claude.ai Linear connector still on deprecated /sse transport 6 weeks past removal date; reconnecting in Connectors UI does not migrate it

[Rumbles]

Preflight Checklist

• I have searched existing issues and this hasn't been reported yet
• This is a single bug report (please file separate reports for different bugs)
• I am using the latest version of Claude Code

What's Wrong?

The claude.ai-managed Linear connector (mcp__claude_ai_Linear__*) is exposed in Claude Code sessions with a server notice declaring its transport deprecated past the removal date, and every tool call to it is rejected client-side as a "pre-removal deprecation signal." The connector is unusable in this state, but it still appears as connected and its tools are still advertised as available.

Verbatim server notice from the session's MCP instructions block:

## claude.ai Linear
## Server notice

DEPRECATED TRANSPORT — REMOVAL DATE 2026-04-08 (past). The /sse transport on this server is deprecated. Replacement endpoint: https://mcp.linear.app/mcp. Migration guide: https://linear.app/docs/mcp.

Every call to a mcp__claude_ai_Linear__* tool then fails with:

Tool call rejected as a pre-removal deprecation signal. DEPRECATED TRANSPORT — REMOVAL DATE 2026-04-08 (past). The /sse transport on this server is deprecated. Replacement endpoint: https://mcp.linear.app/mcp. Migration guide: https://linear.app/docs/mcp.

Reconnecting the connector from claude.ai → Settings → Connectors (disconnect + reconnect, OAuth re-auth completed successfully) does not change the behavior — the next Claude Code session still emits the same deprecation notice and still rejects all calls. The claude.ai-managed connector definition appears to still resolve to /sse even for fresh reconnections.

There is no local linear entry in ~/.claude.json mcpServers (so this is not a duplicate-registration case like #57697 — the only registration is the claude.ai-managed connector, confirmed via claudeAiMcpEverConnected: ["claude.ai Linear", ...]).

What Should Happen?

One of:

1. The claude.ai Linear connector definition should be migrated to https://mcp.linear.app/mcp on the Anthropic side so that reconnecting (or even just a fresh session) picks up the new endpoint. Currently the connector appears to still resolve to /sse six weeks past the stated removal date, and the user has no way to fix this client-side.
2. If the connector is intentionally being retired, it should be removed from the claude.ai Connectors UI / hidden from Claude Code sessions, rather than being advertised as connected with a usable tool surface that rejects every call.
3. At minimum, the deprecation rejection should be surfaced to the user at session start (not just to the model in a system reminder), with explicit guidance that reconnecting in claude.ai will not fix it and pointing at the local-MCP workaround below.

Error Messages/Logs

Server notice (emitted on every session start):

DEPRECATED TRANSPORT — REMOVAL DATE 2026-04-08 (past). The /sse transport on this server is deprecated. Replacement endpoint: https://mcp.linear.app/mcp. Migration guide: https://linear.app/docs/mcp.

Tool-call rejection (on every mcp__claude_ai_Linear__* call):

Tool call rejected as a pre-removal deprecation signal. DEPRECATED TRANSPORT — REMOVAL DATE 2026-04-08 (past). The /sse transport on this server is deprecated. Replacement endpoint: https://mcp.linear.app/mcp. Migration guide: https://linear.app/docs/mcp.

Today's date when this was observed: 2026-05-22 (~6 weeks past the stated removal date).

Steps to Reproduce

1. Connect the claude.ai Linear connector through claude.ai → Settings → Connectors (no local linear entry in ~/.claude.json).
2. Start a Claude Code session and observe the server notice for claude.ai Linear reporting the /sse transport as past its removal date.
3. Attempt any mcp__claude_ai_Linear__* call (e.g. get_issue, save_issue). Every call is rejected with the deprecation message above.
4. In claude.ai → Settings → Connectors, disconnect Linear, then reconnect (OAuth completes successfully).
5. Start a fresh Claude Code session. The deprecation notice and rejections persist; behavior is unchanged.

Workaround

Bypass the broken claude.ai-managed connector by registering Linear as a local MCP server pointing directly at the new endpoint:

claude mcp add --transport http --scope user linear https://mcp.linear.app/mcp

Restart Claude Code; on first use the local MCP flow will prompt for OAuth. A parallel mcp__linear__* tool family becomes available and works correctly against the new /mcp endpoint. The broken mcp__claude_ai_Linear__* tools remain advertised in parallel and still reject calls — it's worth disconnecting the claude.ai connector from the Connectors UI to suppress the deprecation notice in future sessions, though that doesn't always remove the connector from the session immediately.

This is the inverse of the #57697 workaround: that issue resolved a duplicate-registration case by removing a local SSE entry; here you have to add a local HTTP entry because the claude.ai-managed definition is the broken one.

Claude Model

Opus 4.7 (1M context)

Is this a regression?

Unknown — the connector previously worked in earlier sessions; the rejection behavior appears to have started after the 2026-04-08 removal date.

Last Working Version

No response

Claude Code Version

2.1.148 (Claude Code)

Platform

Anthropic API

Operating System

macOS 26.5 (Darwin 25.5.0 arm64)

Terminal/Shell

zsh

Additional Information

• No local linear entry in ~/.claude.json mcpServers at the time of the bug; claudeAiMcpEverConnected lists "claude.ai Linear".
• Closest prior issue is [BUG] Duplicate Linear MCP registration (local SSE + claude.ai connector) silently exposes both backends; no warning when a server's deprecation removal date has passed #57697 (closed, completed), but that was about duplicate local-SSE + claude.ai registration with the local entry as the broken one. This report is the inverted, single-registration case where the claude.ai-managed connector itself is the broken one.

[github-actions]

Closing for now — inactive for too long. Please open a new issue if this is still relevant.