Agent access · experimental

Read the corpus. Know what is live.

The public corpus is available through a live hosted MCP and ordinary machine-readable files. Reads are open. Writes create private, rate-limited review jobs and require an allowlisted token. This guide shows how to connect, prompt an agent, and interpret what comes back.

Connect to the MCP → Use the static corpus

Availability

What works right now

Live

Static read access

Corpus index, object JSON, claims graphs, RO-Crates, verification records, content negotiation, and FAIR Signposting.

Live

Hosted MCP

Streamable HTTP at https://datasetpapers.com/mcp, with ten open read tools and three trust-preserving prompts.

Gated

Write jobs

Propose, fork, verify, and job-status tools persist private jobs for allowlisted collaborators. The generation runner is not connected.

A successful write response means “recorded for review,” not “analysis completed” or “published.” No MCP call can put an object directly into the public corpus.

Quickstart

Give an agent a reliable path through the corpus

Fetch the corpus index

Start here to discover the available papers and their canonical slugs.
```
curl -L https://datasetpapers.com/papers/index.json
```
Choose a paper and fetch its research object

Replace {slug} with a slug from the index.
```
curl -L https://datasetpapers.com/papers/{slug}/object.json
```
Read the claims and evidence links

The claims graph provides atomic statements, typed relationships, and supporting evidence references.
```
curl -L https://datasetpapers.com/papers/{slug}/claims.json
```
Check independent re-execution

Do this before treating a published number as reliable. A claim's self-reported status is not independent verification.
```
curl -L https://datasetpapers.com/papers/{slug}/verification.json
```

Prompt an agent

A ready-to-use instruction

Paste this into an agent that can fetch public URLs. It keeps discovery and verification in the right order.

Use https://datasetpapers.com/papers/index.json to discover the available analyses.

For the paper you select:
1. Read object.json for the complete research object.
2. Read claims.json for atomic claims and evidence links.
3. Read verification.json before relying on any numerical result.
4. Treat engine-reported verification_status separately from independent re-execution.
5. Cite the paper URL and the specific claim identifier where possible.

This corpus is an experimental prototype, not a product or authoritative source. Report any missing evidence or uncertainty.

Machine surfaces

What each URL gives you

Surface	URL	Use it for
Corpus index	`/papers/index.json`	Discover papers, slugs, identifiers, and top-level metadata.
Research object	`/papers/{slug}/object.json`	Read identifiers, source data, components, claims, gate results, verification, and exports.
Claims graph	`/papers/{slug}/claims.json`	Work with atomic claims and typed evidence, derivation, and containment relationships.
RO-Crate	`/papers/{slug}/ro-crate-metadata.json`	Traverse the packaged research object as JSON-LD.
Verification	`/papers/{slug}/verification.json`	Inspect the independent re-execution environment, inputs, comparisons, and result.
Discovery	`/.well-known/mcp.json`	Read the authoritative MCP availability status and machine-surface links.

HTTP discovery

Content negotiation and Signposting

An agent can request a machine representation from the human paper URL instead of constructing a file path.

Request the object JSON

curl -L \
  -H "Accept: application/json" \
  https://datasetpapers.com/papers/{slug}/

Request the RO-Crate

curl -L \
  -H "Accept: application/ld+json" \
  https://datasetpapers.com/papers/{slug}/

Paper responses also advertise FAIR Signposting Link headers for canonical identifiers, metadata, licence, authors, and downloadable components.

Trust model

How an agent should interpret a result

`verification_status`

A status reported with a claim or version. Do not assume this alone means the result was independently reproduced.

`verification.json`

The independent re-execution record. A matched result means the published analysis code reproduced its declared values from pinned inputs.

`trust_distance`

The number of provenance hops from raw data. More hops mean more transformations to inspect, not automatically lower quality.

Experimental status

Formats, identifiers, and fields may change. Treat the corpus as evidence to inspect, not as an authoritative source or decision system.

Hosted MCP · live

Connect an MCP client

Add the endpoint below as a remote or Streamable HTTP MCP server. No account, key, or custom header is needed for read tools.

https://datasetpapers.com/mcp

Generic client configuration

Clients name this setting differently, but the server entry is the same:

{
  "mcpServers": {
    "datasetpapers": {
      "type": "http",
      "url": "https://datasetpapers.com/mcp"
    }
  }
}

Check it without an MCP client

Service health

curl -L \
  "https://datasetpapers.com/mcp?health=1"

Authoritative discovery

curl -L \
  https://datasetpapers.com/.well-known/mcp.json

How to interact after connecting

Ask in ordinary language. Your client discovers the tools and chooses when to call them. Useful starting instructions:

Use the datasetpapers MCP to search for analyses about metabolic completeness.

For the best match:
1. Fetch the datasetpaper with claims, figures, provenance, and attribution.
2. Keep verification_status and trust_distance attached to every claim.
3. Separate confirmed, re-executed, unverified, null, and inconclusive results.
4. Cite the version ARK and individual claim ARKs.
5. Remember this corpus is an experiment, not an authoritative source.

Other examples: “Get claim ark:/…/claim-1 and critique its evidence,” “Find other datasetpapers built on this source dataset,” or “Use the design_fork prompt, but do not submit a write job.”

What the read tools do

Discovery and retrieval: search_datasetpapers, get_datasetpaper, get_claim, list_claims, and get_dataset. Graph and packaging: get_provenance, get_rocrate, get_croissant, find_related, and query_graph. Reads are restricted by database row-level security to published objects.

Check MCP health Read discovery JSON Read the full specification

Authenticated writes · limited experiment

Submitting a proposal, fork, or verification

Write access is allowlisted and tokens are issued to existing experiment collaborators out of band. It is not a public sign-up flow. Add the token only to a client you trust:

{
  "mcpServers": {
    "datasetpapers": {
      "type": "http",
      "url": "https://datasetpapers.com/mcp",
      "headers": {
        "Authorization": "Bearer ${DATASETPAPERS_TOKEN}"
      }
    }
  }
}

`propose_analysis`

Checks that the pinned source is open and has a recognised derivative-friendly licence, then queues the question.

`fork_datasetpaper`

Records the exact parent version and proposed question. It does not mint a fork or run code yet.

`submit_verification`

Queues evidence for review. It does not immediately change a public claim’s verification status.

`get_job`

Returns only jobs owned by the same credential, with their audit events and honest runner or review state.

Each credential is limited to 25 submissions per 24 hours and five active jobs. Supply an idempotency_key when retrying so a network failure cannot create a duplicate. Generation and fork jobs remain queued until the separate sandboxed executor is connected.