Unable to connect Microsoft Foundry UI agent to mcp server - 400 BadRequest

Question

Unable to connect Microsoft Foundry UI agent to mcp server - 400 BadRequest

Nutpapon Limsupavanich 20

Hi all

My team was trying to connect the agent in Microsoft Foundry UI to ours mcp server. At first time the connection was work and the agent is able to access mcp but when we recreate a connection, it showed this error. Do you have idea what happened. messageImage_1789090432566

simon 0 Reputation points

2026-09-11T08:52:50.2233333+00:00

We are also seeing the same error today on a different Foundry resource. Tool name DP, first use of that name, nothing to delete under Connected resources at project or account level.

The same wizard worked on this resource on 7 September. Registering the same connection through the ARM connections API still succeeds, so this looks like a change in the connector gateway rather than a leftover artifact. Please escalate.
SRILAKSHMI C 19,730 Reputation points Microsoft External Staff Moderator

2026-09-11T17:17:50.06+00:00

Hello @Nutpapon Limsupavanich

Thank you for reaching out to Microsoft Q&A.

Based on the error, the failure is occurring while Microsoft Foundry is creating/updating the MCP custom connector, rather than when the MCP server is being invoked.

The key error is:

ConnectorGatewayCreateOrUpdateCustomConnector failed ... 400 BadRequest

and specifically:

The custom connector ... uses the DirectInvoke invocation mode, which requires a version 3.0 API definition.

This indicates that the connector is currently being processed with the DirectInvoke invocation mode, but the API definition associated with the connector is not being recognized as a version 3.0 definition.

As the error indicates, there are two supported options:

Update the connector/API definition to a supported version 3.0 API definition while continuing to use DirectInvoke.

Use the ProxyInvoke invocation mode instead.

Since you mentioned that an identical MCP configuration was working previously and the same MCP server works from other clients, I would not conclude at this point that the MCP server itself is unavailable. The failure shown here occurs during the connector registration/update stage in Foundry.

For reference, the current Microsoft documentation for connecting MCP servers to Foundry Agent Service is available here:

Connect agents to MCP server endpoints - Microsoft Foundry

I Hope this helps. Do let me know if you have any further queries.

Thank you!

1 answer

Your answer

simon 0 Reputation points

2026-09-11T08:52:50.2233333+00:00

We are also seeing the same error today on a different Foundry resource. Tool name DP, first use of that name, nothing to delete under Connected resources at project or account level.

The same wizard worked on this resource on 7 September. Registering the same connection through the ARM connections API still succeeds, so this looks like a change in the connector gateway rather than a leftover artifact. Please escalate.
SRILAKSHMI C 19,730 Reputation points Microsoft External Staff Moderator

2026-09-11T17:17:50.06+00:00

Hello @Nutpapon Limsupavanich

Thank you for reaching out to Microsoft Q&A.

Based on the error, the failure is occurring while Microsoft Foundry is creating/updating the MCP custom connector, rather than when the MCP server is being invoked.

The key error is:

ConnectorGatewayCreateOrUpdateCustomConnector failed ... 400 BadRequest

and specifically:

The custom connector ... uses the DirectInvoke invocation mode, which requires a version 3.0 API definition.

This indicates that the connector is currently being processed with the DirectInvoke invocation mode, but the API definition associated with the connector is not being recognized as a version 3.0 definition.

As the error indicates, there are two supported options:

Update the connector/API definition to a supported version 3.0 API definition while continuing to use DirectInvoke.

Use the ProxyInvoke invocation mode instead.

Since you mentioned that an identical MCP configuration was working previously and the same MCP server works from other clients, I would not conclude at this point that the MCP server itself is unavailable. The failure shown here occurs during the connector registration/update stage in Foundry.

For reference, the current Microsoft documentation for connecting MCP servers to Foundry Agent Service is available here:

Connect agents to MCP server endpoints - Microsoft Foundry

I Hope this helps. Do let me know if you have any further queries.

Thank you!

Answer 1

Hi @Nutpapon Limsupavanich

Straight up: this exact error code isn't documented anywhere publicly, so there's no official fix to point you to. But the error itself tells us what happened, and there's a practical way around it.

What the error means: when you add an MCP tool with OAuth Identity Passthrough, the Foundry portal silently provisions a custom connector inside a Connector Gateway namespace behind the scenes. The namespace (cg-1b8ab8ec...) comes from your project, and the connector name comes from your connection name, which is why it says '1b8ab8ec...-test'. Your first creation registered that connector. When you deleted and recreated the connection with the same name, the portal ran CreateOrUpdate against the leftover artifact from round one, which is registered in DirectInvoke mode, and the gateway's validation now rejects it for lacking a v3.0 API definition. Neither the invocation mode nor the API definition version is exposed anywhere in the Foundry UI, so this is not a misconfiguration on your side. It's a stale backend artifact colliding with the recreation.

What to do:

Recreate the connection with a different name (anything other than "test"). A new name produces a new connector identity in the namespace and sidesteps the stale artifact completely. Given the first creation worked fine, this should unblock you immediately.

If you need the original name back: delete the old connection everywhere it appears (Management center > Connected resources, at both the project level and the account/hub level), wait 10 to 15 minutes for the gateway to propagate the cleanup, then retry. If it still throws the same error after that, the connector artifact is orphaned server side and only Microsoft can purge it.

If the orphaned case is what you hit, reply on your thread so a moderator can escalate it, and include the full error text, the namespace ID (cg-1b8ab8ec...), your region, and the timestamp. This part of the new Foundry is preview-grade plumbing, and these gateway artifacts aren't customer-manageable, so it's a legitimate product-side cleanup request.

As an alternative path while the portal flow is stuck, the code-first route (attaching the MCP server through a toolbox with the connection defined in YAML, or via the REST/SDK flow) is documented and doesn't depend on the portal's add-tool dialog.

References:

https://learn.microsoft.com/en-us/azure/foundry/agents/how-to/tools/model-context-protocol

https://learn.microsoft.com/en-us/azure/foundry/agents/how-to/tools/connectors