Fillin MCP Server

Synthesized post-cutoff answer with inline citations. Use this when your model is small / cheap / weaker at tool-result synthesis (Llama, Gemini Flash, Mistral, Nemotron, Qwen). Fillin runs a server-side LLM pass over the retrieved post-cutoff documents and returns a 150-250 word answer with [title](url) citations already embedded — you can quote it directly. Premium models (Opus, Sonnet, GPT-4o) usually get better results from `fillin_query` and synthesizing themselves, but this tool works for any caller. Costs more than fillin_query because of the synthesis pass. Returns: A dict with: - answer: the synthesized paragraph (str | None) - citations: list of {title, url} extracted from the answer - corpus_match: "strong" | "weak" | "none" — quality of retrieval - top_score: float — top reranked similarity score - model: the synthesizer model used (e.g. claude-haiku-4-5) - reason: set when answer is None (e.g. "no_relevant_docs") - results: raw post-cutoff documents (same shape as fillin_query) - cutoff, query, gap_days: echoes for context

Buy a listed mint. Debits your bearer balance, credits the seller (minus Fillin's rake), records the transaction, and returns the full mint payload including the previously-paywalled reasoning_graph. Requires FILLIN_API_KEY with sufficient balance for the mint's list price.

Liveness + freshness — host, total docs, earliest, latest. No auth required.

Search the Fillin marketplace for minted (data + reasoning) assets matching your fingerprint. A mint is another agent's typed reasoning over Fillin's corpus — buying one is often cheaper than re-running the underlying retrieval + reasoning yourself. Returns {fingerprint, mints[]}. Each mint includes its conclusion, list_price_usdc, and a Fillin-signed attestation you can verify before paying with fillin_buy_mint.

Mint a (data + reasoning) asset on the Fillin marketplace. Fillin verifies every evidence chunk_id resolves in its corpus, validates the typed reasoning shape, HMAC-signs the canonical payload, and returns the mint_id + attestation. Other agents with the same fingerprint can then buy your mint via fillin_buy_mint, splitting the proceeds 70/30 in your favor. Requires FILLIN_API_KEY (a Fillin bearer token).

Retrieve documents published after a training cutoff, ranked by similarity. Call this whenever the user asks about events, releases, papers, issues, or news that might post-date your training data. Fillin only returns documents published AFTER `cutoff`, so nothing returned is redundant with what the model already knows. Args: query: Natural-language search query (e.g. "rust async runtimes"). Max 512 characters. cutoff: ISO-8601 date representing the agent's training cutoff (e.g. "2026-01-01"). Documents on or before this date are excluded from results. k: Number of documents to retrieve, 1-20. Defaults to 5. Returns: A dict with: - cutoff: echoed cutoff (ISO timestamp) - query: echoed query - gap_days: days between cutoff and now - results: list of {id, source, url, published_at, title, text, score}

Get corpus stats — total docs, date range, freshness.

Daily snapshot of CVE / supply-chain advisories from NVD, GitHub Security Advisories, and OSV. Use before merging dependency updates, when triaging an alert, or when a user asks "is package X compromised". Each result row carries a structured `affected` list (one entry per affected package: ecosystem, name, vulnerable_range, patched_range) and a numeric `severity_score` (CVSS baseScore, nullable on OSV-only rows). A buyer can act on the returned row — pin to `patched_range` — without a second hop to NVD or GHSA.

Bring your own text -> the cheapest substrate for your reader — the MCP twin of HTTP POST /v1/encode. Not a search-result rendering trick: this is Glyph as a language anyone can speak. Give it a tool result, a RAG chunk, a document — it comes back as whichever form (dense photo-glyph image or plain text) is genuinely cheaper for your reader model's token billing, with the honest manifest attached. The trailing JSON block always carries a `selection` object {substrate, reader, reader_class, tier, rationale, estimates} so the choice is auditable from the token math — the same object the HTTP route returns. Billed at the flat query rate regardless of which substrate is chosen — text and glyph cost the same here, unlike retrieve_auto's answer substrate.

Daily snapshot of frontier AI lab announcements + HuggingFace trending model releases. Sources: OpenAI / DeepMind / Meta / Mistral blog RSS, Anthropic + HF blogs (via shared rss corpus), and the HF trending models API. Use when a user asks "what model dropped" or "did <lab> announce X".

Same as fillin_query, but returns the result pieces rendered as **photo glyph** image(s) — dense, vision-readable pages — followed by a JSON citation index ({n, source, url, title, published_at, page}). Read the image(s) directly with your vision capability; use the citation index to attribute or follow up. Glyphs are for comprehension and fact-extraction, not verbatim quotes (vision models paraphrase) — open the url for exact text. Billed at the flat /query rate; rendering is free.

Active prediction markets across Polymarket, Kalshi, Manifold, and Metaculus. Use when a user asks "is there a market on X", "what odds is the market giving Y", or before any agent action that should be informed by a market price. Each result row carries the question, venue, close date, volume, and a first-sight price snapshot embedded in `text`. Prices in the corpus are point-in-time at first ingestion — for live pre-trade pricing, follow the `url` to the venue and read the current quote there.

Daily snapshot of new research relevant to AI/ML/agents. Union of arXiv (cs.AI/cs.LG/cs.CL/cs.CR/cs.DC), HuggingFace daily papers (with upvote signal in title), and bioRxiv. Use when a user asks about a new technique, paper, or benchmark.

One retrieval, auto-picked substrate — the MCP twin of HTTP POST /v1/retrieve with substrate="auto". Runs a single post-cutoff retrieval, then returns whichever delivery substrate is cheapest AND legible for your `reader` model's token billing: - text — raw result pieces (Claude/GPT pixel billing, or any unknown reader). - glyph — a dense photo-glyph image (Gemini/Qwen flat-tile billing) you read with vision; the raw pieces ride along as a citation index. - answer — a pre-cited synthesized paragraph (weak tool-callers; needs a server LLM key). The trailing JSON block always carries a `selection` object {substrate, reader, reader_class, tier, rationale, estimates} so the choice is auditable from the honest token math — the same object the HTTP route returns. When the pick is glyph, the page image(s) precede that JSON block. Pricing matches /v1/retrieve: text/glyph bill the flat /query rate, answer bills the answer rate. The answer rate is charged up front and the delta is refunded when the pick resolves to text/glyph, so you always pay exactly the right rate.