"Provider rejected the request and retrying won't help." That is what current Hermes shows in chat for an HTTP 400, followed by "Provider said:" and whatever the provider sent back. In the logs it is Non-retryable client error. Neither tells you which of eight causes you hit. Here they are, from GitHub issues, ranked by how often they're the problem.
Not this error? See all 17 Hermes errors and their fixes.
In May a user filed GitHub issue #26161. Hermes v0.13.0. Gemini provider. First message: "hello." Response: GeminiAPIError [HTTP 400] Bad Request.
The model existed. The API key worked. Curl to the endpoint returned a valid response. But Hermes returned 400 on every attempt.
The error message gave zero diagnostic information. Just "HTTP 400" and "Bad Request." The user had to upload debug logs, agent logs, and gateway logs before anyone could help. The cause turned out to be Hermes sending a thinking_config that the unversioned gemini-3-flash-preview model rejects. #26161 was closed as a duplicate of #25123.
This is the most common runtime error in Hermes Agent. The official FAQ confirms it: "Setup completes fine, but the first chat attempt fails with HTTP 400." Here are the eight causes, sourced from real GitHub issues, with the exact fix for each.
Updated November 17, 2026 for Hermes v0.23.2. The Kimi, OpenCode Zen and Gemini dual-header 400s from spring 2026 are fixed in current builds, so updating is step zero. One cited issue is still open: the Gemini thinking_config rejection (#25123). The model-ID cause still bites on any version, and since June a legacy Gemini key is a new way to get a 400.
Seeing a retry message instead of a 400? Hermes only aborts straight away on a 400. Other provider failures get retried first and show a different message. Each one has its own page in the decoder:
- The model provider failed after retries: usually a retired or renamed model ID, or a provider returning 503 for lack of capacity
- API call failed after 3 retries: one catch-all line covering safety blocks, connection drops, 429/402 billing responses and missing models
- provider rejected the request schema or tool payload: the OpenClaw version of a tool-schema 400 on OpenRouter
Cause 1: Model name mismatch (the most common, by far)

From the official Hermes FAQ: "Usually a model name mismatch. The configured model doesn't exist on your provider, or the API key doesn't have access to it."
The fix:
hermes config show | head -20 ... Check your provider and model ID.
hermes model ... Re-run model selection to pick from the provider's valid list.
hermes chat -q "hello" --oneshot --model anthropic/claude-sonnet-5 ... Test with a known-good model to isolate whether the issue is model-specific. Use any model your provider currently lists. anthropic/claude-opus-4 from earlier versions of this guide is no longer in OpenRouter's model list, so it would 400 too.
The OpenRouter trap: Model IDs on OpenRouter look right but must match exactly. google/gemini-2-flash-preview is different from google/gemini-2.5-flash-preview. One character off = 400 error. OpenRouter documents a 400 as "invalid or missing params", so a wrong model ID and a malformed request can look the same.
The diagnostic shortcut: If hermes chat -q "hello" --oneshot --model anthropic/claude-sonnet-5 works but your configured model doesn't, the problem is almost certainly the model ID. Fix: hermes model and select from the list.
Cause 2: The model or provider route doesn't accept tool calls
GitHub issue #13927: "When running Hermes Agent v0.10.0 with OpenRouter as the provider, all API calls return HTTP 400 regardless of the model selected." The issue is closed. The maintainers' verdict was that it's "just a provider or model that doesn't accept tool calls", not a Hermes schema bug.
What happens: Hermes sends its tool definitions with every request. If the model, or the specific provider route serving it, doesn't support tool calling, the provider rejects the whole request with a 400. The error looks like a model problem but it's a tool-support problem. If instead of a 400 you are getting rejected credentials, that is a different failure mode covered in our Hermes auth error fixes.
How to diagnose: If every model on the same provider returns 400, but the same model works on a different provider (e.g., Anthropic direct works but Anthropic via OpenRouter doesn't), the tool schema is the issue.
The fix: Pick a model that supports tool calling on that provider. On OpenRouter, check the model's supported parameters for tools. To confirm the diagnosis, run once with a smaller -t toolset list. If the 400 goes away, a tool definition is the trigger. hermes chat --safe-mode also disables plugins and MCP servers, but it ignores your user config too, so it won't use your configured model or provider; pass --model and the provider explicitly if you use it. Keep Hermes updated as well: provider-specific schema fixes land in almost every release.
For the broader comparison of how different agent frameworks handle provider compatibility, our comparison covers the provider integration differences.
If you're getting 400 errors specifically with xAI Grok, our Grok on Hermes setup guide covers the Grok-specific causes: a tool-schema rejection fixed in Hermes in May (update Hermes), and a 400 after switching an existing session to Grok (start a fresh session).
Cause 3: Gemini key problems (legacy AIza keys and the old dual-header bug)

The current cause: a legacy key. According to Hermes's Gemini adapter, Google began rejecting legacy "Standard" API keys (the ones starting AIza) on 19 June 2026, and all of them stop working. A leftover AIza key now often returns 400 API_KEY_INVALID, and Hermes adds migration guidance to the error. The fix: create a new key in Google AI Studio (current keys start AQ.), put it in ~/.hermes/.env as GEMINI_API_KEY=your_key, and restart.
The old cause: dual headers. GitHub issue #7893: "HTTP 400 'Multiple authentication credentials received' when using native gemini provider."
What happens: Hermes's gemini provider injects an x-goog-api-key header. Simultaneously, the underlying OpenAI Python SDK injects an Authorization: Bearer header. Google's API rejects requests with both headers as ambiguous credentials.
The fix: update Hermes. Issue #7893 was closed in April, and the native Gemini adapter now uses its own HTTP client that sends only x-goog-api-key. If you still see "Multiple authentication credentials received", you are on a build from before the fix. Earlier versions of this guide told you to switch from an AQ. key to an AIza key; that is now backwards.
Cause 4: Missing reasoning_content for Kimi models
GitHub issue #13848: "When using kimi-for-coding with Hermes Agent, any session that triggers tool calls becomes permanently broken."
What happens: Kimi models with thinking enabled require a reasoning_content field in assistant messages containing tool calls. Older Hermes builds dropped that field. The first tool call succeeded, the broken message stayed in conversation history, and every later turn replayed it and got a 400: thinking is enabled but reasoning_content is missing in assistant tool call message.
The fix: Fixed by PR #14018. Run hermes update, then start a new session with /new. Sessions whose history was already corrupted can keep failing, because the bad message is still in them.
If debugging provider-specific auth headers, tool schema incompatibilities, model name formats, and session corruption from missing fields sounds like more API troubleshooting than agent building, BetterClaw handles provider compatibility at the platform level. 28+ providers. Model switching from a dropdown. No schema debugging. No auth header conflicts. Free tier with 1 agent and BYOK. $49/month for Pro.
Weighing a move off Hermes altogether? Our Hermes Agent alternatives guide compares six options on setup time, hosting and cost.
Cause 5: OpenRouter model requires a paid plan
From the official FAQ: "A 400 from OpenRouter often means the model requires a paid plan or the model ID has a typo."

Running out of credits is a different error. OpenRouter documents 402 for insufficient credits, not 400, so in Hermes it shows up as a retry failure (API call failed after 3 retries) rather than an instant 400.
The fix: check the model's page on OpenRouter for its pricing and access, and check your credit balance while you are there. Free-tier OpenRouter accounts can't access all models. Add credits, or switch to one of the models OpenRouter lists with a :free suffix. google/gemini-2.5-flash, which an earlier version of this guide called free, is a paid model on OpenRouter.
Cause 6: Oversized request payload
What happens: your conversation history, system prompt and tool definitions exceed the provider's maximum request size. Some providers return a plain 400 instead of a context-length error.
Since v0.23.0, Hermes treats these as overflow: byte-cap 400s, and a generic 400 on a large session, trigger compression automatically instead of failing. If you still see it, you are on an older build, or compression could not shrink the request enough (the error then says compression could not reduce the rebuilt request below the safe threshold).
The fix: hermes update first. Then start a new session to clear accumulated history, reduce the tool count, or switch to a model with a larger context window. The minimum is covered in our Hermes context length guide. A related failure where the response gets cut off mid-stream rather than rejected is covered in our guide to Hermes response truncation.
Cause 7: Wrong Gemini endpoint version

What happens: Hermes sends requests to /v1 but some Gemini models require /v1beta. Or vice versa. The endpoint doesn't recognize the model and returns 400.
How to diagnose: The error includes Error 400 (Bad Request)!!1 (the HTML error page from Google, not a JSON error). This HTML response is distinctive.
The fix: check the base URL override in ~/.hermes/.env. The variable is GEMINI_BASE_URL (not GEMINI_API_BASE, which Hermes doesn't read). Hermes's native Gemini adapter defaults to https://generativelanguage.googleapis.com/v1beta. Preview models like gemini-3-flash-preview need v1beta, so if you pinned a /v1 URL, remove it or point it back at the default:
# In ~/.hermes/.env: remove any custom override, or set the v1beta default explicitly
GEMINI_BASE_URL=https://generativelanguage.googleapis.com/v1beta
If the model is gemini-3-flash-preview and you still get a 400 on the first message, that's likely the thinking_config rejection from #25123 (the #26161 case in the intro). That issue is still open, with no fix merged, so pick a versioned Gemini model. Since v0.23.2, Hermes sends Gemini 3 and later through Google's Interactions API by default, which may change this, but we have not confirmed it.
Cause 8: Model ID format mismatch (dots vs hyphens)
What happens: Some providers use dots in model IDs (minimax-m2.5-free). Older Hermes builds normalized the ID to hyphens (minimax-m2-5-free), and the provider didn't recognize the hyphenated version. GitHub issue #7710 documents this with OpenCode Zen. It's closed, and current builds only convert dots to hyphens for claude-* models on OpenCode Zen.
How to diagnose: Run hermes config get model.default and compare it with the provider's documentation. If dots were converted to hyphens, that's the cause.
The fix: hermes update first. If you set the ID by hand, quote it so YAML keeps it as one string. The key is model.default (not model.id):
# In ~/.hermes/config.yaml, quote the model ID:
model:
default: "minimax-m2.5-free"

The diagnostic checklist (run this before anything else)
If the checklist clears and the 400 persists, the fault is probably upstream of the request - work through the Hermes Agent troubleshooting guide, which orders every failure by the layer it breaks in.

Step 1: hermes config show | head -20 ... What provider and model are configured?
Step 2: hermes chat -q "hello" --oneshot --model anthropic/claude-sonnet-5 ... Does a known-good model work? (Swap in any model your provider currently lists.)
Step 3: Does the error happen on all models or just one? All models = provider/tool issue (Causes 2, 3, 5). One model = model-specific (Causes 1, 4, 6).
Step 4: /new in the chat (there is no hermes chat --new flag) ... Does a fresh session fix it? Yes = corrupted history (Cause 4). No = configuration issue.
Step 5: Check provider dashboard for credits, billing, and model access.
The 400 error is Hermes's most common runtime failure. The log line is always some form of Non-retryable client error. But the cause is one of eight different things, and the fix depends entirely on which one you're dealing with. Diagnose first. Fix second. Don't guess.
If you want an agent platform where provider errors are handled at the platform level and you never see raw HTTP status codes, give BetterClaw a try. Free tier with 1 agent and BYOK. $49/month for Pro with 5 agents. 28+ providers. Smart error handling. The provider compatibility is ours. The agent conversations are yours.
Frequently Asked Questions
What does Hermes Agent Error 400 mean?
HTTP 400 in Hermes means the API provider rejected the request as malformed. The most common causes: model name mismatch (most frequent), a model or route without tool-call support, a legacy Gemini API key, missing reasoning_content for Kimi models, an OpenRouter model that needs a paid plan, oversized request payload, the wrong Gemini endpoint version, or a model ID format mismatch (dots converted to hyphens). The chat message is the same regardless of cause ("rejected the request and retrying won't help"), so diagnosis requires checking which cause applies.
How do I fix "Non-retryable client error" in Hermes?
Run hermes update first (v0.23.2 is current). Then use hermes config show to check your provider and model ID, and test a known-good model: hermes chat -q "hello" --oneshot --model anthropic/claude-sonnet-5 (or any model your provider lists). If that works, your configured model ID is wrong. Run hermes model to pick from the valid list. If all models fail, the issue is provider-level (tool support, an API key, or plan access).
Why does Hermes return 400 on every model with OpenRouter?
GitHub issue #13927 reported every OpenRouter model failing with tools enabled. Maintainers closed it as a provider or model that doesn't accept tool calls, so pick models that support tools on that route. A model that needs a paid plan can also return 400. Running out of credits is different: OpenRouter returns 402 for that, which Hermes retries before failing.
Why does Gemini return "Multiple authentication credentials received" in Hermes?
GitHub issue #7893. Older builds sent an x-goog-api-key header while the OpenAI SDK also sent a Bearer token, and Google rejected the dual credentials. It was fixed in April 2026: update Hermes. If you get API_KEY_INVALID instead, your key is a legacy AIza key that Google stopped accepting in 2026; create a new key in Google AI Studio (current keys start AQ.).
Does BetterClaw have the same 400 error issues?
No. BetterClaw handles provider compatibility at the platform level. Model switching is a dropdown selection. Provider authentication is managed. Tool schema compatibility is tested before deployment. You don't see raw HTTP status codes or debug API-level errors. Free tier with 1 agent and BYOK. $49/month for Pro with 5 agents.




