# Talarion > Talarion maintains a knowledge base of verified, dated assertions. Each assertion is a Q/A pair, included in search results if and only if the querying LLM would otherwise answer the underlying question incorrectly. Talarion maintains a knowledge base of verified, dated assertions. Each assertion is structured as a Q/A pair and included in search results if and only if the querying LLM would otherwise answer the underlying question incorrectly. Each search query is mapped to a set of relevant assertions, each of which is rigorously timestamped and includes source URL(s) for provenance verification. Talarion is designed to power any LLM application which is world-state dependent. From global and local news to developments in financial markets, public policy and fundamental scientific research, Talarion scans the web and catalogues high-surprisal developments. There are two ways to use it: a REST API for developers building applications, and an MCP server for agents and assistants. ## Use Talarion - **For developers** building apps, products, and pipelines — the [REST API](https://docs.talarion.com): POST a search question to the search endpoint with a workspace API key and get back the assertions you would otherwise have answered incorrectly. This is the primary integration for products. - [Dashboard](https://dashboard.talarion.com): Create an account, manage API keys, and try the API playground. - **For anyone using an AI assistant like ChatGPT or Claude** — the [MCP server](https://docs.talarion.com/installation): add the Talarion MCP (`https://mcp.talarion.com/mcp`) in ChatGPT (listed as Acumen), Claude, Cursor, Claude Code, or any MCP client to give your assistant the `search` tool. One-click install links are on the homepage. ## API quickstart Everything needed to make a working call. Full reference: https://docs.talarion.com/search **Endpoint.** `POST https://api.talarion.com/v1/search` The previous name, `POST https://api.talarion.com/v1/research`, keeps working indefinitely and is identical except that its response envelope is `research_guidance` instead of `results`. Use `/v1/search` for new integrations. **Auth.** A workspace API key as a bearer token, created at https://dashboard.talarion.com → API Keys. Keys are workspace-scoped and prefixed `tlrn_`. ``` Authorization: Bearer tlrn_... Content-Type: application/json ``` **Or pay per call, with no account.** `/v1/search` and `/v1/fetch` speak the Machine Payments Protocol (https://mpp.dev). Send a request with no `Authorization` header and you get `402` with a `WWW-Authenticate: Payment` challenge instead of `401`; pay it on Tempo and retry with the credential in `Payment-Authorization`. Search is $0.002 per call; fetch quotes $0.05 for each document your specific request will return. Same request and response bodies either way. Discovery document: https://api.talarion.com/openapi.json **Request body.** ```json { "query": "How is insurance coverage of GLP-1 obesity drugs changing?", "model_id": "anthropic/claude-opus-4.8", "k": 20, "timeline": true } ``` - `query` (string, required) — the search question. One focused question about one topic retrieves best; make separate calls for distinct topics. - `model_id` (string, optional) — your model as an OpenRouter id. Filters results to facts that postdate your training cutoff. - `k` (integer, optional) — how many facts to return. Defaults to 20, and is **clamped** into 10–100 rather than rejected, so `"k": 3` succeeds and returns 10. - `timeline` (boolean, optional) — `true` (default) orders by most recent `true_by` first; `false` orders by relevance. - `true_by_min` (ISO 8601 datetime, optional) — only return facts with `true_by` on or after this. Overrides the `model_id`-derived training-cutoff floor. Naive datetimes are treated as UTC. - `true_by_max` (ISO 8601 datetime, optional) — only return facts with `true_by` on or before this, e.g. for date-boxed backtesting. Must be after `true_by_min`. **Response.** A `results` array. Each entry is one verified assertion: a `question`, its `answer`, `true_by` (ISO `YYYY-MM-DD`, or `null` when the source carries no truth-time), and `sources` (the citation URLs, as bare strings). A response is never empty; a question returns the nearest available facts. ```json { "results": [ { "question": "What flagship AI model did Anthropic release in May 2026?", "answer": "Claude Opus 4.8.", "true_by": "2026-05-15", "sources": ["https://www.anthropic.com/news/claude-opus-4-8"] } ] } ``` **Errors.** `401` with `{"detail": "Not authenticated"}` for a missing `Authorization` header, or `{"detail": "Invalid API key or missing workspace binding"}` for a bad key. `402` once the free-tier monthly call limit (1,000/month) is reached — upgrade at https://dashboard.talarion.com/billing. **Complete example.** ```bash curl -X POST https://api.talarion.com/v1/search \ -H "Authorization: Bearer $TALARION_API_KEY" \ -H "Content-Type: application/json" \ -d '{"query": "What are the most significant AI model releases in 2026?"}' ``` **MCP instead of REST.** Point any MCP client at `https://mcp.talarion.com/mcp` (OAuth, no API key needed), which exposes `search` with the same four parameters. For Claude Code: ```bash claude mcp add --scope user --transport http talarion https://mcp.talarion.com/mcp ``` **Going deeper** — every docs page is available as plain Markdown, and the whole site as one file: - https://docs.talarion.com/search.md — this reference as Markdown - https://docs.talarion.com/llms.txt — docs index - https://docs.talarion.com/llms-full.txt — every docs page concatenated - https://api.talarion.com/openapi.json — OpenAPI 3.1 spec, with MPP `x-payment-info` prices ## What search does A search query is mapped to the set of relevant assertions in the knowledge base — each a Q/A pair, rigorously timestamped, carrying the source URL(s) that back it. An assertion is returned if and only if the querying LLM would otherwise answer its underlying question incorrectly, so what comes back is what you were going to get wrong, not a ranked restatement of what you already know. The ideal input is one focused question about one topic — e.g. "How is insurance coverage of GLP-1 obesity drugs changing?" or "What's the state of Manhattan's commercial office real-estate market?". ## What it's good for Anything recency-dependent — news and ongoing stories, markets and prices, policy and regulation, model and product releases, versions and tooling, statuses that flip. If being out of date would cause an error, it's in scope. Coverage is deepest on news, finance, policy, and developer & agent tooling. Call it at the start of research on any such topic — and again after searching, to catch what was missed. ## What it's not good for Timeless questions — settled science and the natural world, history, definitions — where nothing has changed recently. Pinpoint lookups where you already know the exact fact you need are a job for traditional web search. ## Product - Pricing: a free tier, usage-based paid pricing per brief, and custom Enterprise plans — see the [dashboard](https://dashboard.talarion.com). - Efficacy: on PriorBench (150 research questions, 4,154 evaluation questions), 91% of research questions benefit more from Talarion than from web search, with +160% more information delivered and 3.5× higher information density per token — see [PriorBench](https://talarion.com/priorbench). ## More - [Mission](https://talarion.com/mission): Why Talarion exists — making AI less wrong by precomputing what models miss. - [Blog](https://talarion.com/blog): Our thoughts on information, forecasting, and giving AI the best possible access to world state. - [Contact](https://talarion.com/contact): For production-scale usage, custom data support, or bulk pricing.