ChatGPT custom MCP OAuth shows zero tools despite successful pairing

[wey-gu]

Summary

A Discord user reports that Nowledge Mem 0.10.54 can complete Remote MCP OAuth with ChatGPT custom MCP App, but ChatGPT shows zero available actions.

Source

Discord Chinese channel, message 1535519574680936538 on 2026-08-08.

Reported environment:

• Nowledge Mem Desktop / nmem-server 0.10.54
• macOS Apple Silicon
• ChatGPT Plus, Developer Mode enabled
• Cloudflare Tunnel exposing https://<domain>/mcp
• Built-in Remote MCP OAuth

Reported symptoms:

1. ChatGPT creates the App and namespace appears.
2. OAuth succeeds; Nowledge Mem records an approved ChatGPT client and issues tokens.
3. ChatGPT shows no available app actions.
4. Refresh / reconnect still ends with connection setup problem and tool count stays 0.
5. Local direct /mcp/ initialize succeeds; tools/list returns 67 tools; memory_search and get_memory_by_id are present, but core read tool annotations were observed as null.

Confirmed code finding

Desktop MCP runtime still builds most tools from nmem-rs/golden/snapshots/mcp-tools-full.json; that snapshot has stale annotations: null for core shared reads including memory_search and read_working_memory, even though nmem_wire::tools already declares read-only annotations.

This is a real compatibility gap for ChatGPT-style read-only MCP hosts, not just user confusion.

Immediate fix

Patch prepared in codex/fix-mcp-read-tool-annotations:

• shared tools now get runtime annotations from nmem_wire::tools, not stale golden nulls;
• targeted regression covers memory_search and read_working_memory;
• external-agent payload budget test adjusted to keep the annotated slim payload under 9.5KB.

Follow-up work still needed

• Validate with real ChatGPT custom MCP after the patch ships.
• Consider a first-class chatgpt-readonly / OAuth read-only profile so ChatGPT does not receive the full 67-tool desktop surface.
• Revisit whether Remote MCP OAuth default scopes should include write scopes for generic ChatGPT clients.
• Check whether offline_access / refresh-token discovery wording matters for ChatGPT compatibility.
• Improve unauthorized remote MCP responses so users see standard 401 diagnostics instead of Cloudflare-looking 502s where possible.

[wey-gu]

Update:

• Discord reply sent: message 1535551449600364565.
• Fix PR opened: nowledge-co/mem#858.

The PR fixes the confirmed immediate compatibility gap: desktop MCP shared read tools now advertise runtime annotations from the shared registry instead of stale golden nulls. It intentionally leaves the broader ChatGPT read-only profile / OAuth scope shaping / offline_access discovery questions open in this issue for follow-up validation.

[wey-gu]

Update: the immediate runtime-annotation fix has been merged in nowledge-co/mem PR #858.

What this fixes:

• shared MCP tools no longer inherit stale annotations: null from the desktop golden snapshot;
• core read tools such as memory_search and read_working_memory now get their runtime annotations from the shared registry (readOnlyHint, destructiveHint: false, openWorldHint: false, etc.).

Still tracked here after the patch release:

• verify the real ChatGPT custom MCP flow against the shipped build;
• decide whether ChatGPT should get a smaller read-only tool profile by default;
• review OAuth default scopes / offline_access behavior;
• clean up non-auth diagnostics that can show 502 instead of a clear 401/authorization-required response.

[nierclaw-sudo]

Real ChatGPT validation update from the original reporter's environment:

• Nowledge Mem Desktop / nmem-server: 0.10.55
• ChatGPT Plus, Developer Mode, custom MCP App
• Remote MCP OAuth completed successfully
• Requested scopes were reduced to read-only: mem:read, threads:read, library:read
• The App shows connected, but ChatGPT still reports 0 available actions
• Refreshing after OAuth and recreating the App did not change the result
• A ChatGPT Work conversation could select the App namespace, but the session control layer reported status: not_installed / no callable functions

Local direct /mcp/ on 0.10.55 still returns 67 tools. Current manifest counts:

• 23/67 tools have non-null annotations
• 13/67 have readOnlyHint: true
• 44/67 still have annotations: null
• memory_search now has the expected read-only annotations
• get_memory_by_id still has annotations: null

Important validation note: ChatGPT once claimed in text that a tool call succeeded, but there was no tool event and the Sources panel was empty. A follow-up inspection confirmed no NMEM result was actually obtained.

So PR #858 fixed part of the manifest, but the shipped 0.10.55 build still does not expose actions in a real ChatGPT custom MCP flow. This may need a fully validated read-only tool subset / complete manifest cleanup, plus investigation of authenticated remote tools/list and the ChatGPT App session installation state.

[wey-gu]

Thanks for the careful real-environment validation — and sorry that 0.10.55 still showed 0 actions. The earlier fix (#858) only fixed part of the descriptor problem; your follow-up made it clear that the shipped surface was still too ambiguous for ChatGPT.

I opened nowledge-co/mem#886 with the fuller fix:

• Remote MCP OAuth now defaults to read-only scopes: mem:read, threads:read, library:read.
• ChatGPT/OAuth tools/list gets a bounded read-only surface instead of the full desktop compatibility catalog.
• All surfaced runtime tools must have explicit readOnlyHint, destructiveHint, and openWorldHint values.
• get_memory_by_id is now included in mem:read, so the flow can be search -> exact Memory read.
• mem_fs is advertised as read-only under read-only OAuth, and write / rm / delete require the explicit mem:write scope.

Local targeted tests are passing. After this ships, the safest validation path is to refresh/recreate the ChatGPT connector once, because ChatGPT may keep a scanned/cached tool manifest from the previous build.

[wey-gu]

Update: nowledge-co/mem#886 has been merged to main.

This is the fuller ChatGPT custom MCP compatibility fix beyond #858:

• OAuth/ChatGPT remote MCP now gets a bounded read-only tool surface by default instead of the full desktop catalog.
• Every exposed tool in that surface has explicit MCP annotations.
• get_memory_by_id is included in the read scope, so the intended flow is search -> exact memory/thread/library read.
• write/delete-style operations stay behind explicit write scopes.

Next validation step: after the next patch build containing #886 is shipped, please refresh/recreate the ChatGPT custom MCP App once before testing again, because ChatGPT can cache a previously scanned manifest. We are keeping this issue open until a real shipped-build validation confirms that ChatGPT shows callable actions.