anthropics / anthropics/claude-ai-mcp

claude.ai web does not render MCP progress notification messages for remote connectors

Offen
#804 0 Kommentare 2 Reaktionen 0 zugewiesene Personen Auf GitHub ansehen
bug
Vorherrschende Sprache
Keine Sprachdaten
Sterne
471
Forks
77
PR-Merge-Kennzahlen
Keine gemergten PRs in 30 T.

Beschreibung

### What happened?

Our remote MCP connector (custom server, OAuth 2.1, Streamable HTTP) emits spec-compliant notifications/progress during a long-running tool call: monotonically increasing progress values with human-readable message text ("Analyzing Google Analytics data...", "Finalizing response..."). Tool runs legitimately take 30-200 seconds.

On claude.ai web, the tool call renders as a collapsed chip with a spinner for the entire duration. The progress notifications are delivered (confirmed at the transport level with an MCP client against the same server) but the message text is never displayed anywhere in the UI, expanded or collapsed.

### What did you expect to happen?

The latest progress message renders inside or under the running tool chip, updating as notifications arrive. The MCP spec says the message field "SHOULD provide relevant human readable progress information", which exists precisely for client display. Claude Code had the same gap and fixed it (progress text renders as of v2.1.153, per [anthropics/claude-code#51713](https://github.com/anthropics/claude-code/issues/51713)); requesting parity on claude.ai web.

### Steps to reproduce

1. Add a custom remote connector (Streamable HTTP) whose tool takes 30+ seconds and emits notifications/progress with a progressToken, increasing progress, total, and a message string.

2. In a claude.ai web chat, ask a question that invokes the tool.

3. Observe the tool chip shows only a spinner for the full duration; no progress message text appears at any point.

### Area

MCP Connector (adding/managing servers)

### MCP Server (if applicable)

Custom remote server (FUSE)

### Error messages or logs

```shell
None; this is silent non-rendering, not an error. Server-side logs confirm each notification is sent before the tool result.
```

### Additional context

Impact: any connector with legitimately long-running tools is a black box for claude.ai users, who cannot distinguish a working run from a stalled one. This pushes connector developers toward worse patterns (chunked fake-fast tools, or abandoning claude.ai). Progress notifications also serve as client keep-alive signals, so servers are incentivized to send them even though users never see them. Happy to provide a test connector endpoint on request.

Beitragsleitfaden

Für dieses Repository ist kein Beitragsleitfaden indexiert

Rechercherichtung

No source files, tests, or code entry points are named. Start by reproducing the issue with a remote Streamable HTTP MCP connector that emits progress notifications during a long-running tool call; done means the latest human-readable progress message appears and updates in the running tool chip.

Vom Indexierungsmodell aus dem Issue-Text verfasst.

Bewertung

Bereich
api, frontend
Issue-Typ
Bug
Schwierigkeit
4/5
Geschätzter Aufwand
3-5 Tage
Aktivitätsstatus
Ruhig
Klarheit
Größtenteils klar
Anfängerfreundlichkeit
35/100

Neue Issues direkt in Ihr Postfach

Eine kurze Übersicht über anfängerfreundliche GitHub-Issues.