Quickstart
Get your first traces flowing in about five minutes. Python and TypeScript SDKs, no agents or sidecars to run.
1. Install the SDK
Add the package to your project. The SDK is dependency-light and has no native build step.
$ pip install vectornpm install @vector/sdk. The API mirrors the Python examples below.2. Set your API key
Grab a key from the console under Settings → API keys, then export it. The SDK reads it automatically.
$ export VECTOR_API_KEY=vk_live_...
3. Wrap your client
Pass your existing model client to trace(). Every call made through the wrapped client is captured automatically, with zero changes to your call sites.
from vector import trace from openai import OpenAI client = trace(OpenAI()) resp = client.chat.completions.create( model="gpt-4o", messages=[{"role": "user", "content": "hi"}], ) # this call is now traced, end to end
Open the console and your first trace is already there, broken into spans for retrieval, prompt assembly, the model call, and any tools.
SDK reference — evals
Attach scored checks to a prompt so a regression fails loudly instead of shipping quietly.
Registers an eval that runs on every new prompt version.
| Argument | Description |
|---|---|
| name | Human label shown in the console. |
| fn | Callable returning a score between 0 and 1. |
| dataset | Optional named set of test cases to run against. |
| threshold | Score below this marks the version as failing. |
SDK reference — prompts
Pull a versioned prompt from Vector instead of hardcoding it, so edits and rollbacks happen without a deploy.
Fetches a managed prompt and fills its variables.
| Argument | Description |
|---|---|
| slug | The prompt's unique handle, e.g. support-reply. |
| version | A pinned version number, or latest. |
| variables | Values to interpolate into the template. |
Framework integrations
First-class wrappers ship for the tools you already use:
- OpenAI and Azure OpenAI
- Anthropic
- LangChain and LlamaIndex
- Vercel AI SDK
- Any HTTP client, via the manual span API