Skip to main content

Connect Codex, opencode and Hermes Agent

What you get: three more coding agents use your company's models through Ambit, with the same personal API key, access rules, budgets and audit log as Claude Code and Cursor.

Every snippet on this page was run against a real Ambit install, with a real model, except the Hermes MCP note, which comes from the vendor page and was not run. The list of what was not tested is at the end of the page.

What was tested​

ToolVersionRequest formatResult
Codex CLI0.159.3OpenAI Responses (/v1/responses)Works. One shell tool call, answer correct.
opencode1.18.34OpenAI Chat Completions (/v1/chat/completions)Works. One file-read tool call, answer correct.
Hermes Agentv0.13.0 (2026.5.7)OpenAI Chat Completions (/v1/chat/completions)Works with the config below. The first config we tried failed; see Hermes Agent.

Test setup: Ambit from the Quick start on kind, the model nvidia/nemotron-3-super-120b-a12b:free at OpenRouter, a personal key with the scope LLM and MCP, on 2026-10-14. Only this one model was tried. Each tool made one prompt that needs one tool call, and every successful call is in the Audit log with status 200.

Before you start​

You need the cluster from the Quick start, the model, a rule that allows it, and a personal API key (see Connect Claude Code and Cursor, step 1). The addresses below are the quick-start addresses. On your own cluster, use your own.

  • LLM address: http://llm.localtest.me:8080
  • MCP address: http://mcp.localtest.me:8080/mcp
  • Model name for clients: nvidia/nemotron-3-super-120b-a12b:free. Clients send the provider's model name, not the name of the Model object.

Codex CLI​

Codex speaks only the OpenAI Responses format. Ambit serves /v1/responses.

Add this to ~/.codex/config.toml:

model = "nvidia/nemotron-3-super-120b-a12b:free"
model_provider = "ambit"

[model_providers.ambit]
name = "Ambit"
base_url = "http://llm.localtest.me:8080/v1"
env_key = "AMBIT_API_KEY"
wire_api = "responses"

Put the key in the environment and run a prompt:

export AMBIT_API_KEY=ambit_...
codex exec --skip-git-repo-check --sandbox read-only "Run the shell command 'cat note.txt' and tell me what it prints."

Codex ran the command and printed the file. Notes from the test:

  • Codex prints Model metadata for ... not found. Defaulting to fallback metadata. It does not know your model. The warning is harmless.
  • The model is not an OpenAI model, and the call still worked. Ambit sent the Responses request on to OpenRouter and OpenRouter answered. We did not check whether Ambit or OpenRouter converts the format, so we cannot say that every provider works this way. Try your own provider with one prompt before you give Codex to your team.
  • When codex exec runs from a script, it can wait for more input on standard input. Add < /dev/null to the command. We saw one run stop without output until we did this, and cannot say for sure that standard input was the cause.

MCP. Add the Ambit MCP gateway with the key as a static bearer token:

[mcp_servers.ambit]
url = "http://mcp.localtest.me:8080/mcp"
bearer_token_env_var = "AMBIT_API_KEY"

codex mcp list shows ambit ... enabled ... Bearer token. In the test, Codex called time_get_current_time (the time server of the quick start, step 9) and returned the time. Codex then asked the gateway for MCP resources and got Method not found. The gateway serves tools only. Codex shows the error and goes on.

opencode​

opencode needs a custom provider that uses the OpenAI-compatible package. Put this in opencode.json in your project, or in ~/.config/opencode/opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "model": "ambit/nvidia/nemotron-3-super-120b-a12b:free",
  "provider": {
    "ambit": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Ambit",
      "options": {
        "baseURL": "http://llm.localtest.me:8080/v1",
        "apiKey": "{env:AMBIT_API_KEY}"
      },
      "models": {
        "nvidia/nemotron-3-super-120b-a12b:free": { "name": "nvidia/nemotron-3-super-120b-a12b:free" }
      }
    }
  }
}
export AMBIT_API_KEY=ambit_...
opencode run "Use the read tool to read note.txt and tell me what it says."

The first start in a new data directory did not answer for more than 5 minutes in our test, and we stopped it. The second start answered normally. We did not find the cause. If the first start seems stuck, stop it and start again.

MCP. Add a remote server with the key as a header:

{
  "mcp": {
    "ambit": {
      "type": "remote",
      "url": "http://mcp.localtest.me:8080/mcp",
      "headers": { "Authorization": "Bearer {env:AMBIT_API_KEY}" },
      "oauth": false
    }
  }
}

opencode mcp list shows ambit connected, and the model called ambit_time_get_current_time. The same configuration without "oauth": false also worked, because the gateway accepts the static header and never asks for OAuth. We keep "oauth": false, so that opencode does not start an OAuth sign-in if a gateway ever answers 401.

Hermes Agent​

Hermes Agent (Nous Research) uses the Chat Completions format with a custom provider. Put this in ~/.hermes/config.yaml:

model:
  default: nvidia/nemotron-3-super-120b-a12b:free
  provider: custom:ambit
custom_providers:
  - name: ambit
    base_url: http://llm.localtest.me:8080/v1
    key_env: AMBIT_API_KEY
    api_mode: chat_completions
export AMBIT_API_KEY=ambit_...
hermes chat -Q -t file -q "Use a tool to read the file note.txt in the current directory and tell me what it says."

What failed first. Our first config put provider: custom, base_url and key_env: AMBIT_API_KEY directly in the model: block. Hermes ran, sent no key, and ended with Error: authorization failed. The gateway answered 403. In Hermes, key_env works inside custom_providers not in the model: block. The config above is the one that worked. The other way, api_key: in the model: block, puts the key itself into the file, and we did not test it.

We did not test Hermes with MCP. Hermes reads MCP servers from mcp_servers in the same file, with url and headers; see the Hermes MCP guide. Treat that as not tested by Ambit yet.

Not tested​

  • Hermes with MCP, and Hermes with the anthropic_messages format.
  • A real Claude model behind /v1/messages.
  • Codex with a model of an OpenAI-type provider. Only an OpenRouter model was used.
  • Claude Code without ANTHROPIC_SMALL_FAST_MODEL or the ANTHROPIC_DEFAULT_*_MODEL variables.
  • Any model other than nvidia/nemotron-3-super-120b-a12b:free.

Check it in the console​

Open Audit and search for your email address. Every call of the three tools is listed with the model, the token counts and the key that made it. Tool calls through the MCP gateway are listed with the tool name.