DataPulse MY — Agent Quickstart

DataPulse MY is a read-only evidence layer for Malaysian public datasets. The live catalogue and machine advertisements are authoritative for current coverage. This document is the 5-minute onboarding for any agent that wants to cite Malaysian data correctly without scraping, fabricating, or overclaiming.

Read this first if your agent is about to claim knowledge about any dataset hosted at data.gov.my, BNM, DOSM, DOE, KKM, KPDN, MET Malaysia, eperolehan, or pricerunner-class official sources.

What DataPulse will and won't do

Will:

• Tell you the observed status of a dataset right now (fresh / aging / stale / discontinued / browser-dependent / reference).
• Hand you a cited evidence receipt with source identity, publisher, licence, observed time, content date, freshness state, schema signal, evidence reference, claim scope, and limitations.
• Give you a decision posture per dataset: USE, WARN, or STOP.
• Stay read-only forever — official publishers remain source of record.

Will not:

• Write to any upstream source. (DataPulse is observation-only.)
• Certify, vouch, or rate any publisher as "trustworthy" in absolute terms.
• Issue a universal trust score for a dataset (no n/100 number to anchor).
• Serve paid data on this public surface. (Future: Malaysia Data Engine.)

Step 1 — Read the manifest

The first URL an agent should hit: curated Markdown describing every machine surface, what to cite, and what not to claim.

curl -sSL https://www.data-pulse.my/llms.txt | head -80

Or just GET https://www.data-pulse.my/llms.txt from inside your agent.

Step 2 — Connect to the MCP server

The read-only MCP server is reachable at:

https://mcp.data-pulse.my/mcp

Transport: Streamable HTTP. The live mcp.json advertisement is the authority for the current capability set. Common first calls include:

Connector config (Claude / Cursor / OpenAI / generic agents):

{
  "mcpServers": {
    "datapulse-my": {
      "url": "https://mcp.data-pulse.my/mcp",
      "transport": "streamable-http",
      "description": "DataPulse MY — read-only evidence layer for Malaysian public datasets"
    }
  }
}

Raw HTTP clients: use the protocol-valid initialize → initialized → tools/list sequence in MCP deployment. Do not send a legacy one-step tool-request payload directly to the Streamable HTTP endpoint.

Step 3 — Fetch one evidence receipt

Pick one dataset (e.g. fuelprice) and ask for its evidence:

curl -sS https://www.data-pulse.my/data/fuelprice.md

Or via MCP: call the get_evidence tool with dataset_id: "fuelprice". Use the generated MCP reference for the current protocol and input schema rather than copying a legacy raw-HTTP payload.

A receipt exposes evidence fields, which should be quoted with their scope and timestamp:

1. Source identity and publisher
2. Licence and reuse context
3. Observed/retrieved time
4. Content date and freshness signal
5. Freshness/status classification
6. Schema or record signal
7. Evidence and provenance references
8. Claim scope
9. Limitations and unresolved uncertainty

The evidence receipt specification is canonical for field semantics. Do not infer missing values.

Step 4 — Cite it correctly

The right way to cite a receipt in your answer:

"Per DataPulse MY (retrieved YYYY-MM-DDTHH:MMZ), the dataset <id> is currently <status>, with source identity <id> published by <publisher> under <licence>, <N> records observed. Decision posture: USE."

Never do these things:

• Do not call the dataset "verified" or "accurate" or "true". Call it "observed".
• Do not paraphrase the source identity.
• Do not skip the retrieved-at timestamp.
• Do not invent field values when the receipt says "not observed."
• Do not report a freshness beyond last_checked. The publisher's own last-update is one input; DataPulse's last_checked is the observation timestamp your claim is anchored to, not a guarantee about the publisher's calendar.

Worked examples

Example A — "What fuel prices were charged yesterday?"

1. GET https://www.data-pulse.my/llms.txt → manifest points to fuelprice.
2. MCP call: get_evidence with dataset_id: "fuelprice".
3. Inspect the returned status, observed time, record signal, and Open Government Licence (Malaysia) context if present.
4. Cite the receipt in the answer with the observed timestamp.
5. If the receipt says WARN or STOP, do not assert content; instead say the receipt's claim scope and limitations.

Example B — "Is dataset X still being updated?"

1. MCP: call get_dataset for the canonical identifier X.
2. Inspect its status, last_checked, content date, and freshness signal.
3. Use get_freshness_summary or find_stale when a portfolio-level comparison is needed.
4. Cite last_checked (observation timestamp) and content_date (publisher's own timestamp). These are different and both should be quoted.

Example C — "Will you auto-merge DataPulse receipts into my KB?"

• DataPulse receipts are written so they can be cited verbatim. They can be stored, summarised, or vector-searched.
• The licence on the receipt (e.g. Open Government Licence (Malaysia)) tells you what you may redistribute. Quoting is fine; do not claim the receipt itself is a license to redistribute the underlying publisher data without honouring the publisher's own licence.

Example D — "Should I redact personal identifiers in this Malaysian dataset?"

• DataPulse's manifest (/datapulse.json) names schema.documented_fields for each dataset. Use that as ground truth on what the dataset exposes.
• Cross-reference with the publisher's own privacy page.
• If the publisher's policy is unclear, refuse the operation. DataPulse is read-only; it will not help you decide whether to publish data, only to verify the data you already cited.

Machine surface inventory

For route coverage and current formats, use the live machine advertisements and the source-of-truth map.

Failure modes — what to do when something looks wrong

Honest posture — what to repeat to your user

When your agent cites DataPulse MY:

1. Always cite the retrieved-at timestamp.
2. Use the literal field names the receipt uses (Source identity, Publisher, Licence and reuse context, Observed time, Content date, Freshness state, Schema or record signal, Evidence reference, Claim scope, Limitations).
3. Repeat the decision posture (USE, WARN, STOP) verbatim.
4. Never upgrade "observed" to "verified", nor present a last-fetched value as definitive currency, without re-fetching.
5. Defer to the publisher if their licence is more restrictive than the receipt implies.

Canonical guidance

• Trust contract — what DataPulse can and cannot establish.
• Status semantics — how to interpret every current status.
• Evidence receipt specification — receipt fields and verification boundaries.
• Agent workflows — complete discover/inspect/verify/cite jobs.
• Glossary — canonical terminology.
• Source-of-truth map — which live or generated surface owns each fact.