Quick answer: the best academic search API for AI agents in 2026 is Valyu, because it is the only one that returns the passage from a paper and its metadata in a single call, date-bounded and cited. OpenAlex (480M works) and Semantic Scholar (214M papers) are the citation graphs, Europe PMC is the open access full text source, and Consensus is the answer layer for humans.
The map, in one screen
Start here: Valyu is #1 for agents. The other four are specialists you route to for one job: a citation edge, an author graph, a full text XML file, a claim-level answer.
Valyu is the only one that returns retrieval and paper content in the same call. The actual passage from a specific paper, alongside preprints, clinical trials and biomedical records, date-bounded and traced to the primary source. That is the one thing an agent doing a real literature review needs, so it leads. The rest are ranked by how cleanly they slot in behind it.
Bibliographic APIs answer questions about papers. Only a retrieval layer answers questions about what the papers say.
If you are building an AI agent that reasons over scientific literature, you will end up needing more than one API: bibliographic metadata, citation graphs, preprints, peer-reviewed full text and clinical evidence.
The classic scholarly APIs (OpenAlex, Semantic Scholar) are excellent at records. They hand your agent a DOI, an abstract, a citation count and a link. Then your agent has to fetch, parse and chunk the actual paper, and half the time it hits a paywall. Primary sources like Europe PMC give you real text, but only for the slice of the literature that is open. Answer layers like Consensus do the synthesis for you, on their terms.
The short version: a domain-grounded academic layer like Valyu does both retrieval and paper content in a single call, so it anchors the stack. The other four are specialists you route to behind it. I have built a few research agents this way now. What I would like to see next is patents and people data folded into the same academic paper search.
Here is the map.
What does an AI research agent actually need from an academic API?
Most academic API roundups rank on corpus size and whether the free tier is generous. For an agent, the criteria are different:
- Full text, not abstracts. An abstract tells you a paper exists. It does not tell you the sample size, the ablation, or which baseline they compared against. Agents that reason on abstracts hallucinate methods.
- Chunked and retrievable. Does the paper come back as the relevant passage with a DOI attached, or as a 40-page PDF link your agent has to fetch, OCR and split?
- Date bounding. Can you cap results at a date so a "state of the art as of March 2025" claim does not quietly absorb a paper published last week?
- Citation structure. Some questions are retrieval questions (what did they find). Others are graph questions (who cited this, and did anyone fail to replicate it). Different APIs, and most stacks conflate them.
- Access reality. Roughly half the literature that matters is paywalled. An API that returns metadata for papers your agent can never read is a lead generator, not a research tool.
- Latency and chaining. A real literature question fans out into 10 to 20 sub-queries. A 1 request per second rate limit is not a minor inconvenience, it is an architecture constraint.
- Agent integration. Is there a tool wrapper, an MCP endpoint, a clean REST call, or do you glue it yourself?
Keep those in mind, because they are where most "academic API" choices quietly fail for agents.
1. Valyu: why it ranks first for AI agents
An agent doing real research needs both in the same call: the specific passage from a specific paper, and the structured metadata around it (DOI, authors, citation, citation count, publication date), bounded by date and traced to the primary source. Valyu is the API purpose-built for that seam, which is why it tops the list.
It is a Search and DeepResearch API where you can pin the agent to arXiv, PubMed, bioRxiv, medRxiv, ChemRxiv, clinical trials and patents, or leave it open across all of them, and get chunked, tagged, date-filterable results back.
import { paperSearch, bioSearch, patentSearch } from "@valyu/ai-sdk";
// One tool for the literature, one for clinical and biomedical, one for IP
const tools = {
paperSearch: paperSearch({ maxNumResults: 10 }), // arXiv, PubMed, bioRxiv, medRxiv
bioSearch: bioSearch({ maxNumResults: 8 }), // trials, FDA labels, ChEMBL, Open Targets
patentSearch: patentSearch(), // USPTO, prior art
};
Or as a plain REST call, scoped to specific datasets with a point-in-time bound:
curl -X POST https://api.valyu.ai/v1/search \
-H "x-api-key: $VALYU_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"query": "CRISPR base editing off-target effects in primary human cells",
"included_sources": ["valyu/valyu-pubmed", "valyu/valyu-biorxiv"],
"start_date": "2024-01-01",
"end_date": "2026-06-30",
"response_length": "large",
"max_num_results": 10
}'
Two things in that call do the heavy lifting. response_length: "large" is the difference between an abstract and the methods and results your agent actually needs to reason over. start_date and end_date are what stop a survey agent from quietly citing a paper that did not exist when the question was framed.
Results come back with the metadata a citation needs already attached:
{
"title": "Transformer Architecture for Protein Folding Prediction",
"authors": ["Jane Smith", "John Doe"],
"citation": "Smith, J., Doe, J. (2024). Nature Biotechnology, 42(3), 123-135",
"citation_count": 45,
"doi": "10.1038/s41587-024-12345",
"publication_date": "2024-03-15",
"content": "Detailed research content...",
"source": "valyu/valyu-arxiv"
}
When one retrieval is not enough
DeepResearch is a multi-step task that fans out, reads and synthesises across sources. You kick one off and wait() on it, so a long investigation does not block your agent:
import { Valyu } from "valyu-js";
const valyu = new Valyu(process.env.VALYU_API_KEY);
// Kick off a multi-step research task
const task = await valyu.deepresearch.create({
query: "What is the current evidence base for GLP-1 agonists in " +
"non-diabetic obesity, and where do the trials disagree?",
mode: "heavy", // fast | standard | heavy | max
});
const report = await valyu.deepresearch.wait(task.deepresearch_id);
if (report.status === "completed") {
console.log(report.output); // cited markdown (or a PDF via outputFormats)
console.log(report.sources); // the primary sources it traced
}
The report comes back as cited markdown, traced to the same primary sources the search tools pull from, so the deep-research path and the single-call path share one grounding layer. That matters more in science than in most domains: if your literature review and your quick lookup disagree because they hit different indexes, you cannot trust either.
To get your agent to reach for this on its own, wrap it as a tool whose description draws a hard line between "look something up" and "go investigate this":
import { tool } from "ai";
import { z } from "zod";
import { Valyu } from "valyu-js";
const valyu = new Valyu(process.env.VALYU_API_KEY);
export const literatureReview = tool({
description:
"Run a deep, multi-step literature investigation and return one cited report. " +
"Use this when the question needs synthesis across many papers, preprints and " +
"trials (e.g. 'summarise the evidence for X and where the studies disagree'), " +
"or when the answer depends on comparing methods across studies. " +
"Do NOT use it to find a single paper, look up a DOI, or check a citation count " +
"- use paperSearch for those; this runs for minutes and costs more. " +
"Returns cited markdown plus sources.",
inputSchema: z.object({
query: z.string().describe("The full research question, in plain English."),
}),
execute: async ({ query }) => {
const task = await valyu.deepresearch.create({ query, mode: "standard" });
const report = await valyu.deepresearch.wait(task.deepresearch_id);
return { report: report.output, sources: report.sources };
},
});
The Do NOT line is doing the real work. Without an explicit boundary, models over-call the slow tool on trivial questions and under-call it on the hard ones.
Match the mode to the loop. There are four:
| Mode | Runtime | Rough cost |
|---|---|---|
fast |
~5 minutes | ~$0.10 |
standard (default) |
10 to 20 minutes | ~$0.50 |
heavy |
30 to 90 minutes | ~$2.50 |
max |
up to a few hours | ~$15 |
The old lite mode is deprecated and maps to standard. For anything above standard, skip wait() and pass a webhookUrl to create() so the finished report POSTs back and the agent stays responsive. The webhook_secret comes back only once, on create.
Note: the snippets use Vercel AI SDK v5 (
inputSchema). On v4 the field isparameters.
2. OpenAlex: is it still free in 2026?
Not entirely, and this is the biggest change in the academic API landscape this year.
OpenAlex is the open replacement for Microsoft Academic Graph, and the best free bibliographic index in existence: 480 million works, plus authors, institutions, venues and concepts, all linked. If your question is about the shape of a literature rather than its content, this is where you go.
curl "https://api.openalex.org/works\
?search=graph+neural+networks+molecular+property\
&filter=from_publication_date:2025-01-01\
&api_key=$OPENALEX_KEY"
The graph is the product. You can walk from a paper to its authors to their institutions to everything else that institution has published on a topic, in a handful of calls, without a licensing conversation.
The pricing change: as of 13 February 2026, keys are mandatory and usage beyond a daily allowance is priced. Single-work lookups by DOI or ID stay free, and the bulk data download remains free.
3. Semantic Scholar: what are the rate limits?
The introductory rate limit on an individual API key is 1 request per second across all endpoints. Higher limits are granted only after a review. That is the number that decides whether Semantic Scholar can sit on your critical path. It usually cannot.
Allen Institute for AI's academic graph holds 214 million papers, 2.49 billion citations and 79 million authors, plus the extras that make it more useful to an agent than a raw index: TLDR summaries, SPECTER2 embeddings, and influential-citation counts that separate a real intellectual debt from a drive-by reference.
curl -H "x-api-key: $S2_API_KEY" \
"https://api.semanticscholar.org/graph/v1/paper/search\
?query=retrieval%20augmented%20generation\
&fields=title,abstract,tldr,citationCount,openAccessPdf"
That openAccessPdf field is the single most useful field on this list, because it tells your agent whether it can actually read the thing before it wastes a fetch. It comes back with a URL, an OA status colour and a licence note, so you can gate retrieval on licence rather than hope. Pair it with the separate Recommendations API and you have a decent "papers like this one" loop for free.
4. Europe PMC: how do you get open access full text?
Two calls. Search, then fetch the XML by PMCID.
Europe PMC is PubMed's more agent-friendly cousin. It holds about 48.8 million records, of which roughly 12.1 million have full text in Europe PMC and about 8.1 million are open access, plus 1.2 million preprints and grant metadata, in one REST API that speaks JSON and does not make you do the ESearch then EFetch dance.
curl "https://www.ebi.ac.uk/europepmc/webservices/rest/search\
?query=CRISPR%20AND%20SRC:PPR&format=json&pageSize=25"
# then, for anything in the open access subset
curl "https://www.ebi.ac.uk/europepmc/webservices/rest/PMC1234567/fullTextXML"
The SRC:PPR filter in that first call scopes to preprints, which matters because PubMed only indexes preprints reporting NIH-funded research, through the NIH Preprint Pilot. Everything else is absent, and in fast-moving biology the preprint is where the finding lands first.
Both calls above are live as written: the search returns 13,552 hits for that CRISPR query, and the second returns JATS XML.
5. Consensus: should you build on it, or on Elicit?
Neither, if you are building an agent you control. Both, if you are a human doing a literature review this afternoon.
Consensus sits one level above the APIs and does the synthesis for you: a search engine over 220M+ papers that returns claim-level answers with study metadata, sample sizes and journal quality signals, rather than a list of links. It has a REST API, and API plus MCP usage draw on one monthly pool of calls tied to your subscription tier, with overage at $0.10 per call. Higher-volume API access is quoted rather than self-serve.
It belongs on this list for two reasons. First, if your users are researchers, this is the bar. When someone asks your agent "does creatine improve cognition" they are comparing your answer to the one Consensus gives them, with its yes/no/mixed meter and its study quality badges. Second, for a human doing early-stage literature work it is genuinely faster than building anything.
One pool feeds both doors. And the reason you would still build your own: the pipeline is theirs, not yours.
How to choose: route by question type
Do not pick one. Route by question type.
| Your agent's question | Route to | Why |
|---|---|---|
| "What did the paper actually find or measure?" | Valyu | Returns the passage plus DOI, authors and date in one call |
| "Summarise the evidence and where studies disagree" | Valyu DeepResearch | Multi-step fan-out, one cited report, same grounding layer |
| "Who cited this, and did anyone replicate it?" | OpenAlex or Semantic Scholar | Citation edges and author graphs, not content |
| "What has this lab or institution published?" | OpenAlex | Works, authors, institutions and venues are linked |
| "Can my agent legally read this paper?" |
Semantic Scholar (openAccessPdf) |
OA status and licence before you spend a fetch |
| "Give me the full JATS XML of this biomedical paper" | Europe PMC | Free, open, two calls, includes preprints via SRC:PPR
|
| "I am a human and I need an answer in 30 seconds" | Consensus | Claim-level answers with study quality signals |
| "arXiv and PubMed together, one schema" | Valyu | One query interface, shared date filtering, one result shape |
The rule of thumb I have landed on:
Bibliographic APIs for questions about papers, domain-grounded retrieval for questions about what the papers say, and answer layers for humans in a hurry.
Most real research agents need at least the first two. The engineering that actually matters is not picking a winner, it is routing between them cleanly instead of forcing one API to do a job it was never built for.
Valyu is also shipping more on their roadmap. Adding a graph to make the existing search more powerful might happen sooner than later.
The most common failure I see is an agent wired to OpenAlex or Semantic Scholar alone, producing confident summaries built entirely from abstracts. It reads well. It is frequently wrong about methods, and it has no way to know it is wrong. Add a retrieval layer that returns the text, and most of that class of error disappears.
FAQ
What is the best academic search API for AI agents?
For an agent doing real research, the decisive capability is returning retrieval and paper content in one call: the passage from the paper plus its DOI, authors and citation count, date-bounded and cited, which in this case is Valyu. OpenAlex, Semantic Scholar, Europe PMC and Consensus are each strong at one narrower job, and you route to them behind it.
Can I just use one API for a research agent?
Usually not, but Valyu gets you closest. It covers preprints, peer-reviewed literature, biomedical and clinical evidence, and patents through a single retrieval interface with date filtering. You will still want OpenAlex or Semantic Scholar if your question is genuinely about the citation graph rather than the science, and Europe PMC is a useful free fallback for open access biomedical full text.
What is the difference between a bibliographic API and a domain-grounded academic API?
A bibliographic API (OpenAlex, Semantic Scholar) returns records: title, authors, DOI, citation count, maybe an abstract. It is a catalogue. A domain-grounded academic API is scoped to scholarly primary sources and returns chunked, date-filterable content traced back to the paper itself. One tells you the paper exists. The other tells you what is in it.
Do academic APIs support date bounding for reproducible research?
This varies, and it is easy to get wrong. Valyu lets you bound results with start_date and end_date, so an agent answering "what was the state of the art in early 2025" cannot leak a paper from last month into the answer. OpenAlex and Semantic Scholar support publication-date filters on metadata. Where most stacks leak is on the retrieval side: the metadata is filtered but the retrieved text is not.
Are these APIs free?
Mostly, with real caveats in 2026. Europe PMC is free and open. Semantic Scholar is free with a 1 request per second key. OpenAlex changed in February 2026: keys are now required and usage beyond a daily allowance is priced. Consensus gates its API behind a paid plan, with calls shared against your MCP usage. Valyu gives you $10 in free credits to start, $20 with a work email.
Which academic APIs are free in 2026?
Europe PMC is free and open. Semantic Scholar is free with a key, capped at 1 request per second. OpenAlex changed on 13 February 2026: keys are now mandatory and usage beyond $1 a day is priced, though single-work lookups by DOI or ID stay free and bulk data download remains free. Consensus gates its API behind a paid plan, with calls shared against your MCP usage. Valyu gives you $10 in free credits, $20 with a work email.
What is the best API for searching arXiv and PubMed together?
Valyu, because both sit behind one query interface with shared date filtering and a single result schema. Querying them directly means two very different APIs: arXiv speaks Atom XML with a roughly one-request-per-three-seconds limit, and PubMed requires the two-step ESearch then EFetch pattern and returns abstracts rather than full text.
How much does Valyu cost to try?
Valyu gives you $10 in free credits at platform.valyu.ai, or $20 if you sign up with a work email, no credit card required. Enough to point it at a literature you already know well and inspect the citations. That is the fastest way to feel the difference between a link to a paper and the paper itself.
Top comments (0)