GitHub Repo
Open WebUI Community Proposes MCP Error Message Fix to Improve Feedback for Model Retries
An October 22 report from the Open WebUI community said that when MCP tools fail, the backend converts the content list to a Python string representation, making errors harder for models to interpret. The proposed fix has been updated to handle content without text, but had not been merged as of the time of review.

On October 22, the Open WebUI community reported that MCP error handling in v0.12.3 converts the content list returned by the server directly into a string, then passes it into the model’s next-turn context. This mixes feedback about tool failures with Python object representations and extra fields; the reporter observed that the model may consequently submit the same malformed parameters again. The issue report and proposed fix continued to be updated on the 23rd, but had not been merged as of the time of review. Issue report
The reported trigger is a model generating parseable JSON whose contents do not conform to the tool schema, such as omitting the required sectioncode or passing a value of the wrong type. After the MCP server responds with isError: true, Open WebUI’s MCPClient.call_tool passes the entire content list to an exception object; the outer layer then wraps it in an error result using str(e). As a result, the message originally contained in a text block is surrounded by quotation marks and fields such as annotations and meta. Reproduction steps and code path
This concerns whether an agent can use failure feedback to correct its next operation. The MCP specification distinguishes protocol errors from tool execution errors; the latter can be returned through isError together with content blocks. From an engineering perspective, if a client does not properly extract text, even a server that provides a useful explanation may reduce the model’s chance of recognizing the problem and correcting its parameters. However, there are currently no quantitative tests showing how this issue affects overall task success rates. MCP Tools specification
The corresponding PR #32942 proposes extracting text from the server response. If an error result contains only an image, audio, or resource, it would provide a brief error message to avoid sending empty feedback. This still cannot recover a reason the server did not disclose: the .NET SDK example in the report returns only a generic failure message, and correcting the client’s formatting will not automatically reveal which parameter is missing. Deployment teams can first inspect the tool results actually sent to the model, and track the proposal’s merge status, release, and cross-SDK validation. Proposed fix