# Parrot (parrot.markets)

> Free data for Kalshi and Polymarket mention markets: transcripts, word history and calibrated probabilities, for people and AI agents. Hubs: earnings calls, Trump, the Fed, White House briefings, news and talk shows, politics, sports broadcasts, business keynotes, entertainment, podcasts, world leaders. Transcripts and history first; prices are an opt-in layer (markets=1). Anonymous and read-only: no key, no signup, no OAuth.

REST: https://parrot.markets/api/v1/...  MCP (streamable HTTP, stateless): https://parrot.markets/mcp
    claude mcp add --transport http parrot https://parrot.markets/mcp

## Start here: one question, one call (sizes in bytes of the default TSV answer, measured on the full data 2026-10-21; '<=' is a budget the endpoint enforces)
1. What is listed in a hub?  GET /api/v1/upcoming?category=trump  ~1,700 (every hub: /api/v1/upcoming ~5,600; <=2,800 a hub-week)  (MCP upcoming)
2. Everything about the next occasion  GET /api/v1/brief?entity=donald-trump  <=3,200: the words that fit, ranked by p; '# rows 1-48 of 54 · more: offset=48' pages the rest (limit=0 = all)  (MCP brief)
3. Will they say this word?  GET /api/v1/word?entity=donald-trump&word=Tariff  ~290, one line (<=480)  (MCP word)
4. Who / what is this?  GET /api/v1/entity/fomc-press-conference  ~1,000 (<=1,600)  (MCP entity)
5. How did this word settle before?  GET /api/v1/history?entity=donald-trump&word=Crypto  ~870 for 8 listings (~70 a listing)  (MCP history)
6. Where was it said?  GET /api/v1/quotes?symbol=NFLX&word=Advertising  ~1,500 (~250 a row, <=5 rows)  (MCP quotes)
7. Search what was said  GET /api/v1/search?q=tariff&limit=3  ~1,500 (10 transcripts ~4,600)  (MCP search + fetch)
8. What changed since my last call?  GET /api/v1/changes?since=<as_of>  ~3,100 a page of 50 rows (~60 a row)  (MCP changes)
Two calls answer "Will the Fed chair say 'uncertainty' at the next press conference?": upcoming?category=fed&days=14 (~500 bytes), then word?entity=fomc-press-conference&word=Uncertainty (~300). An empty upcoming says when the next occasion is ('# next beyond the window').
ChatGPT deep research / connectors: MCP search(query) -> {results: [{id, title, url, text}]}, then fetch(id) -> {id, title, text, url, metadata}; ids are entities, appearances, k:/p: events and <entity>/words/<word>.

## Every hub: category=, an example entity and word (brief?entity=<id> answers once upcoming lists an occasion)
- earnings: GET /api/v1/entity/netflix · GET /api/v1/word?entity=netflix&word=Live%20Event  (transcripts: quotes, tn/th; symbol=NFLX works too)
- trump: GET /api/v1/entity/donald-trump · GET /api/v1/word?entity=donald-trump&word=Tariff  (the hub is the entity; ~20 open events a week: rallies, remarks, windows)
- fed: GET /api/v1/entity/fomc-press-conference · GET /api/v1/word?entity=fomc-press-conference&word=Uncertainty  (scope role:fed-chair = whoever chairs that day; see role-fed-chair)
- white-house: GET /api/v1/entity/white-house-press-briefing · GET /api/v1/word?entity=white-house-press-briefing&word=Border  (scope role:press-secretary)
- politics: GET /api/v1/entity/jd-vance · GET /api/v1/word?entity=jd-vance&word=Border
- news: GET /api/v1/entity/world-news-tonight · GET /api/v1/word?entity=world-news-tonight&word=Trump  (anchors and talk shows)
- sports: GET /api/v1/entity/nba · GET /api/v1/word?entity=nba&word=Triple%20Double  (leagues: the announcers of each game)
- business: GET /api/v1/entity/elon-musk · GET /api/v1/word?entity=elon-musk&word=Tesla  (CEOs, keynotes, posts)
- entertainment: GET /api/v1/entity/love-island-usa · GET /api/v1/word?entity=love-island-usa&word=Bombshell
- podcasts: GET /api/v1/entity/the-joe-rogan-experience · GET /api/v1/word?entity=the-joe-rogan-experience&word=Elon
- world: GET /api/v1/entity/prime-ministers-questions · GET /api/v1/word?entity=prime-ministers-questions&word=Reform
Transcripts (tn/th, quotes, search) by hub: of the word events settled in the 30 days to 2026-10-23, these had text: earnings calls 60 of 61, News 48 of 49, Podcasts 13 of 17. Series that meet less often, over their own last settled occasions: Fed: FOMC press conference 7 of its last 8 occasions, latest 2026-09-16 (other Fed events 0 of 6 in those 30 days). Other hubs show settlement history and p until their transcripts land and pass our checks (/accuracy); blank tn/th = no transcript yet, or the market's speakers are not attributed in the text; never zero.
Clips: earnings calls with verified timings (Castify MP3 at the second); other sources use quotes.

## Ids
- Hubs (category=): earnings, trump (one entity: donald-trump), news, politics, sports, business, entertainment, podcasts, world, fed, white-house.
- Entities are stable slugs: donald-trump, kevin-warsh, nvidia, the-last-word, role-fed-chair. Earnings companies also take symbol=NFLX (an alias of entity=). List a hub: GET /api/v1/entities?category=news
- Events: k:<Kalshi event ticker>, p:<Polymarket event slug>. Appearances: <entity>:<YYYY-MM-DD>:<kind>.
- Words: Kalshi syntax 'Tariff / Tariffs (3+ times)' (Polymarket's 'Tariff 3+ times' works too), or a key ('ai|artificial intelligence', key=); any word, listed or not. A word answers its own history; when the next market lists it inside a wider word, word says 'next <date> <ev> as "Russia / Ukraine"' (next=1 answers that market instead). Plurals, possessives and spoken numbers count; thresholds are never merged.
- Entities resolve by id, name, alias or a unique surname ('Warsh'); an ambiguous one is a 404 that lists 'did you mean' ids.

## Formats
- Tables: format=tsv (default: '# ' head lines with key value pairs and as_of, a header row, tab-separated rows, empty cell = none) | json ({meta, cols, rows}) | md. word: format=text (one line) | json.
- fields=a,b picks columns (unknown names: 400 listing the allowed ones); fields=full gives every column (icons on upcoming and entity). markets=1 adds px (mid 0-1 of the listing venue; may be delayed).
- brief: the head gives each venue's speaker scope; a scope column appears only when a word differs. Default columns leave out a column empty on every row (tn/th before transcripts, p before it is published) and move a column with one value on every row to the head ('same on every row: p_src hist'; json: meta.same).
- Every table and word answer has as_of and an ETag: send If-None-Match to get 304. Poll changes?since=<as_of> instead of re-pulling lists (when a page says more, pass its next cursor back as since=). CORS is open (*) for GET.
- Full text: GET /api/v1/appearances/<id>/text for public-domain / OPL sources (Fed, White House, Congress, UK Parliament); MCP fetch(<appearance id>) includes it.
- Errors: JSON {"error", "detail"}: 400 bad parameter (one short line), 404 unknown id, 403 rights, 429 rate limit, 503 query_timeout. Unknown parameters are ignored and named ('# ignored params: x').
- Search: FTS5 words, "phrases", OR / NEAR (<= 6 alternatives), NOT after a word, prefix* (3+ letters, <= 2 a query; a prefix matching most of the index is refused). '# more: offset=N' pages.

## Columns (one dictionary for every endpoint)
- w: the word as listed by the venue
- key: normalised word key: alternates joined by |, #N = threshold
- thr: count threshold (3+ times = 3)
- scope: who must say it (named, company_reps+operator, anyone, host, role:fed-chair, ...)
- v: venues listing it: K Kalshi, P Polymarket
- n: settled listings (settlement history)
- y: YES among n
- rate: y / n (always next to n)
- l10: last <=10 results oldest -> newest, Y / n
- last_y: date of the last YES
- tn: transcript history: appearances checked (the last <=8 with text of the occasion's kind: rallies for a rally); blank = no transcript yet, or the market's speakers are not attributed in the text (tx / tx_reason say which), never zero
- th: of those, appearances where the word reached thr under scope
- tx: word: the transcripts behind tn/th, '8 rally 2026-07-29..2026-09-17' ('any' = every kind; '*' = too few of the occasion's kind, so every kind)
- tx_reason: word json: why tn/th are blank: no_text | scope_unattributed
- p: Parrot's calibrated probability that the word is said (blank until published)
- p_src: p's inputs: hist = settlement history, tx = transcripts, blend = both, prior = a base rate (no settled history of the word: speaker / category rate; ranked after the rest)
- cnt: after the event: count under scope
- cnt_all: after the event: count over every speaker
- res: yes no void pending; blank = open
- t: media seconds of the first counted hit
- clip: yt:<video id>@<start>-<end> (whole seconds) or mp3:<source id>@<start>-<end> (Castify MP3, 1 dp)
- px: mid price 0-1, only with markets=1
- ent: entity id
- app: appearance id <entity>:<date>:<kind>
- ev: event id k:<TICKER> / p:<slug>
- date: occasion date (ET / local)
- time: HH:MM ET, or bmo / amc for earnings
- n_w: listed words
- top3: word:p<p>:n (calibrated p; n 0 = a prior) or word:r<rate>:n (settled rate, no p yet): the 3 likeliest (word = the listing's first alternate, #N = threshold)
- kind: occasion kind: rally remarks presser keynote earnings_call window_week posts_window ...; '<kind> by' = a deadline market ('the next press briefing', open until date)
- win: upcoming: a window's span start..end, or ..deadline for a 'by' market
- next: word / entity: the next open listing 'date ev'; 'next by <date>' = a deadline market; next_as = the listed word when that market lists a wider word ('Russia / Ukraine' for Ukraine)
- name: display name
- hub: category id
- sub: facet within the hub
- tier: entity: T1 official / labelled text, T2 captions / ASR, T3 settlement history only; capped by the category's settlement QA (/accuracy): below 97% a T1 shows T2, below 90% 'in review'
- role: speaker role in the transcript (company analyst operator host guest reporter ...)
- evs: every event id of the occasion
- icon: icon path
- strike: strike id
- status: appearance / event status
- streak: current run, Y3 / n2
- spk: speaker entity id or printed label
- sec: section: prepared qa monologue interview segment post
- q: quote <=25 words, match in «»; blank for counts-only sources (Internet Archive TV captions) and speakers the captions do not name
- counted: 1 = counts under the market's scope
- src: canonical source URL
- text_kind: official labelled captions asr licensed
- ts type id op: changes feed: time, entity|appearance|source|event|history, id, add|upd|del

## Rules
- Probability: p is Parrot's calibrated model with p_src (settlement history, shrunk toward the speaker's and the category's base rates, blended with transcript history where transcripts are validated). It is a base rate, not a forecast; market prices never feed it.
- Full text only for public-domain and openly licensed sources (US government, UK Parliament); everything else gives counts and quotes of <= 25 words / 200 characters: <= 5 quotes per word (quotes), <= 2 per transcript (search), never overlapping or touching. Internet Archive TV captions are counts and timings only (q blank). Clips are links to the publisher's video (youtube-nocookie) or, for earnings calls, Castify's public MP3 at the right second (D4); never rehosted media.
- Pending text: a transcript that has not arrived says so ('# text pending: <source> (state)' in appearances/<id>, quotes and search; entity says how many later appearances are pending).
- Counts follow each venue's speaker rule (scope): e.g. Kalshi earnings count company speakers (+ the operator when its rules say so), Polymarket counts anyone.
- Parrot data is not financial advice: rates and probabilities are base rates from history, not forecasts. People who run Parrot trade on Kalshi and Polymarket, including mention markets (parrot.markets/disclosure).
- Licence: Parrot-derived data (counts, history, p) is CC BY 4.0: credit parrot.markets; transcript rights stay per source.

## More
- https://parrot.markets/llms-full.txt: every endpoint, parameter and MCP tool
- https://parrot.markets/agents: add Parrot to Claude Code, Claude, ChatGPT, Cursor, VS Code, Windsurf, Codex CLI or Gemini CLI, with live sample answers
- https://parrot.markets/api-guide: API limits, formats and errors; https://parrot.markets/docs: API reference (HTML; https://parrot.markets/docs.md as plain text); https://parrot.markets/openapi.json: OpenAPI
- https://parrot.markets/methodology.md: counting rules, attribution, the p model
- https://parrot.markets/sources.md: sources and rights per category
- https://parrot.markets/terms: terms of use