Skip to content
vavio

Why does Hermes reject my model or ignore its tools?

If Hermes refuses to start with a context-window error, check the window your model server actually exposes. If Hermes answers but prints a tool request as ordinary text, check the server’s tool-calling configuration. These failures need different fixes; paying for a larger hosting computer alone does not identify either cause.

First identify the symptom

What you see First check What a successful retest shows
An error naming a context window below the minimum The model server’s active context setting and the window Hermes detects Hermes starts and a simple prompt gets a reply
A reply containing tool-shaped JSON, with no tool running The server’s chat template and tool-call parser Hermes records an actual tool invocation and its result
A real tool runs but returns an access or authentication error That tool’s credentials, permissions and inputs The same small task returns the intended result

A normal chat reply only tests the first row. It does not establish that the model can use tools.

Reproduced: Hermes refuses a 32,768-token window

On 11 November 2026, we ran unmodified Hermes Agent v0.23.0, commit 2bd378ed1b8d6bf2192cb26e0a55ff3091e6bd19, against an isolated local test endpoint. It advertised a model named repro-context with a 32,768-token window. Hermes stopped during initialization with this error:

Model repro-context has a context window of 32,768 tokens, which is below the minimum 64,000 required by Hermes Agent.

The endpoint received metadata requests, but no chat-completion request. This happened before any model could answer the prompt.

We then changed the test endpoint’s advertised window to 65,536, started a new Hermes process with a fresh configuration directory, and repeated the prompt. Hermes detected 65,536, sent one chat-completion request, and returned CONTEXT_RETEST_OK.

This test verifies Hermes’s startup check and recovery when endpoint metadata changes. The endpoint supplied a fixed reply; it ran no model and exercised no tools. It does not prove that a real model can fit a 65,536-token conversation or complete your job.

Hermes’s pinned startup code checks a 64,000-token minimum, with an exception for an explicit positive context setting on the LM Studio provider. The 65,536 value above was our retest setting, not the exact minimum.

Fix the served window, then restart and retest

For a model you host yourself, check the model’s supported context, available memory and server settings together. Configure a window of at least 64,000 that the model and machine can actually handle. Hermes’s error names examples: llama.cpp uses -c, vLLM uses --max-model-len, and Ollama uses its context configuration. Follow the provider guide for your server, restart it, then start a fresh Hermes session and retry a simple prompt.

Check the window per active session. For example, the guide warns that llama.cpp parallel slots divide the configured total context: -c 64000 -np 4 leaves only 16,000 per slot.

If a hosted provider really supplies a smaller window, choose a suitable model. If the window is large enough but metadata is wrong, verify the true value before setting Hermes’s model.context_length override. An override changes what Hermes believes; it allocates no extra server memory or context capacity.

Once the reply works, retry one small task that requires a tool. Confirm the tool ran and inspect its output before restoring your full workflow.

Documented: tool calls arrive as text

A model server can return convincing text about using a tool without returning a structured tool call. Check the server before changing your prompt repeatedly.

  • llama.cpp: its function-calling documentation enables OpenAI-style function calling through llama-server --jinja. The model’s chat template also matters. Hermes’s guide describes raw tool JSON appearing as text when this setting is missing.
  • vLLM: its tool-calling documentation requires --enable-auto-tool-choice and a --tool-call-parser appropriate to the model for automatic tool choice. Select the parser for your actual model family; hermes is not a universal parser simply because your agent is called Hermes.

After changing the relevant setting and restarting, repeat the same small tool task. Look for a structured invocation, an actual tool result and a final answer based on that result. More fluent prose is not sufficient evidence.

We checked these parser requirements against documentation; we did not reproduce the missing-flag failures on real llama.cpp or vLLM servers for this article. Passing the context check also does not establish model quality on multi-step work.

What this means for hosted Hermes

Vavio hosts Hermes on a managed computer; the model provider you choose bills inference separately. The tested provider route on Vavio is OpenRouter with GLM 5.3. This article’s local endpoint experiment does not qualify every provider or local model for Vavio, and the hosting plan supplies no GPU for running model weights.

If you already run your own endpoint, fix and retest that endpoint first. If you want help choosing a provider route, see using Hermes without an OpenAI or Anthropic account. If your decision is whether to manage the server yourself, compare VPS and managed Hermes hosting and check the current hosting offer.

Sources

Written by Vavio with AI assistance. Startup refusal and recovery reproduced locally with official Hermes; tool-parser advice checked against upstream documentation.

More from the Vavio blog: Hermes Agent guides