PR #7 now combines Agent Skills format compatibility with two no-cost workflow demonstrations and a revised landing page. It does not install or execute the upstream Genesis catalog.
doe-mcp demo dataset-card: two geothermal catalog records, draft
metadata and an explicit owner-review list. No assertion of Genesis or
Croissant schema validity, file integrity, reuse rights or AI suitability.doe-mcp demo microbiome: a shale-study search followed by five samples
from a returned study ID. Sequence files are not retrieved.The default is recorded replay. Both new commands also completed with
--live on October 20 without credentials or a model. A public metadata
record does not grant access to the underlying dataset. Separate AI
clients may charge for inference.
The DOE release and PNNL announcement establish the October 15 public release. The existing research archive and upstream source review informed the integration choice. Indexed LinkedIn announcements from PNNL and PPPL were found; no individual post permalink was verified. Targeted X searches yielded no relevant verifiable posts. Social content was not used to establish technical capability.
Facility operations, data movement and metered model services are excluded. The facility-status demo is deferred: its existing adapter combines status and planned-outage fetches under the status fetch metadata. That needs a separate provenance review before presenting a combined status demo.
Initial code review read PR #7 at 69d332f, including the shared skill
reader, its consumers, conformance tests and record truncation changes.
No new blocking finding in that implementation. The final follow-up review
checked data dependencies, empty-result refusal, publisher-text escaping,
source references, cost claims and generated output. It corrected the
microbiome summary so a missing total is shown as unknown.
The inherited CI liveness failure is an OSTI timeout also reported on main. Replay tests do not establish future publisher uptime or model answer accuracy. No paid service was used and the PR is left unmerged.
Added rulemaking and outage-history, bringing the walkthrough count to
twelve. The rulemaking demo follows a returned document number and keeps
actions, dates, docket IDs and FERC exclusions visible. The outage finder
separates releases, documentation and denominator metadata; it calculates
no outage rate. Both support recorded replay and completed live keyless
reads during verification. The landing page now highlights six workflows.
All 717 tests pass, including empty-result refusal, rulemaking dependency, unknown-date preservation, denominator-vintage visibility and safe links. Ruff and registry validation pass. Schema and site regenerated. Tablet, phone and desktop browser checks cover the new picker entries and links. No new dependency, paid service or upstream code was introduced.
The latest Claude parent-session notes and PR review threads were checked. The catalog-kind count and microbiome follow-up findings were already fixed. The deployed guide and reference HTML pages both returned HTTP 200; the static local preview does not run the GitHub Pages Markdown renderer.
VIPS search returned HTTP 401, “Valid API key required”. Its manifest now records API-key access and proposed status with a named blocker; live tool calls refuse before fetching. The site no longer advertises VIPS as an active source. September adapter recordings remain covered by explicit historical replay tests, with no change to the default runtime gate. No credentials, paid services or new dependencies were added.
PR #6 added outage releases, NERSC status and microbiome records. PR #7 added recorded walks across those sources and the research catalogs. PR #8 put the walks on the landing page and marked the VIPS key requirement in the reference. A reader could already trace a record. Preparing a citation list or counting missing measurements still meant working through the JSON.
| Server | What its maintainers document | Choice for DOE-MCP |
|---|---|---|
| US Census Bureau | Its geography tool resolves names to candidate FIPS codes. Aggregate queries specify a dataset and vintage. | Say exactly what a place filter selected. The NMDC query matches text containing “Washington”; it cannot establish a state boundary. |
| NASA Earthdata | Separate tools find collections, check for granules and look up citations. | Give readers citation fields they can take away. A catalog hit by itself says nothing about whether anyone read the paper. |
| cyanheads EIA server (community project) | Its query results carry units and match totals. Its guide distinguishes null from zero. | Keep numeric zero in the soil worksheet. Leave unreported values blank, name the unit and count missing fields only on the returned page. |
This comparison covers documented workflows. It does not measure server uptime, response times or the accuracy of a model using them.
The bibliography worksheet applies the Genesis literature-search skill’s bibliography step to DOE PAGES. The soil sample field check applies the data-exploration skill’s missing-value check to NMDC. Neither runs the upstream scripts. Each worksheet row points to an evidence ID in the downloadable JSON. Both examples run from recordings without a model or key.