codes 7203     # Tachibana e-shiten BRiSK Next; cookies from BRISK_ESHITEN_COOKIES (JSON) or `brisk login`
brisk live --site matsui --codes 7203       # Matsui (or --site monex, --site smbcnikko); cookies from BRISK_MATSUI_COOKIES (JSON) or `brisk login`; not yet tried live
brisk live --site sbi --output days         # runs day after day; keeps each recording in days/ (e.g. sbi_20261013_085800-113000) and prints a {"day": {...}} line for it on stderr
brisk live --site sbi                       # records, uploads and serves briskapi.Live() (--no-gateway: serves nothing), with a line on stderr every 10 s (a minute in a log): the day, how long it has recorded, its updates and their rate; and, while it saves a recording, how far that has got
brisk record --output recordings/s1         # record the demo (uploaded unless you opt out); stderr says when it is packaged and saved, and how the upload goes
brisk list --date 20210927 --source historical_mock   # your recordings (--timing: your timing reports; --deleted: what you deleted in the last 30 days)
brisk pull PREFIX --output recordings/downloaded   # use a prefix returned by brisk list
brisk delete PREFIX                         # remove one of your uploads from your archive (restorable for 30 days)
brisk restore PREFIX                        # bring back an upload deleted in the last 30 days (also a timing report's key)
brisk prune recordings/s1                   # free disk: remove a local recording your archive holds
brisk export recordings/s1 --nautilus catalog      # a NautilusTrader v2 catalog, an instrument at a time (or --jsonl FILE, which holds the recording in memory: tens of GB for a broker's session)
brisk query XTKS.BRISK --schema ohlcv-1m --symbols 7203 --start 2026-10-13   # the cloud API: CSV to stdout (--encoding json|parquet, --output FILE, --end, --limit)
brisk consent [--accept | --revoke]         # show or change automatic uploads (--no-upload: this run only)
brisk upload recordings/s1                  # retry uploading a recording (stderr: bytes sent, the archive's checks, then stored)
brisk identity [--recovery-code | --recover CODE]  # your owner ID; its recovery code; use it on another computer

Each command has --help. pull verifies everything before writing and never overwrites an existing folder.

Each command is the command-line form of a Python call:

brisk live --site sbi (or another broker's site) runs until you stop it, normally keeping each trading day's morning and afternoon sessions as separate recordings, with extra splits at 80 million quote rows. Started after the close or on a holiday, it says so and waits for the next trading day. Unless you opt out, it uploads its market recordings as they are made, and each day's timing summary, without asking; --no-upload-market-data stops uploading the recordings and --no-upload both (add --output DIR to keep them on disk). brisk list --source brisk_live (every broker's) lists your broker recordings. Every brisk live also serves its feed to briskapi.Live() on this computer (Live records) unless you pass --no-gateway.

Saving a recording takes minutes, and shows. When a session ends, brisk live says so (session ended; waiting for the next trading day) while the recording is still being saved. Saving reports on stderr (every 5 s on a terminal, 30 s in a log; stdout stays the quotes of --quotes): which recording (sbi 2026-10-13 from 09:00:00), its stage (preparing, writing batches, quotes and trades, finalizing) with the rows written of the total, then saved to PATH in 2m35s only once it is on disk, or Packaging failed after ... and why (no update came until 12:00:03; nothing to save for a stretch in which the stream sent nothing, such as a reconnection in the lunch break). The upload has lines of its own: bytes sent, the archive's checks, and uploaded; the archive stored it. A save that fails is kept in ~/.local/share/brisk/recordings and says so the same way (keeping what was recorded in ..., then saved to ..., with brisk upload PATH to retry an upload that failed). brisk record, package and upload tell their packaging and upload on stderr likewise, and their JSON results stay on stdout.

Your private archive

Uploading gives you a free, off-computer copy of every complete session, checked on arrival (a complete demo replay must match the reference). Once a recording is in your archive, brisk prune DIR deletes the local copy (after checking its SHA-256 against your archive) and brisk pull brings it back, on this computer or any other. Uploading also gives you the cloud API and its OHLCV bars, which exist only there. Each upload is filed under archive/OWNER/YYYYMMDD/FROM-TO-SHA256 (the prefix brisk list returns), FROM and TO being the recording's first and last source clock (JST, HHMMSS), so a day's recordings sort by time.

Uploads are private: only you can list, download or delete them. There is no account. The first time it needs one, the tool makes your identity, a random secret, and the archive files your uploads under a one-way hash of it, your owner ID. The secret is kept twice, in your system's credential store (the macOS Keychain or Windows Credential Manager) and in an owner-only file (~/.config/brisk/identity.json); each restores the other. When it is made, the tool shows its recovery code, 28 letters and digits: write it down. brisk identity --recovery-code shows it again, and brisk identity --recover CODE uses it on another computer (or set BRISK_IDENTITY). Keep it private: whoever holds it is you to the archive, and without it or a saved copy your uploads can't be reached. The archive's operator can access what it stores, as with any hosted service; see PRIVACY.md.

Uploads are on by default. The first time you record or start a live session from the command line, the tool shows what is uploaded and asks once (Enter accepts, n declines), and saves your answer. After that, every clean, complete session (or recording of a live feed) uploads automatically, without asking again. Where nothing can ask, in the Python API or a command run without a terminal, sessions upload by default and the tool says so once per process (a warning in Python), until you decide with brisk consent --accept or --revoke (briskapi.consent(accept=True) or revoke=True).

• What a demo session uploads: local timing measurements (including your computer's clock, which shows when you recorded). A demo replay's market data is the same for everyone, so the archive keeps it once: a demo upload sends only its own timing table (about 0.2 MB instead of about 29 MB), the archive joins it to its verified copy of the demo's market tables, and brisk pull returns the complete recording, verified as before. (The very first demo upload to an archive sends the whole recording.) Your IP address is used only for rate limits, as a keyed hash.
• What a broker session uploads: a timing summary (percentiles of decode time, data age at receipt and frame spacing, a stall count, the frame count, the trading date and the first and last minute; briskapi.Archive().timing() lists yours) and, once a trading day has ended, that day's decoded market recording: securities master, prices, quantities, books, trades, codes and local timing. Cookies, tokens and connection diagnostics are never uploaded. The day's trades also give you OHLCV bars in the cloud, built from your own recordings only (see below).
• Opting out: --no-upload for one run (upload=False in Python); --no-upload-market-data (upload_market_data=False) uploads a broker session's timing summary but not its market recording (--output DIR keeps that); BRISK_UPLOAD=0 for every run in that environment; brisk consent --revoke for good.
• Deleting and restoring: brisk delete PREFIX removes an upload from your archive, so you can no longer list or download it. For 30 days you can bring it back: brisk list --deleted shows what you deleted (with deleted_at and restore_until) and brisk restore PREFIX (also a timing report's key) restores it. After 30 days it can no longer be restored and is removed.
• Partial replays (--limit-frames) are never uploaded. A broker feed you stop mid-day ends that day there and uploads what is left of it, and the part of a day before a lost connection is uploaded once the feed reconnects. A day a crash or forced kill cuts short is repaired with its end and uploaded too (the next time a feed starts, after a crash).

Querying your archive in the cloud

Uploads also give you a cloud API over your own recordings, shaped like Databento's historical API (the same names and parameters). Your archive identity is the API key: there is no signup and no key to create, and briskapi.Historical() finds your identity the way uploads do (and makes one if you have none).

import briskapi

client = briskapi.Historical()
trades = client.timeseries.get_range("XTKS.BRISK", start="2026-10-13", end="2026-10-17",
                                     symbols=["7203"], schema="trades")
trades.to_df()      # a pandas DataFrame indexed by ts_recv; also to_ndarray(), to_csv(), to_json(), to_parquet(), replay()
client.timeseries.get_range("XTKS.BRISK", start="2026-10-13", symbols=["7203"],
                            schema="ohlcv-1m").to_df()     # 1-minute bars (see below)
client.timeseries.get_range("XTKS.BRISKDEMO", start="2021-09-27T08:59", end="2021-09-27T09:03",
                            symbols=["7203"], schema="imbalance")    # the demo, for any identity
client.metadata.get_dataset_condition("XTKS.BRISK", start_date="2026-10-01")   # each date and the clock windows you recorded

get_range(dataset, start, end=None, symbols='ALL_SYMBOLS', schema='trades', limit=None, path=None) returns a store; path also writes its Parquet file. client.metadata has list_datasets(), list_schemas(dataset), list_fields(schema), list_publishers(), get_dataset_range(dataset), get_dataset_condition(dataset, start_date, end_date) (each date available or missing, and the clock windows your recordings cover, in JST), get_record_count(...) and get_cost(...). get_cost is always 0: the cloud API is free, with quotas of 60 requests a minute and 2 GB a day per identity, and 120 requests a minute and 4 GB a day per IP address.

• Datasets follow Databento's VENUE.FEED naming (XTKS is the Tokyo Stock Exchange). XTKS.BRISK is your own live recordings, whichever broker's BRiSK they came through. XTKS.BRISKDEMO is the public demo (2021-09-27), open to every identity.
• Schemas: trades; mbp-10 (ten levels per side; BRiSK gives books, not order events, so action and side are N); definition; imbalance (opening and closing auction state: indicative price and volume, market-order imbalance); brisk-quotes (every BRiSK quote field); and the OHLCV bars ohlcv-1s, ohlcv-1m, ohlcv-1h and ohlcv-1d. list_fields(schema) lists a schema's columns.
• Times: a date means that JST trading day, a time without a zone is JST and an integer is UTC nanoseconds. end is exclusive and defaults to the end of the start's trading day. A request covers at most 31 days and 2,000 symbols.
• Symbols are JPX codes. instrument_id is fixed per code: digits-only codes are their number (7203), others are 100,000 plus the code's base-36 value (167A is 154694).

Historical's OHLCV bars come only from the cloud, never from files on your computer (Live builds them from the stream). The archive builds your bars for each trading day from your own current live recordings of that day, counting each trade once, whenever you upload, delete or restore one, so a query reads them ready-made. They hold the bars your recordings cover: the security, and the bar's whole span within the trading sessions (09:00 to 11:30 and 12:30 to 15:30 JST). A recording from 08:55 to 11:31, for example, gets the morning's 1-minute and 1-hour bars but not the day bar. So uploading is how you get bars. No one else's uploads enter yours, and yours enter no one else's. (XTKS.BRISKDEMO's bars are built from the demo's own trades.) See PRIVACY.md (policy version 8). Ticker.candles() is something else: BRiSK's own bars from your broker, which need a broker login.

On the command line, brisk query writes CSV to stdout. --encoding json or parquet changes that (Parquet needs --output FILE), --output FILE writes a file instead, and --end and --limit bound the request:

brisk query XTKS.BRISK --schema ohlcv-1m --symbols 7203 --start 2026-10-13 > toyota-1m.csv

The HTTP API behind both is described in docs/cloud-api.md. The same records stream live, from the feed on your computer, through briskapi.Live (above).

More documentation

The documentation site has these documents and an API reference with an example of each call and what it returns.

• docs/data-format.md: every column of recordings and cloud API records, and how to read them
• docs/cloud-api.md: how the cloud API works
• ARCHITECTURE.md: how it works, data format, archive integrity and limits
• PRIVACY.md: privacy policy
• CONTRIBUTING.md: development, tests and releases
• tools/brisk_mock/README.md: field definitions, timing and latency, checking the SBI master layout, and the optional Rust collector
• infra/README.md: deploying an archive
• THIRD_PARTY.md: decoder, data and pybrisk attribution

Software is MIT licensed.