Bring Your Own LLM

Overview

A voice agent normally answers with the model in its voice_model setting. With bring your own LLM you register your own OpenAI-compatible chat completions endpoint, point an agent at it with custom_llm_id, and every reply on that agent's calls is generated by your model. CallMissed still runs the call: speech recognition, turn-taking, the voice, tools, transcripts and phone carriage.

Your endpoint receives the conversation as a standard streamed chat completions request, tool definitions included, and streams its reply back. Your model's tokens are yours: CallMissed does not bill them. A call that runs on your own LLM pays a per-minute platform fee in their place; speech recognition, voice and phone minutes bill as they do on any call.

Authentication

Every endpoint on this page takes Authorization: Bearer cm_your_api_key. Reading endpoints needs the bots:read scope; creating, changing, deleting and testing needs bots:write. A key without the scope gets 403.

The request we send

For each reply, CallMissed sends one POST to {base_url}/chat/completions:

POST https://llm.example.com/v1/chat/completions
Content-Type: application/json
Accept: text/event-stream, application/json
Authorization: Bearer your-model-key

{
  "model": "your-model-id",
  "stream": true,
  "stream_options": { "include_usage": true },
  "messages": [
    { "role": "system", "content": "You are the front desk for..." },
    { "role": "user", "content": "Hi, I want to book a table for two." }
  ],
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "check_availability",
        "description": "...",
        "parameters": { "type": "object", "properties": { "...": {} } }
      }
    }
  ]
}

• model is always the model you registered, whatever the agent sends.
• The auth header is the header_name and header_value you registered (Authorization by default). It is the only credential your endpoint sees.
• tools is present when the agent has tools. Answer a tool call the OpenAI way (delta.tool_calls); the agent runs the tool and sends you the result on the next request.
• temperature is sent when the agent sets llm_temperature.

Your endpoint must reply 200 with Content-Type: text/event-stream and Server-Sent Events chat completion chunks, ending with data: [DONE]:

data: {"id":"c1","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"role":"assistant","content":"Sure, "}}]}

data: {"id":"c1","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"content":"for what time?"}}]}

data: [DONE]

Keep replies short and conversational: they are spoken aloud on a live call, and the caller hears nothing until your first chunk arrives.

Timeouts and failures

timeout_seconds (1 to 10, default 10) is how long we wait for your endpoint to connect and for each chunk of the reply, the first one included. A reply that stops streaming for longer is cut off there. A single reply is also capped in total length and duration.

Any non-200 status, a timeout, a redirect or a connection failure counts as a failed reply. What happens next depends on platform_fallback:

The status and body your endpoint returned are not passed back to the caller or stored. Use POST /api/v1/custom-llms/{id}/test to check an endpoint.

Address rules

base_url must be https:// with a public host name or address. It cannot carry a username, password, query string or fragment; a pasted full .../chat/completions URL is trimmed to its base. A host that resolves to a private, loopback, link-local or cloud metadata address is refused when you save it, and the address is checked again on every request, so a DNS change cannot point the agent at a private network. Redirects are not followed.

Endpoints

Fields

Responses carry key_last4 (the last four characters of header_value, or null when the value is shorter than 12 characters) in place of the value, plus id, created_at and updated_at. An account can register up to 50 endpoints.

Register an endpoint

curl -X POST https://api.callmissed.com/api/v1/custom-llms \
  -H "Authorization: Bearer cm_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Support model",
    "base_url": "https://llm.example.com/v1",
    "model": "support-v3",
    "header_name": "Authorization",
    "header_value": "Bearer your-model-key",
    "timeout_seconds": 8,
    "platform_fallback": true
  }'

201 returns the endpoint:

{
  "id": "6f1c2a0e-6c8f-4a51-9d0e-2f7e1b9c4d11",
  "bot_id": null,
  "name": "Support model",
  "base_url": "https://llm.example.com/v1",
  "model": "support-v3",
  "header_name": "Authorization",
  "key_last4": "-key",
  "timeout_seconds": 8,
  "platform_fallback": true,
  "enabled": true,
  "created_at": "2026-10-08T09:30:00Z",
  "updated_at": "2026-10-08T09:30:00Z"
}

List, get, update, delete

# List (optional ?bot_id=, limit 1-100, offset)
curl https://api.callmissed.com/api/v1/custom-llms \
  -H "Authorization: Bearer cm_your_api_key"

# Get one
curl https://api.callmissed.com/api/v1/custom-llms/{id} \
  -H "Authorization: Bearer cm_your_api_key"

# Update: send only what changes. "header_value": null removes the stored header.
curl -X PATCH https://api.callmissed.com/api/v1/custom-llms/{id} \
  -H "Authorization: Bearer cm_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{"model": "support-v4", "header_value": "Bearer rotated-key"}'

# Delete
curl -X DELETE https://api.callmissed.com/api/v1/custom-llms/{id} \
  -H "Authorization: Bearer cm_your_api_key"

A delete returns 204, or 409 while an agent's custom_llm_id still points at the endpoint; the message names those agents. An unknown id, or one in another account, is 404.

Test an endpoint

curl -X POST https://api.callmissed.com/api/v1/custom-llms/{id}/test \
  -H "Authorization: Bearer cm_your_api_key"

Sends one short streamed request ("Reply with the single word OK.") exactly as a call would, and reports whether a well-formed chat completion chunk came back. The reply's text is not returned or stored.

{ "ok": true, "status_code": 200, "latency_ms": 412, "error": null }

On failure ok is false, status_code is your endpoint's status (or null when it could not be reached) and error says what went wrong in general terms, for example "The endpoint returned HTTP 401".

Point an agent at it

Set custom_llm_id in the agent's config:

curl -X PATCH https://api.callmissed.com/api/v1/bots/{bot_id}/config \
  -H "Authorization: Bearer cm_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{"values": {"custom_llm_id": "6f1c2a0e-6c8f-4a51-9d0e-2f7e1b9c4d11"}}'

To go back to the agent's own model, unset it: {"unset": ["custom_llm_id"]}.

custom_llm_id cannot be combined with:

• a voice_tier: a tier runs its own model at a flat per-minute price;
• a speech-to-speech voice_model (such as the realtime or managed voice models), which listens and speaks itself and so has no separate reply step.

Either combination is refused with 422 and a message saying which setting to change. The agent's voice_model stays relevant as the fallback model when platform_fallback is on.

Each call starts with your endpoint's current settings, so a changed model, header or timeout applies from the next call. An endpoint pinned to one agent with bot_id cannot serve any other agent's calls.