The Stochastic Parrot
Span-grounded audits of how US outlets covered the same story, quotes checkable at the source.
我該用這個嗎
品質與安全性
根據工具定義與協定合規性的自動化分析。
上下文成本
這是每次將伺服器的工具載入模型上下文時所消耗的約略 token 數量。數量越高,可用於其他工作的注意力就越少。
安裝
一鍵安裝
將以下內容加入你的 `claude_desktop_config.json` 檔案:
{
"mcpServers": {
"stochastic-parrot": {
"url": "https://thestochasticparrot.com/api/mcp"
}
}
}遠端端點
https://thestochasticparrot.com/api/mcpstreamable-http它能做什麼
工具清單
工具(15)
🟢search_corpus(query, limit)
Full-text search over every published piece (titles, deks, context lines, outlet names, and source headlines). Returns matching pieces with slug, title, URL, and a highlighted snippet, best matches first. Use the slug with get_audit for full detail.
輸入結構描述
{
"type": "object",
"properties": {
"query": {
"type": "string",
"description": "Search terms, e.g. 'McConnell hospital photo'. Terms are matched as whole words, all required; plain words work best (no special query syntax needed)."
},
"limit": {
"type": "integer",
"description": "Maximum results to return (default 25, hard cap 100)."
}
},
"required": [
"query"
]
}🟢get_audit(slug)
Fetch one piece's full detail JSON by slug: headline, kind, date, context, dek, outlets, verified contradictions with verbatim spans, naming/framing splits, and the full source list. Slugs come from search_corpus or list_audits.
輸入結構描述
{
"type": "object",
"properties": {
"slug": {
"type": "string",
"description": "The piece's URL slug, e.g. 'mcconnell-proof-of-life-photo'."
}
},
"required": [
"slug"
]
}🟢list_audits(outlet, kind, since, until, limit)
List published pieces (summaries: slug, title, kind, date, counts of hard contradictions and naming/framing splits), newest first. All filters optional and combinable.
輸入結構描述
{
"type": "object",
"properties": {
"outlet": {
"type": "string",
"description": "Only pieces whose sources include this outlet, matched exactly but case-insensitively, e.g. 'Reuters' or 'Fox News'."
},
"kind": {
"type": "string",
"description": "Only pieces of this kind: 'audit' (discrepancy audit), 'coverage' (coverage brief), 'dispatch', or 'solo' (solo audit)."
},
"since": {
"type": "string",
"description": "Only pieces dated on or after this ISO date (YYYY-MM-DD)."
},
"until": {
"type": "string",
"description": "Only pieces dated on or before this ISO date (YYYY-MM-DD)."
},
"limit": {
"type": "integer",
"description": "Maximum results to return (default 25, hard cap 100)."
}
}
}🟢list_corrections
The desk's public corrections ledger: every logged error on a published piece — what was wrong, what changed, and when — in original ledger order.
輸入結構描述
{
"type": "object",
"properties": {}
}🟢get_chain(chain_id)
Story chains: multi-piece storylines the desk has covered more than once. Call without arguments to list the available chain ids; pass chain_id for one chain in full (its pieces in order, claim ledger, and reopen conditions).
輸入結構描述
{
"type": "object",
"properties": {
"chain_id": {
"type": "string",
"description": "A chain id from the no-argument listing. Omit to list all chains."
}
}
}🟢verify_piece(slug)
Provenance record for one piece: SHA-256 hashes of its canonical JSON and published HTML as of the last corpus sync. Re-hash the live artifacts at the piece's URL and compare to confirm what you read is what the desk published.
輸入結構描述
{
"type": "object",
"properties": {
"slug": {
"type": "string",
"description": "The piece's URL slug, e.g. 'mcconnell-proof-of-life-photo'."
}
},
"required": [
"slug"
]
}🟢get_framing_index
The Framing Index: across the cross-outlet corpus, how often disagreements between outlets are hard contradictions vs naming/framing splits — totals, rate, monthly trend, per-outlet appearances. The desk's core longitudinal measurement.
輸入結構描述
{
"type": "object",
"properties": {}
}🟢get_coverage_record
The capture-coverage record: what the desk's observatory actually monitored and when, including gaps. Read this before treating any silence in the corpus as meaningful — the desk's own rule.
輸入結構描述
{
"type": "object",
"properties": {}
}🟢get_claim_ledger(chain)
A story chain's claim ledger: specific verbatim claims outlets made, each tracked through states (open, corroborated, contradicted, corrected, quietly_abandoned, resolved) with a receipt for every transition. Empty claims list = the desk has not logged claims for that chain yet.
輸入結構描述
{
"type": "object",
"properties": {
"chain": {
"type": "string",
"description": "Chain slug, e.g. 'mcconnell' or 'fauci'."
}
},
"required": [
"chain"
]
}🟢get_storyboards(chain)
A chain's per-newsroom storyboards: for each outlet, every piece it filed on the story entry-by-entry, the desk's published READ of what each piece was built to do (objective label, motive with its mandatory innocent read, confidence), the label arc, and the desk's 'story, as built' synthesis where one exists.
輸入結構描述
{
"type": "object",
"properties": {
"chain": {
"type": "string",
"description": "Chain slug, e.g. 'fauci'."
}
},
"required": [
"chain"
]
}🟢list_letters
Letters to the Desk: reader mail, excerpted and initialed, with the desk's replies in its own voice. The desk's only interactive surface with its readership.
輸入結構描述
{
"type": "object",
"properties": {}
}🟢get_boxscore(day)
The Daily Box Score for one day — the desk's frozen per-day stat snapshot. Requires the day; days are listed at /boxscore/ on the site.
輸入結構描述
{
"type": "object",
"properties": {
"day": {
"type": "string",
"description": "The day, YYYY-MM-DD."
}
},
"required": [
"day"
]
}🟡submit_letter(text, from, key)
Submit a letter to the desk. If the desk answers, the exchange may be published on the Letters page (excerpted, initialed) — through the desk's normal editorial process, never directly. Requires the reader key.
輸入結構描述
{
"type": "object",
"properties": {
"text": {
"type": "string",
"description": "The letter (plain text, up to ~4000 chars)."
},
"from": {
"type": "string",
"description": "Initials or short byline, e.g. 'M.P.'."
},
"key": {
"type": "string",
"description": "The reader key issued by the desk's operator."
}
},
"required": [
"text",
"key"
]
}⚪report_error(slug, description, key)
Report a suspected error in a published piece (a miscount, a quote that reads wrong, a broken claim). Verified catches are fixed and logged on the public corrections ledger, credited 'Caught by a reader.' Requires the reader key.
輸入結構描述
{
"type": "object",
"properties": {
"slug": {
"type": "string",
"description": "Slug of the piece, if known."
},
"description": {
"type": "string",
"description": "What looks wrong, and where."
},
"key": {
"type": "string",
"description": "The reader key issued by the desk's operator."
}
},
"required": [
"description",
"key"
]
}⚪propose_development(chain, tip, key)
Tip the desk that a story moved — a development on an existing chain or a story worth auditing. Lands in the same proposed-developments queue the observatory's triggers feed; the desk decides what, if anything, to file. Requires the reader key.
輸入結構描述
{
"type": "object",
"properties": {
"chain": {
"type": "string",
"description": "Chain slug, if the tip extends a chain."
},
"tip": {
"type": "string",
"description": "The development, with a source if you have one."
},
"key": {
"type": "string",
"description": "The reader key issued by the desk's operator."
}
},
"required": [
"tip",
"key"
]
}社群
證據