Skip to content

Use with Claude Desktop and claude.ai

This page explains how to get query results, full evidence and workflow instructions in local Claude Desktop and through a claude.ai custom connector, and how to check that a setup actually works. Set up the connection first with Connect an MCP client.

What has and has not been checked

Be clear about which claims rest on which evidence:

Surface What was checked Status
Server version 0.4.0, over local STDIO and local Streamable HTTP Automated offline tests and smoke checks: tool text for all five progression tools, read_evidence, get_workflow_instructions, native prompts and resources, paging, size limits, progression details in framework listings, citation links in component text, and one request per link type in workflows Passed locally, 2026-10-07. This tests the server, not a client.
MCPB bundle, version 0.4.0 STDIO smoke test run from the unpacked bundle Passed locally, 2026-10-07; not an installation in Desktop
MCPB installation in Claude Desktop, version 0.4.0 Install and use on macOS; removal and same-version replacement Install and use observed by a user on macOS, 2026-10-08. Removal and same-version replacement not checked on any OS; see the installation guide
Claude Desktop, earlier development build Observed by a user on 2026-10-03: the connector started and listed six frameworks; all nine prompts appeared under Add from curriculum-knowledge-graph, and the support-planning prompt rendered as an attachment; the Catalog resource attached; the resource menu offered only Catalog and could not find a progression link. User observations, not a test of this version
Claude Desktop, version 0.4.0, on macOS Observed by a user on 2026-10-08 through both the checkout configuration (claude_desktop_config.json) and the .mcpb bundle: the connector started, frameworks were listed, and prompts and resources appeared. Steps 1–13 of the walkthrough below, including 9a, were run, and the user checked the composed, cited answers in steps 11–13. The resource menu listed more than Catalog. User observation, 2026-10-08; no failures observed. Desktop app version not recorded.
claude.ai custom connector, version 0.4.0 Observed by a user on 2026-10-08 against the public deployment at https://kg-for-ed-global-mcp.up.railway.app/mcp, which was confirmed to run 0.4.0 on that date. Same results as the Desktop row above, including walkthrough steps 1–13 and their composed answers. The server's prompts showed in claude.ai, and server resources could be attached. User observation, 2026-10-08; no failures observed. Plan type and connector name not recorded.

The local setup guide is written for macOS, and the 2026-10-08 Desktop observations were made on macOS. The 2026-10-03 observations and 2026-10-07 smoke evidence above do not record an OS, so they establish no OS-specific Desktop pass. Windows and Linux Desktop setup is undocumented and untested here; the manifest's three declared platforms do not change that.

On 2026-10-08 a user ran the walkthrough workflows, including the cited, composed answers in steps 11–13, in Claude Desktop and in claude.ai, and checked the answers. These are a user's observations, not automated tests. Rendering a prompt, reading a resource or passing a smoke test is still not the same as a completed workflow.

Supported routes

You need Claude Desktop (local) claude.ai custom connector
Progression query results Ordinary tool text. Each of the five progression tools returns its complete bounded result as JSON text, including edges, statements, confidence, warnings, links and cursors. Same tools and text.
Next page of a search or direct-connections result Ask Claude to call the same tool with page.nextRequest unchanged. Same.
Full provenance and supporting evidence Attach a resource from the connector's menu where it is offered (only Catalog on an earlier build; more resources on 0.4.0). For any other link, ask Claude to call read_evidence with the exact URI. Attach a server resource in claude.ai (observed working, 2026-10-08), or ask Claude to call read_evidence.
Standard and learning-component evidence Workflow instructions list exact links and per-record link patterns (EVIDENCE LINKS); Claude fills in IDs from tool text and reads them with read_evidence. Same.
Workflow instructions Native prompt from Add from curriculum-knowledge-graph (observed working on an earlier build; prompts also appeared on 0.4.0, 2026-10-08). Alternative: ask Claude to call get_workflow_instructions. The server's prompts showed in claude.ai (observed, 2026-10-08). Alternative: get_workflow_instructions covers the seven teaching, study and progression workflows.

get_workflow_instructions does not cover administrator_alignment_review or cross_framework_comparison; use those as native prompts. See Evidence and workflow access tools for both tools.

Known client limits that matter here:

  • On an earlier build (2026-10-03), the Desktop resource menu listed only Catalog and could not find a progression link by URI. On 0.4.0 (2026-10-08) it listed more than Catalog; whether it can find a progression link was not recorded. Use read_evidence for links returned by tools.
  • Anthropic documents a tool-result limit of roughly 150,000 characters for connectors. This server keeps each tool result, text and structured copy combined, under 100,000 characters and 1 MiB. Pages and paths stop early rather than exceed that, and say so.
  • Remote connections come from Anthropic's cloud, not your machine, so the hosted endpoint must be reachable from the internet.

Claude Desktop walkthrough

Run this in a new Desktop conversation with the curriculum-knowledge-graph connector enabled, against the 0.4.0 server from this repository or a 0.4.0 bundle. Each step gives a message to send and what you should see. Claude may summarise in its reply; to check details, expand the tool call to see the raw result, or ask Claude to quote the fields named.

The walkthrough uses one fixed diagnostic case from the Nigeria curriculum:

Item Value
Framework ID nigeria-nerdc-mathematics-primary-1-3
Snapshot ID nigeria-nerdc-mathematics-primary-1-3@undated+bc5e769ed26f
Relationship ID 0129f5d5-42fd-52cb-bcf2-ec07c47103e7 ("write correctly number 1-5" builds towards "write correctly number 6-9")
Target standard node ID e399b510-48bb-58ee-abda-61460a5a853b ("write correctly number 6-9", PRIMARY ONE)

Run status. Every tool call below was replayed against the same server code and data through an in-process MCP client on 2026-10-07, and the expected results were checked. On 2026-10-08 a user ran steps 1–13, including 9a, in Claude Desktop on macOS against 0.4.0 and observed no failures. The results are in the table at the end; add your own there.

Before you start, check that Desktop runs the current server. In step 1, each framework should show Routing graph types: and Included graph types: lines. If it shows a single Graph types: line instead, Desktop is running an older copy. For a bundle, follow the removal and reinstall steps in Claude Desktop installation (unverified in this project). For a checkout added in Desktop's configuration file, verify its paths, fully quit Desktop and reopen it.

1. Discovery

Use curriculum-knowledge-graph to list all available frameworks.

Expected: six framework snapshots, including Nigeria Mathematics with the snapshot ID above. Each snapshot shows Routing graph types: academic_standards and Included graph types: academic_standards, learning_components, learning_progressions. Each package line shows learning_progressions=available and its stored link counts; for Nigeria these are builds_towards=189 and relates_to=297.

Call list_frameworks with graphTypes ["learning_progressions"] and list the snapshot IDs.

Expected: the same six snapshots. The filter matches the graph types a package includes. The routing type stays academic_standards: it is the type tools use to pick a package, not a list of what the package contains.

Call get_capabilities and tell me how many tools, prompts and resource templates it reports, and which graph types are included.

Expected: 19 tools (including read_evidence and get_workflow_instructions), nine prompts, one fixed resource and 14 templates. Included graph types: lists all three types, and each package has a Learning progressions: block with hasLearningProgressions: true and its two stored counts.

2. Exact relationship

Call get_learning_progression with {"request":{"frameworkId":"nigeria-nerdc-mathematics-primary-1-3","snapshotId":"nigeria-nerdc-mathematics-primary-1-3@undated+bc5e769ed26f","relationshipId":"0129f5d5-42fd-52cb-bcf2-ec07c47103e7"}} and show me the relationship type, both endpoint statements, the confidence, the warnings and the provenance link.

Expected in the JSON text:

  • relationshipType buildsTowards, from 12b603df-b6f0-50b3-914d-5d655a2dbaf1 ("Pupils should be able to: write correctly number 1-5;") to the target standard;
  • judgment.confidence 0.72 with the notice that it is a model judgment;
  • warningCount 9, five warningExcerpts and warningsOmitted: true, and rationaleExcerpted: true — the tool shows excerpts, and the full record is in step 7;
  • relationshipUri, provenanceUri, standardUri for both endpoints, the generated-origin notice and a continuationNotice saying this operation has no continuation.

3. Direct connections

Call get_standard_progressions for node e399b510-48bb-58ee-abda-61460a5a853b in that snapshot three times, with connectionKind "outgoing_builds", "incoming_builds" and "related", and list each connection with its kind and relationship ID.

Expected: no outgoing_builds; two incoming_builds (0129f5d5… from 12b603df…, and c0155efa… from c4885674…, "Count and read correctly from 1-9"); and six related. Each page has isComplete true and no cursor.

Asking for connectionKind "all" returns the same eight connections here in one page. For a standard with many links, an all page lists builds first and can stop at the result-size limit before any related links appear. That is why the workflows ask for each kind separately.

4. Search and pagination

Call search_learning_progressions with {"request":{"frameworkId":"nigeria-nerdc-mathematics-primary-1-3","snapshotId":"nigeria-nerdc-mathematics-primary-1-3@undated+bc5e769ed26f","relationshipTypes":["buildsTowards"],"endpointScope":"either","limit":25}}. Report page.returnedCount, page.stoppingReason and whether page.nextCursor is present.

Expected: 7 relationships, stoppingReason byte_limit, isComplete false, and both page.nextCursor and page.nextRequest present. limit is a maximum; the page stopped early to stay within the result-size limit.

Call search_learning_progressions again with page.nextRequest exactly as returned, and list the new relationship IDs.

Expected: seven different relationship IDs. Following every page gives 27 pages and 189 distinct buildsTowards edges, with no repeats.

5. Upstream traversal

Call traverse_learning_progressions upstream from node e399b510-48bb-58ee-abda-61460a5a853b in that snapshot with maxDepth 3, maxNodes 20 and maxEdges 30.

Expected: five standards and five stored edges, scopeComplete and graphExhausted true, an empty frontier and no truncation reasons. The deepest supports are "count correctly up to 5" (ec0c6a6d…) and "sort and classify number of objects" (b193d5b5…).

6. Connecting paths

Call get_learning_progression_paths from node ec0c6a6d-0a48-5aa8-bdfe-72260dd3b559 to node e399b510-48bb-58ee-abda-61460a5a853b in that snapshot with maxDepth 3 and maxPaths 3.

Expected: two two-hop paths, through c4885674… (41ec8cba…, c0155efa…) and through 12b603df… (7fc3c0e1…, 0129f5d5…); scopeComplete true, graphExhausted false, truncationReasons ["depth_limit"], nextUnreturnedPath null. A path is derived from stored edges; it is not a compulsory teaching order.

7. Full provenance

Call read_evidence with the provenanceUri from step 2. If page.isComplete is false, keep calling with page.nextRequest until it is true. Then summarise the full rationale, all warnings and the producer/checker trace.

Expected: one window, contentStatus full, 12,275 bytes, holding the full rationale, confidence, all nine warnings and the producer/checker trace with hashes. To see continuation in action, ask for "maxContentBytes":4096: the same record then arrives in three windows.

8. Standard and learning-component evidence

Call get_learning_components_for_standard for node e399b510-48bb-58ee-abda-61460a5a853b in that snapshot.

Expected: one supporting learning component, 20507dfe-4d56-575c-b7ea-33b1f9072300, "Write correctly the numbers 6 to 9", support confidence 0.97, labelled as a model-generated decomposition. The text also gives what you need to cite it: the Standard URI, and for the component Support relationship ID: 5f89e75a-70c5-5260-b325-31fcba940574, its Support relationship URI, Direction: component -> standard, the Component URI and the Component provenance URI.

Read the Support relationship URI and the Component provenance URI from that result with read_evidence, and also kgfegmcp://framework/nigeria-nerdc-mathematics-primary-1-3/snapshot/nigeria-nerdc-mathematics-primary-1-3%40undated%2Bbc5e769ed26f/standard/e399b510-48bb-58ee-abda-61460a5a853b/provenance.

Expected: all three complete in one window (1,049, 541 and 3,077 bytes). The support relationship record is a supports relationship from the component to the standard. The standard provenance link follows the patterns in the workflow's EVIDENCE LINKS block (step 11).

9. Progression coverage, validation and unresolved claims

Call read_evidence for the summaryUri, validationUri and unresolvedUri from step 2. Then read the learning-progressions summary for snapshot india-cbse-science-learning-framework-classes-9-10@undated+576740bed2d1.

The CBSE link is:

kgfegmcp://framework/india-cbse-science-learning-framework-classes-9-10/snapshot/india-cbse-science-learning-framework-classes-9-10%40undated%2B576740bed2d1/learning-progressions

Expected: all four are progression records, each complete in one window. The Nigeria summary (4,577 bytes) reports stored counts, eligibility limits and that validation is structural only. The progression validation report is 1,806 bytes. The progression unresolved record (214 bytes) lists no unresolved claims for Nigeria (claims empty, total_needs_review 0). The CBSE summary reports one needs_review claim kept out of accepted edges.

9a. Optional: a long record in several windows

This step only shows paging on a long record. The record is not progression evidence: the /unresolved resource is the standards-hierarchy report, listing parent–child links in the Ghana Mathematics source that could not be resolved.

Call read_evidence for kgfegmcp://framework/ghana-nacca-primary-mathematics-basic-4-6/snapshot/ghana-nacca-primary-mathematics-basic-4-6%402019%2B0b768f7cfaf9/unresolved, keep calling with page.nextRequest until page.isComplete is true, and tell me how many calls it took.

Expected: 42,831 bytes in three windows at the default window size of 16,384 bytes (contentStatus partial until all are joined). The number of windows depends on the window size: with "maxContentBytes":4096 the same record takes 11.

10. A policy outcome

Call read_evidence for the learningProgressionProvenance artifact URI listed in metadata.artifacts from step 2.

Expected: resource_access_denied — this package does not allow bulk artifact reads. Claude should report the denial, not invent the content. Per-edge provenance (step 7) remains readable.

11. Support plan through the native prompt

  1. Open Add from curriculum-knowledge-graph and choose learning_progression_support_plan.
  2. Enter framework_id nigeria-nerdc-mathematics-primary-1-3, snapshot_id nigeria-nerdc-mathematics-primary-1-3@undated+bc5e769ed26f, identifier {"identifierType":"node_id","nodeId":"e399b510-48bb-58ee-abda-61460a5a853b"}, local_context Learners can count to 9 but often write 6 and 9 the wrong way round. and output_language en.
  3. Check the attachment: prompt version 1.4.0, the pinned snapshot, an EVIDENCE ACCESS section and an EVIDENCE LINKS block.
  4. Send it, and let Claude make the listed calls and evidence reads.

Expected answer: it starts from the target standard; cites the incoming edges (0129f5d5…, c0155efa…) as [GENERATED-EVIDENCE / llm_inferred] with their provenance links; keeps related concepts separate; cites component 20507dfe… with its support relationship 5f89e75a…; labels review and practice ideas as generated; treats the classroom note as a teacher report, not a diagnosis; and repeats the NERDC attribution. It reads at most ten distinct edge-provenance records and at most 32 evidence windows, and says so if it had to stop.

12. Same workflow through tools only

Call get_workflow_instructions with {"request":{"workflowName":"learning_progression_support_plan","frameworkId":"nigeria-nerdc-mathematics-primary-1-3","snapshotId":"nigeria-nerdc-mathematics-primary-1-3@undated+bc5e769ed26f","identifier":{"identifierType":"node_id","nodeId":"e399b510-48bb-58ee-abda-61460a5a853b"},"localContext":"Learners can count to 9 but often write 6 and 9 the wrong way round.","outputLanguage":"en"}}, then follow rendered.message exactly.

Expected: rendered.message is the same text as the native prompt in step 11, and instructionsNotice says the server has not run it. Claude then performs the same retrieval and composition.

Call get_workflow_instructions with {"request":{"workflowName":"learning_progression_curriculum_review","frameworkId":"ghana-nacca-primary-mathematics-basic-4-6","snapshotId":"ghana-nacca-primary-mathematics-basic-4-6@2019+0b768f7cfaf9","localGradeLabels":["BASIC 5"],"endpointScope":"either","outputLanguage":"en"}}, then follow rendered.message exactly.

Expected: step 3 of rendered.message lists two separate scans, a buildsTowards scan and a relatesTo scan, each allowed up to three pages. Each page returns seven relationships (stoppingReason byte_limit), so Claude reviews 21 links of each type, and both scans still have a cursor after their third page. The answer should report the two types separately (Ghana Mathematics stores 299 buildsTowards and 300 relatesTo links), say the review is incomplete for both, not add the counts together, and fully inspect at most ten relationships, up to five of each type.

To review one type only, add "relationshipTypes":["relatesTo"] to the request; the native prompt takes relationship_types ["relatesTo"]. The instructions then contain a single relatesTo scan.

Record your results

Step Recorded results Your result (date, client version, pass/limited/fail, notes)
1–10 (and 9a) Tool calls replayed locally and passed, 2026-10-07. Run in Claude Desktop 0.4.0 on macOS, 2026-10-08: no failures observed.
11 Prompt appeared and rendered in Desktop on an earlier build, 2026-10-03; rendering of 0.4.0 checked locally, 2026-10-07. Run in Claude Desktop 0.4.0 on macOS, 2026-10-08, with the composed answer checked: no failures observed.
12 Rendering checked locally (identical to native), 2026-10-07. Run in Claude Desktop 0.4.0 on macOS, 2026-10-08, with the composed answer checked: no failures observed.
13 Rendering and the six scan pages checked locally, 2026-10-07. Run in Claude Desktop 0.4.0 on macOS, 2026-10-08, with the composed review checked: no failures observed.

The 2026-10-08 results are a user's observations; the Desktop app version was not recorded. The same steps gave the same results in a claude.ai custom connector on that date.

Mark a step limited if it worked but Claude only summarised, needed extra instructions, or a client feature was unavailable.

claude.ai remote acceptance checklist

Use this to check the hosted service after each deployment (see Hosted deployment). Deployment timing and sign-off are the operator's decision.

On 2026-10-08 the public deployment at https://kg-for-ed-global-mcp.up.railway.app/mcp was confirmed to run 0.4.0, and a user's claude.ai results against it are recorded in What has and has not been checked. No result from the HTTP smoke in item 1 is recorded for that deployment.

  1. From a repository checkout, run the HTTP smoke against the endpoint:

    uv --directory backend run --locked --no-dev kgfegmcp-http-smoke \
  --url https://kg-for-ed-global-mcp.up.railway.app/mcp

    It must pass before testing in claude.ai. This proves the deployed server, not the claude.ai client.

  2. Add or refresh the custom connector (endpoint URL including /mcp, no authentication) and enable it in a new conversation.

  3. Repeat walkthrough steps 1–10 in claude.ai. Confirm that get_capabilities reports 19 tools, that frameworks show Included graph types: with learning_progressions, and that step 2's tool text shows the edge, statements, confidence, warnings and links, not just a summary.
  4. Check pagination (step 4) by having Claude replay page.nextRequest, and check that the result does not appear cut off.
  5. Check full evidence through read_evidence (steps 7–9a), including a multi-window record.
  6. Record whether claude.ai offers the server's prompts or resources in its interface. If it does, try step 11; if not, note it as an unsupported client feature.
  7. Run step 12 (get_workflow_instructions) and let Claude complete the support plan. Check the answer against the expectations in step 11.
  8. Check the denial in step 10.
  9. Run step 13 and check that the review includes relatesTo links and reports each type separately.
  10. Write down the date, endpoint, connector name, plan type and each result, keeping server results, client observations and untested items separate. For the 2026-10-08 check, the connector name and plan type were not recorded.

Next: First queries