Baseten through the Vigil proxy

One row per call: the model id as Baseten returned it, input and output tokens, cache reads and writes where the response reports them, latency, the HTTP status and the provider's error code when it fails, and the agent named in the X-Vigil-Agent header. Likely prompt injections and personal data in prompts are flagged on the Errors page. A cost spike is flagged when a call costs over three times the agent's seven-day median.

Cost is computed from the proxy's rate registry, which prices 2 models on this platform today. A model the registry has no row for is logged with its cost left empty, never estimated.

• moonshotai/Kimi-K3
• zai-org/GLM-5.3

Price per token, by model: Kimi K3.

Every route has the shape https://api.vigil.tools/{user_id}/{provider}/{path}. For Baseten the provider segment is baseten, so the base URL is https://api.vigil.tools/{user_id}/baseten and a request path such as /v1/chat/completions is forwarded unchanged. The proxy stores the origin only, so the path after the segment is whatever your client already uses, /v1 for most OpenAI-compatible APIs.

Two headers: X-Vigil-Key, your Vigil key from the dashboard, and X-Vigil-Agent, the name the dashboard groups this traffic under. Your Baseten key stays where it was, in BASETEN_API_KEY.

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.BASETEN_API_KEY,   // unchanged
  baseURL: "https://api.vigil.tools/{user_id}/baseten/v1",
  defaultHeaders: {
    // MUST be X-Vigil-Key: Authorization already carries your Baseten key.
    "X-Vigil-Key": "vk_your_vigil_key",
    "X-Vigil-Agent": "my-agent",
  },
});

const completion = await client.chat.completions.create({
  model: "<your model id>",
  messages: [{ role: "user", content: "Hello!" }],
});

from openai import OpenAI

client = OpenAI(
    api_key=os.environ["BASETEN_API_KEY"],  # unchanged
    base_url="https://api.vigil.tools/{user_id}/baseten/v1",
    default_headers={
        "X-Vigil-Key": "vk_your_vigil_key",
        "X-Vigil-Agent": "my-agent",
    },
)

completion = client.chat.completions.create(
    model="<your model id>",
    messages=[{"role": "user", "content": "Hello!"}],
)

curl https://api.vigil.tools/{user_id}/baseten/v1/chat/completions \
  -H "content-type: application/json" \
  -H "Authorization: Bearer $BASETEN_API_KEY" \
  -H "X-Vigil-Key: vk_your_vigil_key" \
  -H "X-Vigil-Agent: my-agent" \
  -d '{
    "model": "<your model id>",
    "messages": [{ "role": "user", "content": "Hello!" }]
  }'

Replace {user_id} and vk_your_vigil_key with the values on your Connect page. The snippets above are generated by the same code as that page.

Vigil adds a routing hint so repeat calls are more likely to land on a machine that already holds your prompt. That raises the hit rate; it cannot mark a cache breakpoint, so there is no guaranteed saving and nothing here is reported as one.

With optimisation On: Vigil adds a cache-affinity hint, which can raise the hit rate but cannot place a cache marker. In Shadow, Vigil measures what caching would have saved and changes nothing. Off records the call and nothing more.

Does Vigil see or store my Baseten API key?

No, not the key itself. Your Baseten key passes through the proxy with the request and is never written to disk. A call record can keep a short one-way digest of the credential, so that one key’s cache is kept apart from another’s; it cannot be turned back into the key. Vigil identifies you by the X-Vigil-Key header and your user id in the URL.

Does routing Baseten through Vigil add latency?

Some. The proxy runs on Cloudflare’s edge, so the added hop is short, but reading the request body to place cache markers takes time, and there is a bounded lookup for your account state. The dashboard shows total latency per call. For long calls, stream: Cloudflare’s edge closes non-streaming connections after roughly 125 seconds.

Which Baseten models does Vigil price?

2 models with a published rate in the registry: moonshotai/Kimi-K3, zai-org/GLM-5.3. A call to any other Baseten model is logged and monitored, with its cost left empty rather than guessed.

• LLM proxy setup — base URLs, headers, modes and errors for every provider
• LLM provider differences — LLM provider differences
• prompt caching providers — Prompt caching providers: inject, steer or observe
• LLM token cost — LLM token cost is a function, not a number
• AI token calculator — price a call from the same rate registry