Egress, end to end

Overview

Egress answers one question: what does it cost to leave a tokenized US stock position, at a stated size, right now. Entering is easy and every venue advertises it. Leaving is the half nobody prices, and on a tokenized equity it is the half that moves, because the share underneath stops trading at the New York close while the token does not.

The project is two halves that never touch. A crawler reads every listed instrument on a fixed interval and appends what it saw to an append-only record. Everything else - the phase table, the validation, the estimator, this site - reads that record afterwards. Collection never imports interpretation, so a change in how a number is read can never change what was recorded.

No credentials anywhere. Every venue endpoint used is public. The crawler runs unattended on a shared runner with nothing to leak. The only secret in the project is the reader key for the desk, and it lives in the serverless function, never in a page.

Quickstart

Python 3.10 or newer, tested on 3.10 through 3.13. No runtime dependencies at all - the standard library does the HTTP, the gzip and the CSV. Ruff and mypy are dev tools, not requirements.

git clone https://github.com/Ritapossible/Egress.git
cd Egress

# one pass over the whole listed universe, written to state/
python -m egress.crawl --once

# what the record holds so far
python -m egress.crawl --coverage

# what leaving 25,000 USDT of TSLA costs, from the live book
python -c "from egress import exitcost; print(exitcost.for_symbol('RTSLAUSDT', 25000))"

# rebuild this site into docs/
python -m egress.page

A continuous record is a long loop rather than a long process: python -m egress.crawl --loop --hours 6 takes a snapshot every 5 minutes for six hours and exits cleanly, which is what the scheduled job on the repository runs.

Architecture

Twelve modules, each with one job, listed in the order data moves through them.

The reader compiles, the code computes. The language model is only ever asked for a symbol and a notional. It is never asked for a price, a spread or a cost, and it cannot select a ticker that is not listed on the venue right now - resolution is a lookup against the live instrument list, so an invented ticker fails loudly instead of being priced.

Data reference

The record lives in state/. Each snapshot is one gzipped CSV under state/snapshots/, and state/manifest.jsonl holds one JSON line per completed snapshot. A snapshot absent from the manifest is skipped on read even if its file exists.

Current record: 8,581 snapshots, 26,834,943 rows, 142 gaps in coverage, across 3,602 listed instruments (3,019 tokenized stocks, 581 crypto pairs, 2 metals).

Silence is not zero. A symbol missing from a snapshot means the venue did not report it, not that its spread was nothing. Missing rows are absent from every median rather than counted as a value.

What "3,019 tokenized stocks" counts. Rows with symbolType == "stock" from GET /api/v3/market/instruments?category=SPOT — the venue's own classification, not ours. All 3,019 are status: online, all are USDT-quoted, and there are 3,019 distinct base coins, so no symbol is counted twice.

Why another count will disagree. The v2 endpoint most people reach for, /api/v2/spot/public/symbols, returns the same 2,710 rows but carries no symbolType field at all. Without it the only way to split the list is the ticker prefix, and that is wrong in both directions: 25 ordinary crypto pairs start with R (RAYUSDT, ROSEUSDT, RLCUSDT — iExec, not Royal Caribbean) and two tokenized stocks carry no prefix (PRESPCXUSDT, PREOPAIUSDT). The prefix heuristic returns 2,150. Ours returns 3,019 because it reads the field.

Bitget Skills

Two are wired, and both are on the Ask path. bitget-mcp-server answers equity_price_quote behind every desk answer — it is where the basis against the listed share comes from, and without it this site could price an exit but not say what the position was worth. bitget-signal is asked news_feed/latest for the name in the question, on a six-second fuse, and every answer carries what it said. It cannot move the number. The cost comes from the venue's own book; the news block is built after that number is already fixed, and a test asserts the whole answer is identical whether the service replies, returns nothing, or dies. Today it returns nothing, and the desk prints that rather than a briefing assembled out of an empty feed.

bitget-signal, probed 2026-10-22T10:44:48+00:00. 3 of its catalog calls answer; 0 of its live-data calls return anything. The aggregator lists 44 feeds and reports zero articles from all of them, and crypto_price fails with the service’s own ConnectTimeout — so the Skill is running and its outbound fetches are not.

It would not help here even if it worked. Of those 44 feeds, the ones that could carry a US stock story at all are: cnbc. The rest are crypto, tech and general news. This cannot explain why a tokenized stock is wide; it is a control, not an explanation, and it is on this page as a measured result rather than a logo.

HTTP API

One endpoint, used by the desk on the front page. It exists because a reader key cannot live in a browser.

POST /api/ask
Content-Type: application/json

  "q": "What does leaving 40,000 USDT of TSLA cost?"

The response states a symbol, the compiled size, the quote walked off the live book, a slicing bound, and a plain-English reading. Every failure is a stated answer rather than an exception: an unreadable question, a ticker that is not listed, or a venue that will not answer each come back as a sentence saying so. Bodies over 4 KB are refused.

Estimated, never observed. Every figure the endpoint returns is marked estimated. It is what the displayed book says an exit would cost, not a fill anyone received.

Command line

Limitations

Stated here rather than buried, because a research tool that hides its limits is a marketing tool.

• A quote is not a fill. Displayed size can be withdrawn before anyone reaches it, and a real order moves the book it was measured against. Every cost here is an estimate of the displayed book.
• The record is short. The snapshot count behind every table is printed next to it. Nothing here is a claim about the future, and a handful of days is not a study.
• Replenishment is unobservable from a snapshot. That is why slicing is answered as a best and a worst case rather than a number.
• Two of the venue's volume feeds disagree on tokenized stocks. Which one is right cannot be settled from outside, so those symbols are excluded from validation and named on the validation page.
• Idle and phantom liquidity look identical. Displayed size going unconsumed could mean it was withdrawn or that nobody wanted it. This data cannot separate them.
• Research only. Not advice, not an offer, not a quote.

Glossary