Claude Code Fixes MCP Result Mix-Ups
Claude Code 2.1.326 fixes an MCP result mix-up and wrong-server tool lookup. See the exact cases and a safe checklist for your own setup.

On November 16, 2026, Anthropic released Claude Code 2.1.326 with fixes for two distinct MCP problems: parallel calls could read another call’s saved result when two long results arrived in the same millisecond, and an exact-name tool search could load a tool from the wrong MCP server. These are Anthropic’s reported fixes, not findings from our own tests. If your workflow uses multiple MCP servers, verify both the server selected for a call and the result returned before relying on a successful tool response.
What changed in the November 16 release
The 2.1.326 release notes describe several MCP corrections. The table separates the documented condition from a proposed check a team can perform with disposable data. The checks are suggestions; Anthropic does not claim they are a certification procedure.
| Documented trigger | Earlier behavior reported by Anthropic | Proposed check after updating |
|---|---|---|
| Two long parallel MCP results arrive in the same millisecond | One call could read the other’s saved result | Make two harmless calls with visibly different outputs; compare each returned value with its request and server |
| Exact-name tool search across MCP servers | Search could load another server’s tool, and may ask again for the requested tool | Give controlled tools distinct names and confirm the loaded server as well as the tool name |
| Two servers have nearly identical names | MCP prompts could disappear | List available prompts from each test server and record which server supplies each |
| An SDK host just removed an MCP server | A call could reconnect to that server or misleadingly request authentication | Remove a disposable test server, then confirm that a later call reports its unavailability without reconnecting |
A tool schema starts with anyOf or oneOf |
Claude could see only the first form of an argument | Try a harmless fixture accepted by an alternative branch; inspect the actual argument type received |
The first row is unusually specific. Anthropic names two long results arriving in the same millisecond. The notes do not say that all parallel calls were mixed up, how often this happened, or which earlier versions were affected. They also do not establish that private data crossed a security boundary in a real deployment. A team should treat the report as a reason to check its own result handling, without assigning an unsupported severity or incident count.
Check server identity as well as the result
The result collision and the wrong-server lookup occur at different points. In the collision described by Anthropic, the requested calls already exist, but a saved result can be read by the other call. In the tool-search case, the client can load a tool belonging to another MCP server when asked for a tool by its exact name. A correct-looking answer alone cannot distinguish those cases.
For a controlled check, use two test servers or two read-only tools that return clearly different, non-sensitive markers. Record the requested server, the exact tool name, a request identifier you assign to the test, and the expected marker before each call. Compare that record with the loaded server and returned result. If the output does not match, stop the dependent workflow and preserve a sanitized trace for the person who manages the MCP configuration. Do not put production credentials or customer data in a test marker.
This is a proposed validation pattern, not an assertion that we reproduced the defect or that a single successful run proves every timing case is fixed. The MCP tools specification describes tools as server-exposed operations identified by names and input schemas. In a multi-server workflow, recording the server alongside the name makes a result easier to audit.
Near-identical server names deserve their own check. The release says they could make MCP prompts disappear; it does not say the server stopped working or that a prompt was sent to the wrong destination. List the prompts you expect from each controlled server before and after the client update. Keep that prompt inventory separate from the tool-result comparison, so a missing prompt is not mistaken for a mixed result.
Schema alternatives and removed servers need separate checks
Anthropic also reports a fix for MCP tools whose input schemas begin with anyOf or oneOf: Claude could previously see only the first branch’s form of an argument. The notes give a concrete type consequence: text such as 42 can now arrive as a number where another branch accepts one. The Claude Code MCP guide discusses root-level schema combinators, while the release notes establish this version-specific behavior.
For a tool you control, define a harmless input whose string and numeric forms produce distinct test output. Capture the type the server actually received and compare it with the branch intended by the request. Do not infer the type from how the final answer displays the value. A real tool may have side effects, so use a fixture or a read-only test endpoint instead of trying this on a production action.
The removed-server fix is different again. Anthropic says a tool call to an MCP server just removed by an SDK host could reconnect to it or report that authentication was needed. A useful check is to remove a disposable server under the host’s normal procedure, then ask for its tool and record whether the client reports the server as unavailable. An authentication prompt in this situation should be investigated against the server inventory before anyone signs in again. The release does not promise a particular error message for every host.
What to verify before relying on a tool result
Start by recording the installed Claude Code version and the MCP servers the workflow expects. Anthropic’s MCP guide explains how Claude Code connects to and manages servers; use the configuration view appropriate to your installation. For each disposable call, record a short row like this:
| Request ID | Expected server and tool | Expected output or type | Observed server and result | Decision |
|---|---|---|---|---|
| Example A | Test server A, read marker | lavender as text |
Fill in after your test | Pass only if both server and value match |
| Example B | Test server B, read marker | mint as text |
Fill in after your test | Pass only if both server and value match |
These are illustrative markers, not results from DailyTech testing. Run the comparison on the actual client and host combination you use. If the result is unavailable, ambiguous, or from a different server, mark it unverified and avoid using it as input to a later action. A tool’s success status is only one signal; the identity and contents of its output matter too.
For adjacent Claude Code changes, our guide to 2.1.321 OAuth redirects and file checks covers sign-in and local permissions, while the 2.1.319 unattended jobs guide covers process status versus completed output. Those checks answer different questions. Our agent monitoring cost guide helps budget a separate monitoring workflow. The AI model tracker tracks models rather than validating this Claude Code client patch; the API cost calculator can estimate a separate API workload from your figures, not the cost of this fix.
Anthropic has not published an affected-user count, failure frequency, measured speed improvement, or dollar saving for these MCP fixes in the release notes. The practical question is narrower: after updating, can your own controlled workflow show that the intended server received the call and that the intended result came back? Keep the evidence tied to each request and avoid treating a maker’s release note as an independent reproduction.
Sources
- Anthropic Claude Code v2.1.326 release notes — release date and reported MCP fixes.
- Anthropic Claude Code MCP guide — current server and schema documentation.
- Model Context Protocol tools specification — tool names and schemas in the protocol.
All pages read on November 17, 2026.
Free toolWhat will 10,000 chats a month cost?Estimate my bill

