hermes-ollama-native

Make a local Hermes Agent talk to Ollama's native /api/chat API instead of its OpenAI-compatible /v1 endpoint — so per-request num_ctx is honored and tool calls stream correctly.

Status: an explicit, temporary stopgap until the native adapter lands upstream in NousResearch/hermes-agent#4505. When that merges, you won't need this package.

The problem

When Hermes runs against a local Ollama server, it speaks Ollama's OpenAI-compatible /v1/chat/completions endpoint. That endpoint has two costly limitations:

1. It silently ignores per-request num_ctx. Ollama maps only a fixed allowlist of OpenAI sampling params to model options; num_ctx isn't one of them. So every model loads at Ollama's 4096 context default no matter what context window Hermes computed — regardless of whether you send num_ctx nested in options or top-level. (Verified through Ollama v0.30.11 / main; tracked in Ollama issue #11409.)
2. It can corrupt streamed tool_calls deltas. Multiple users report that the same local model returns broken tool calls via /v1 but correct tool calls via native /api/chat — which makes an agentic assistant unable to act.

Ollama's native /api/chat endpoint honors options.num_ctx and returns correct tool-call deltas. This package routes Hermes' local-Ollama traffic there.

How it works

hermes-ollama-native ships a drop-in, OpenAI-SDK-shaped client (OllamaNativeClient) that POSTs to /api/chat + /api/embed and translates requests, responses, NDJSON streams, tool calls, multimodal content, reasoning, and token usage back to the OpenAI shape Hermes already consumes. It is modeled directly on Hermes' own gemini_native_adapter.py, so api_mode stays chat_completions and the rest of the agent loop is unchanged.

A small runtime shim wraps two of Hermes' client-construction chokepoints and returns the native client only when the target is positively identified as Ollama:

• Hook 1 — main chat: agent.agent_runtime_helpers.create_openai_client (covers the primary client and the per-request sync + stream clients).
• Hook 2 — side tasks: agent.auxiliary_client.resolve_provider_client / resolve_vision_provider_client (compression, title, vision, web-extract). On by default; set HERMES_OLLAMA_NATIVE_AUX=0 to disable just this.

Detection is defensive by construction: it confirms the endpoint is Ollama with a cached GET /api/version probe before routing, so it never mis-routes a non-Ollama "custom" endpoint such as vLLM, llama.cpp, or LM Studio. Everything is wrapped in try/except and degrades to stock /v1 on any error, so it can never crash agent startup.

⚠️ It monkeypatches private Hermes internals. A future Hermes release could rename those functions; the shim is symbol-anchored and fails safely (reverts to /v1) rather than crashing, but native mode would silently stop engaging until this package is updated. This is why the durable home is upstream #4505.

Install

pip install hermes-ollama-native

Requires Python 3.10+ and a Hermes Agent installed in the same environment. The only runtime dependency is httpx.

Quickstart

Native mode only engages when the HERMES_OLLAMA_NATIVE environment variable is truthy (1/true/yes/on). Pick one of two activation styles:

Option 1 — persistent (recommended). Install a one-line startup hook into the current environment so every Hermes process picks it up automatically:

export HERMES_OLLAMA_NATIVE=1
hermes-ollama-native install-pth
hermes            # run Hermes as usual

Option 2 — explicit. Activate from your own launcher/entrypoint before the agent builds its client (no startup hook, nothing global):

import hermes_ollama_native

hermes_ollama_native.activate()

Either way, point Hermes at your local Ollama as a custom endpoint (the standard local-Ollama config — you do not need to strip /v1 yourself; the adapter handles it):

# ~/.hermes/config.yaml
model:
  default: "gemma4:e2b"
  provider: "custom"
  base_url: "http://localhost:11434/v1"

Verify it worked. With a model loaded through Hermes, check the context Ollama actually allocated:

ollama ps      # CONTEXT should be your configured value, not 4096

CLI

hermes-ollama-native status         # show flag + startup-hook state
hermes-ollama-native install-pth     # persist auto-activation in this environment
hermes-ollama-native uninstall-pth   # remove the startup hook

Configuration

Turning it off needs no "deactivate" step: every hook re-checks HERMES_OLLAMA_NATIVE per call, so unsetting it transparently reverts the agent to /v1 (run uninstall-pth too if you installed the startup hook).

Compatibility & limitations

• Targets a Hermes Agent that exposes agent.agent_runtime_helpers.create_openai_client and agent.auxiliary_client.resolve_provider_client / resolve_vision_provider_client (the structure as of Hermes v0.17.x).
• Patches private internals — see the warning above. Pinned/tested against the Hermes releases noted in the changelog.
• Only routes endpoints positively identified as Ollama via GET /api/version. vLLM / llama.cpp / LM Studio (also provider: custom) are left on /v1.

Development

pip install -e ".[dev]"
ruff check . && ruff format --check .
pytest -q

Status & license

This is a standalone, general-purpose adapter; its durable home is the upstream contribution tracked in NousResearch/hermes-agent#4505. Once that lands, this package becomes unnecessary.

Licensed under the MIT License.