At a glance
ProblemPrompts fail when a GPT-4.1 model is selected under the MCP-only runner
AffectsEnterprise h2oGPTe 1.6.x
WorkaroundNone — the model works in RAG and General Agent sessions, but not under the MCP-only runner
SolutionUpgrade to 1.7.x

Problem

Prompts submitted through the MCP-only runner fail when a GPT-4.1 model is selected. The same model works correctly in RAG and General Agent sessions, and the MCP-only runner works correctly with other model families such as Claude.

Observed error:

API Error: 400 litellm.BadRequestError: You passed in model=<your-gpt-41-deployment>.
There are no healthy deployments for this model.
No fallback model group found for original model_group=<your-gpt-41-deployment>.

Downstream symptom in the runner logs:

Fatal error in message reader: Command failed with exit code 1 (exit code: 1)

The runner terminates because the upstream model call never succeeds.

Affected versions

ProductEnterprise h2oGPTe 1.6.x (observed on 1.6.56 and 1.6.58)
Runner modeMCP-only
ModelGPT-4.1, registered through H2OGPT_MODEL_LOCK
Routing layerLiteLLM Router
Fixed in1.7.x

Cause

On the 1.6.x line, GPT-series models do not resolve reliably through the LiteLLM router when invoked from the MCP-only runner, even when base_model, display_name and litellm_params.model are all consistently configured. The error reports no healthy deployment for the model group, despite the same model working in other session types.

MCP support was substantially reworked in the 1.7 release line, which introduced native Model Context Protocol integration. Upgrading is the supported resolution.

Rule this out first

A genuine model-configuration or Azure deployment mismatch produces the same error. Before concluding this is the product defect, select the same GPT-4.1 model in a RAG or General Agent session:

  • It works there — the configuration is fine, and you are hitting the issue described here.
  • It fails there too — this is a model configuration problem, not the MCP runner. Check your deployment name and credentials.

Workaround

No workaround available

There is no configuration change that makes GPT-4.1 work under the MCP-only runner on 1.6.x. Until you can upgrade, either use a model family that does work under the MCP-only runner (such as Claude), or use GPT-4.1 in RAG or General Agent sessions instead.

Solution

Upgrade to 1.7.x
  1. Confirm the h2oGPTe version in use. If it is 1.6.x, the reworked MCP runner support is not present.
  2. Review the model definition in H2OGPT_MODEL_LOCK and confirm base_model, display_name and litellm_params.model are consistent for the GPT-4.1 entry.
  3. Upgrade Enterprise h2oGPTe to 1.7.x, which includes native MCP server integration. See the 1.7 changelog.
  4. Retest prompt submission with the GPT-4.1 model under the MCP-only runner.

How to verify

  1. Submit a prompt through the MCP-only runner with the GPT-4.1 model selected, and confirm it returns a response.
  2. Confirm no BadRequestError or no healthy deployments entries appear in the runner logs for that request.
  3. Confirm the runner process stays alive rather than exiting with Fatal error in message reader.

Related