[mattzcarey]

Removes the background retry and link-expiry machinery from the MCP client. Servers now behave the way Claude Code's do: a connection makes a few quick attempts (the server's retry options, 3 by default), each one a connect plus discovery. If they all fail, the failure is recorded and the server stays failed across wakes until someone calls reconnectMcpServer(id). A server waiting for authorization keeps the link it was given until the user completes it or reconnects.

Stacked on #2761, part of the work on #2658. The stack had grown background work nobody asked for:

• A durable reconnect schedule (feat(mcp): durable reconnect and authorization-expiry jobs #2662): it retried a failed server every 30 minutes for a day, each attempt waking the object. A server that is down usually stays down.
• A link-expiry job: it recorded links as expired on a timer.
• Link refresh in addMcpServer(): it swapped a dead link for a new one.

We can't know when a link expires. That's the authorization server's call, and when the user clicks a dead link the callback already reports it. main went the other way and recorded nothing, so every wake retried a failed server.

What goes:

• the durable connect job, its backoff policy, the 24h give-up and the reconnect manager option
• the auth-expiry job, the 24h link cap and the expired-link check on wake
• addMcpServer()'s dead-link refresh and the authorization-expired reason
• jobs.ts and the manager's onJob; the MCP client no longer uses the job queue

None of this has been released. Lifecycle logs and drops any leftover mcp job from a preview build.

What stays:

• recorded failures and refusals
• reconnectMcpServer()
• the step-up limit and the mid-session re-authorization checks

addMcpServer() on a registered server returns ready or the pending link as is, and throws for a failed server. The error names reconnectMcpServer(). A stored link that isn't an absolute http(s) URL, which only happens with corrupted storage, still starts a new flow as on main.

The stack now ships one changeset, here. The lower PRs drop theirs, so the release notes describe only what changed for users.

New tests cover the new behaviour:

• 3 attempts hit a down server before it counts as failed, counted by a new 503 counter on the msw server.
• A catalog that fails once is retried. This test fails without the fix.
• A server down on wake stays failed across a second wake without contacting the server, and addMcpServer() reports it without contacting the server either.
• A waiting server's link survives wakes and addMcpServer() after it stops working. Clicking it reports the failure, and reconnecting gives a working link.

The MCP suite (700 tests), typecheck and lint pass locally. CI covers the rest.

[changeset-bot]

🦋 Changeset detected

Latest commit: 5843710

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 2 packages

Name	Type
agents	Minor
@cloudflare/agent-think	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[agent-think]

🟢 agents import sizes: 4 entry points changed, no growth

	Entry point	Exports	Largest gzip change	Size now
🟢	agents/mcp	15 resized	up to -1.3 KiB (-0.35%)	361.6 KiB
🟢	agents/mcp/client	1 resized	-1.2 KiB (-0.71%)	167.5 KiB
🟢	agents	5 resized	up to -1.2 KiB (-0.44%)	272.2 KiB
🟢	agents/chat-sdk	3 resized	up to -1.2 KiB (-0.42%)	274.5 KiB

Changed exports (24)

	Import	Gzip change	Size now
🟢	agents/mcp#getMcpAuthContext	-1.3 KiB (-0.35%)	361.6 KiB
🟢	agents/mcp#McpAgent	-1.3 KiB (-0.35%)	361.6 KiB
🟢	agents/mcp#SSEEdgeClientTransport	-1.3 KiB (-0.35%)	361.6 KiB
🟢	agents/mcp#StreamableHTTPEdgeClientTransport	-1.3 KiB (-0.35%)	361.7 KiB
🟢	agents/mcp#RPC_DO_PREFIX	-1.3 KiB (-0.34%)	361.5 KiB
🟢	agents/mcp#MCP_SERVER_ID_MAX_LENGTH	-1.2 KiB (-0.34%)	361.6 KiB
🟢	agents/mcp#RPCServerTransport	-1.2 KiB (-0.34%)	361.6 KiB
🟢	agents/mcp#DurableObjectEventStore	-1.2 KiB (-0.34%)	361.6 KiB
🟢	agents/mcp#ElicitRequestSchema	-1.2 KiB (-0.34%)	361.6 KiB
🟢	agents/mcp#normalizeServerId	-1.2 KiB (-0.34%)	361.6 KiB
🟢	agents/mcp#WorkerTransport	-1.2 KiB (-0.33%)	365.4 KiB
🟢	agents/mcp#RPCClientTransport	-1.2 KiB (-0.34%)	361.6 KiB
🟢	agents/mcp/client#MCPClientManager	-1.2 KiB (-0.71%)	167.5 KiB
🟢	agents#routeAgentEmail	-1.2 KiB (-0.44%)	272.2 KiB
🟢	agents#getCurrentAgent	-1.2 KiB (-0.44%)	272 KiB
🟢	agents#StreamingResponse	-1.2 KiB (-0.43%)	272 KiB
🟢	agents#Agent	-1.2 KiB (-0.43%)	272 KiB
🟢	agents/mcp#createMcpHandler	-1.2 KiB (-0.29%)	408.8 KiB
🟢	agents#DEFAULT_AGENT_STATIC_OPTIONS	-1.2 KiB (-0.43%)	272 KiB
🟢	agents/chat-sdk#ChatSdkStateAdapter	-1.2 KiB (-0.42%)	274.5 KiB
🟢	agents/chat-sdk#createChatSdkState	-1.2 KiB (-0.42%)	274.6 KiB
🟢	agents/chat-sdk#ChatSdkStateAgent	-1.1 KiB (-0.42%)	273.7 KiB
🟢	agents/mcp#experimental_createMcpHandler	-1.1 KiB (-0.28%)	396 KiB
🟢	agents/mcp#createLegacyMcpHandler	-1.1 KiB (-0.28%)	395.9 KiB

How this works

Each runtime export is bundled on its own, minified, and gzipped. Changes smaller than 100 B, or smaller than 1% and 1 KiB, are ignored. Growth over 10% or 5 KiB is marked 🔴. This report is informational and does not fail CI. The workflow artifact contains every measurement.

_{Compared 3f5beb70 → 58437106 · workflow run · reported by agent-think[bot]}

[pkg-pr-new]

Open in StackBlitz

agents

npm i https://pkg.pr.new/agents@2778

@cloudflare/ai-chat

npm i https://pkg.pr.new/@cloudflare/ai-chat@2778

@cloudflare/codemode

npm i https://pkg.pr.new/@cloudflare/codemode@2778

hono-agents

npm i https://pkg.pr.new/hono-agents@2778

@cloudflare/shell

npm i https://pkg.pr.new/@cloudflare/shell@2778

@cloudflare/think

npm i https://pkg.pr.new/@cloudflare/think@2778

@cloudflare/voice

npm i https://pkg.pr.new/@cloudflare/voice@2778

@cloudflare/worker-bundler

npm i https://pkg.pr.new/@cloudflare/worker-bundler@2778

commit: 5843710

[devin-ai-integration]

Devin Review found 2 new potential issues.

[devin-ai-integration]

🔴 Existing grant blocked after wake

When an agent wakes with a pending step-up, recoverRequest rejects even requests covered by the existing grant. liveStepUps is empty after restore, although the OAuth token and pending link persist.

Learn more

A scope step-up leaves a usable session on the connection while the user decides. restore rebuilds a pending connection as authenticating after a wake, but the in-memory liveStepUps map is not rebuilt. The new guard therefore rejects every call without trying the still-valid grant. A first authorization must still return the pending link immediately, so the guard needs a durable way to distinguish that case from a pending step-up.

Example: Authorize files:read, request write and leave its scope link open, then let the agent wake. read now returns MCPAuthorizationRequiredError instead of succeeding on the existing grant.

Recommended fix: Persist whether the pending flow started on an established session, restore that distinction with the connection, and use it in recoverRequest. Preserve the immediate link response for servers awaiting their first grant.

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[devin-ai-integration]

🔴 Old denial cancels current step-up

When two step-up links exist, declining the older link makes declineStepUp clear the newer link. Approving the newer link then reports stale success without exchanging its code, leaving the scope ungranted.

Learn more

The authorization provider can issue several valid links before the user responds; the server saves the latest link as its active continuation. The callback checks that a state is valid but does not compare it to the active link. Denying an earlier link then clears the active link and moves the connection to ready; isPastAuthorization treats the later approval as already completed. That callback reports success without exchanging its authorization code.

Example: Two write requests produce links A and B. The user declines A and approves B; the agent reports B as authorized, but a later write still needs step-up.

Recommended fix: Match the callback state against the current pending authorization link before restoring the previous connection state. Consume earlier denied states without changing the active flow; keep B available until B completes or is declined.

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[mattzcarey]

Replaced by #2782 (MCP client changes) and #2783 (examples).