You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Part of epic #1065. Companion PR in nemarOrg/docs adding an MCP section to the For-agents guide with client configurations for Claude Desktop, Cursor, and OSA. Companion PR in OpenScience-Collective/osa adding the mcp_servers URL to the NEMAR assistant config. Verify on mcp-test.nemar.org, then enable the production route, then check off the acceptance criteria in #1065.
Definition of done: mcp.nemar.org reachable over Streamable HTTP, listed on docs.nemar.org, and consumed by the OSA NEMAR assistant on staging.
Companion deliverable on the website side: nemarOrg/website#319 asks for /.well-known/mcp/server-card.json on nemar.org (SEP-1649) pointing at https://mcp.nemar.org/mcp, generated from the shared contract. Check it off from this phase's PR.
Phase 5 status: three of four parts landed, one in review.
In-repo (nemar-cli) -- done. ADR 0050 and the Python verification script
merged in #1330, plus the epic review's fixes and a CI lint gate for scripts/mcp/ in #1335. Released as v0.10.1, which provisioned mcp.nemar.org (a custom_domain route, so the hostname did not exist before the
deploy).
Verification -- done, and it reordered the phase. Staging came up first and
every transport-level check passed against mcp-test.nemar.org: 2026-07-28
negotiated from the official Python SDK, all six tools with their inputSchema,
the tools/list cache hint, GET/DELETE 405, OPTIONS 204, a missing Mcp-Method 400, a disallowed Origin 403, describe_dataset end to end, and
the D1 gate refusing a real private dataset (nm000103) with the same wording an
unknown id gets.
But zarr-test.nemar.org's catalog is count: 0 -- the dev environment has no
converted dataset at all -- so the four recording-level tools had nothing to
read there, and the docs examples this phase was supposed to copy from staging did
not exist. Production has 625. Owner decision 2026-09-09: release first, then
document from production. All six tools now verified against mcp.nemar.org.
One correction to this issue's own premise, and it was in my favour: this issue
said to add an MCP section to the For-agents guide, and an earlier note of mine
claimed no such guide existed and a new page was needed. The issue was right and
the note was wrong -- platform/for-agents.md has been published since
2026-09-03, and I had read a stale local checkout instead of origin/main. It is
a section on that page, as originally specified. Every request and response in it
is copied from a real session against mcp.nemar.org.
That PR also retargets four ADR 0025 links to ADR 0049, which supersedes it,
and drops the "any future NEMAR tool server returns a recipe" framing from cost-ladder.md and index-contract.md: that server now exists, so those pages
point at it.
Worth stating plainly because it changes what this part was: OSA's two NEMAR tools
were already dead. Both called nemar.org/api/dataexplorer/datapipeline/..., which returns 404 since the legacy
dataexplorer site was retired. So this replaces them rather than adding to them,
and OSA had no MCP runtime at all -- mcp_servers existed in the config model with
a "(Phase 2)" docstring and nothing read it.
The one real design constraint: CommunityAssistant.__init__ is synchronous and
FastAPI calls it from inside a running event loop, where asyncio.run raises. So
discovery runs in a worker thread with its own loop. Mutation-checked: with a bare asyncio.run the running-loop test fails while the plain synchronous one passes.
Tests stand up a real MCPServer over Streamable HTTP -- no mocks -- and a network-marked tier passes against mcp.nemar.org.
Keeping this open until #352 merges. Epic #1065 is closed: all eight of its
acceptance criteria are verified against production, with two deliberate deviations
called out in its closeout comment.
Phase 5 complete. Definition of done, item by item:
mcp.nemar.org reachable over Streamable HTTP. Released in v0.10.1. Verified against production with scripts/mcp/verify_streamable_http.py --url https://mcp.nemar.org/mcp (the official Python mcp 2.2.0 SDK): initialize, all six tools listed, one real call per tool, the provenance envelope present on each, and the over-cap read_window refusal returning a typed refusal rather than a truncated window. The multi-chunk render_overview path that phase 4 fixed was confirmed on real data: width_px: 800 selects level 3 and reads 3 chunks, returning an 800x1259 PNG.
Listed on docs.nemar.org. nemarOrg/docs#27, merged and live: a ## Tool calling: mcp.nemar.org section on the For-agents guide, with the transport, the six tools, a real session trimmed from the production verification run, the recipe-versus-taste rule, the provenance envelope, the limits, and client configuration for Claude Desktop, Cursor and OSA. Every example is copied from a real session, none from the design doc.
Consumed by the OSA NEMAR assistant.OpenScience-Collective/osa#352, merged: src/tools/mcp_client.py is a real MCP runtime (the repo had none), and the NEMAR assistant now gets its dataset tools from https://mcp.nemar.org/mcp instead of the two hand-written tools that called the retired dataexplorer endpoint and had been returning 404. Two follow-ups also merged: osa#354 pointed the assistant's documentation URL at nemar.org/discover (read from the website's own src/lib/host.ts, not guessed), and osa#355 fixed the documentation-URL check to send a named User-Agent, which had been misreporting a working third-party page as a 403.
Proof from merged develop, not from the branch: the network-marked live tests pass against production (the six tool names are exactly the documented set, a search returns real datasets, a refusal arrives as readable text), and the documentation-URL check passes for all 16 communities.
Two things this phase produced that were not in its plan:
ADR 0050, no WASM in the Worker bundle, with both measured cases and the consequence that a compatibility_date bump silently removes the guard this repo currently relies on.
Activity
neuromechanist commentedon Sep 8, 2026
Companion deliverable on the website side: nemarOrg/website#319 asks for /.well-known/mcp/server-card.json on nemar.org (SEP-1649) pointing at https://mcp.nemar.org/mcp, generated from the shared contract. Check it off from this phase's PR.
fix: correct ADR 0050's platform claim and the docs target
Phase 5: ADR 0050 and a Python MCP client verification (#1330)
neuromechanist commentedon Sep 9, 2026
Phase 5 status: three of four parts landed, one in review.
In-repo (nemar-cli) -- done. ADR 0050 and the Python verification script
merged in #1330, plus the epic review's fixes and a CI lint gate for
scripts/mcp/in #1335. Released as v0.10.1, which provisionedmcp.nemar.org(acustom_domainroute, so the hostname did not exist before thedeploy).
Verification -- done, and it reordered the phase. Staging came up first and
every transport-level check passed against
mcp-test.nemar.org:2026-07-28negotiated from the official Python SDK, all six tools with their
inputSchema,the
tools/listcache hint,GET/DELETE405,OPTIONS204, a missingMcp-Method400, a disallowedOrigin403,describe_datasetend to end, andthe D1 gate refusing a real private dataset (
nm000103) with the same wording anunknown id gets.
But
zarr-test.nemar.org's catalog iscount: 0-- the dev environment has noconverted dataset at all -- so the four recording-level tools had nothing to
read there, and the docs examples this phase was supposed to copy from staging did
not exist. Production has 625. Owner decision 2026-09-09: release first, then
document from production. All six tools now verified against
mcp.nemar.org.Docs -- done, nemarOrg/docs#27, live at
docs.nemar.org/platform/for-agents.
One correction to this issue's own premise, and it was in my favour: this issue
said to add an MCP section to the For-agents guide, and an earlier note of mine
claimed no such guide existed and a new page was needed. The issue was right and
the note was wrong --
platform/for-agents.mdhas been published since2026-09-03, and I had read a stale local checkout instead of
origin/main. It isa section on that page, as originally specified. Every request and response in it
is copied from a real session against
mcp.nemar.org.That PR also retargets four
ADR 0025links toADR 0049, which supersedes it,and drops the "any future NEMAR tool server returns a recipe" framing from
cost-ladder.mdandindex-contract.md: that server now exists, so those pagespoint at it.
OSA -- in review, OpenScience-Collective/osa#352.
Worth stating plainly because it changes what this part was: OSA's two NEMAR tools
were already dead. Both called
nemar.org/api/dataexplorer/datapipeline/..., which returns 404 since the legacydataexplorer site was retired. So this replaces them rather than adding to them,
and OSA had no MCP runtime at all --
mcp_serversexisted in the config model witha "(Phase 2)" docstring and nothing read it.
The one real design constraint:
CommunityAssistant.__init__is synchronous andFastAPI calls it from inside a running event loop, where
asyncio.runraises. Sodiscovery runs in a worker thread with its own loop. Mutation-checked: with a bare
asyncio.runthe running-loop test fails while the plain synchronous one passes.Tests stand up a real
MCPServerover Streamable HTTP -- no mocks -- and anetwork-marked tier passes againstmcp.nemar.org.Keeping this open until #352 merges. Epic #1065 is closed: all eight of its
acceptance criteria are verified against production, with two deliberate deviations
called out in its closeout comment.
neuromechanist commentedon Sep 10, 2026
Phase 5 complete. Definition of done, item by item:
mcp.nemar.orgreachable over Streamable HTTP. Released in v0.10.1. Verified against production withscripts/mcp/verify_streamable_http.py --url https://mcp.nemar.org/mcp(the official Pythonmcp2.2.0 SDK):initialize, all six tools listed, one real call per tool, the provenance envelope present on each, and the over-capread_windowrefusal returning a typed refusal rather than a truncated window. The multi-chunkrender_overviewpath that phase 4 fixed was confirmed on real data:width_px: 800selects level 3 and reads 3 chunks, returning an 800x1259 PNG.Listed on docs.nemar.org. nemarOrg/docs#27, merged and live: a
## Tool calling: mcp.nemar.orgsection on the For-agents guide, with the transport, the six tools, a real session trimmed from the production verification run, the recipe-versus-taste rule, the provenance envelope, the limits, and client configuration for Claude Desktop, Cursor and OSA. Every example is copied from a real session, none from the design doc.Consumed by the OSA NEMAR assistant. OpenScience-Collective/osa#352, merged:
src/tools/mcp_client.pyis a real MCP runtime (the repo had none), and the NEMAR assistant now gets its dataset tools fromhttps://mcp.nemar.org/mcpinstead of the two hand-written tools that called the retired dataexplorer endpoint and had been returning 404. Two follow-ups also merged: osa#354 pointed the assistant's documentation URL atnemar.org/discover(read from the website's ownsrc/lib/host.ts, not guessed), and osa#355 fixed the documentation-URL check to send a named User-Agent, which had been misreporting a working third-party page as a 403.Proof from merged
develop, not from the branch: the network-marked live tests pass against production (the six tool names are exactly the documented set, a search returns real datasets, a refusal arrives as readable text), and the documentation-URL check passes for all 16 communities.Two things this phase produced that were not in its plan:
compatibility_datebump silently removes the guard this repo currently relies on.Epic #1065 is closed with all eight acceptance criteria checked and the two deviations recorded there.