hermes-provider-teamorouter

TeamoRouter model providers for Hermes Agent.

TeamoRouter is an API gateway focused on Claude models, and it also serves GPT, Gemini, DeepSeek, GLM and Grok models behind one API key. Hermes speaks one wire protocol per provider, so this plugin registers two providers on the same key:

Provider	Models	Wire protocol	Endpoint
teamorouter	Claude (claude-*)	Native Anthropic Messages API	https://api.teamorouter.com/v1/messages
teamorouter-openai	GPT, Gemini, DeepSeek, GLM, Grok	OpenAI-compatible Chat Completions	https://api.teamorouter.com/v1

Claude goes over the Anthropic protocol so Hermes applies Anthropic prompt caching (cache breakpoints) and thinking (agent.reasoning_effort) exactly as it does for Anthropic itself.

The plugin registers these two model providers and nothing else: no tools, hooks, middleware, commands, background threads or processes. Hermes core files are not touched.

Maintained by TeamoRouter.

What you get

• --provider teamorouter / --provider teamorouter-openai, both providers in hermes model and /model, and hermes doctor checks, all wired by Hermes from the provider profiles.
• /model pickers built from TeamoRouter's public catalog: Claude models under teamorouter, the other tool-capable chat models under teamorouter-openai. Image models, per-image billing, service-tier and free-tier variants, and models under 64K context are left out. When the catalog is unreachable a short built-in list is shown instead.
• Context windows from the same catalog, so Hermes does not fall back to its generic default for TeamoRouter models.
• Cheap models for Hermes side tasks (compression, titles): claude-haiku-4-5 on teamorouter, gemini-3.5-flash-lite on teamorouter-openai.

Requirements

• Hermes Agent 0.21.5 or newer. The teamorouter provider uses Hermes' Anthropic transport, which needs the anthropic Python package; Hermes installs it on first use (or ships it with its anthropic extra).
• A TeamoRouter API key in TEAMOROUTER_API_KEY. Create an account at teamorouter.com; key setup is described in the API integration guide.
• The plugin itself has no Python dependencies beyond the standard library.

Install

From the Hermes plugin catalog (once listed):

hermes plugins install teamorouter

Or drop the plugin directory into your Hermes home:

git clone https://github.com/2867297362mjb-sketch/hermes-provider-teamorouter.git
mkdir -p ~/.hermes/plugins/model-providers
cp -r hermes-provider-teamorouter/teamorouter ~/.hermes/plugins/model-providers/teamorouter

Use

export TEAMOROUTER_API_KEY="..."
hermes --provider teamorouter --model claude-sonnet-5-5
hermes --provider teamorouter-openai --model gpt-6.1-sol

Or run hermes model, pick TeamoRouter or TeamoRouter (OpenAI-compatible), and choose a model. hermes setup also offers TEAMOROUTER_API_KEY.

If you point model.base_url at your own proxy, the picker comes from Hermes' generic {base_url}/models request instead of TeamoRouter's catalog.

Network calls, credentials and telemetry

• Chat requests go to TeamoRouter through Hermes' own clients with the key from TEAMOROUTER_API_KEY: Claude requests to https://api.teamorouter.com/v1/messages (key in the x-api-key header), other models to https://api.teamorouter.com/v1/chat/completions (key as Authorization: Bearer). Hermes reads the key; the plugin never reads, stores, logs or forwards it.
• Model catalog: the plugin itself makes one kind of request, GET https://api.teamorouter.com/v1/models/pricing, a public endpoint. It is sent without the API key (User-Agent hermes-provider-teamorouter/<version>), cached in memory for an hour, and used only to build the pickers and report context windows. A failed fetch is retried at most every five minutes.
• Source header (teamorouter-openai only): its requests carry a fixed X-Client-Source: hermes-teamorouter-plugin header, which Hermes sends from the provider profile. It holds no user, machine or session data, and TeamoRouter uses it only to count requests that come from this plugin. Hermes does not send profile headers on its Anthropic transport, so Claude requests (teamorouter) carry no header from this plugin; they go out with the Anthropic SDK's own User-Agent (Anthropic/Python <sdk version>), which does not identify Hermes or this plugin.
• No telemetry. The plugin sends no analytics, usage reports or crash data anywhere, and its own code contacts no host other than api.teamorouter.com.
• No files, no processes. The plugin reads and writes no files, stores no credentials and starts no threads or processes.
• Prompts, tool schemas and model output are processed by TeamoRouter and the upstream model vendor serving the chosen model. See TeamoRouter's terms on teamorouter.com.

Development

python3 -m unittest discover -s tests -v

The unit tests use stub providers modules and need neither Hermes nor the network. tests/test_hermes_integration.py runs only when Hermes is importable (a Hermes checkout's interpreter, with PYTHONPATH pointing at it and its anthropic extra installed). It installs the plugin into a temporary HERMES_HOME and drives real Hermes turns against local fake servers, one per wire protocol, so it makes no real network calls either. Check the plugin the way the catalog CI does:

hermes plugins validate ./teamorouter --install-deps

Changelog

0.1.0

First release: teamorouter (Claude, Anthropic Messages) and teamorouter-openai (OpenAI-compatible) providers, catalog-backed pickers and context windows, X-Client-Source header on the OpenAI-compatible provider.

License

MIT. See LICENSE.