# Manual setup — Python (/docs/get-started/manual-python)
Add TokenPolice to a Python app in three steps. You'll need your
[API key and the collector URL](/docs/get-started/api-key) first.
## 1. Install [#1-install]
```bash
pip install token-police
```
Python 3.10 or newer. Provider libraries (OpenAI, Anthropic, …) are auto-instrumented — no
extra install for the common ones. Framework extras are on the
[integrations](/docs/integrations/matrix) page.
## 2. Initialize [#2-initialize]
Call `tp.init()` **once at startup, before you import or construct your LLM client.** That
ordering lets TokenPolice wrap the client automatically.
```python
import os
import token_police as tp
tp.init(
api_key=os.environ["TOKENPOLICE_API_KEY"], # tp_sk_...
base_url="https://collect.tokenpolice.ai", # the collector — always set this
firewall="dry_run", # watch first; flip to "enforce" later
)
# Only now import and construct your LLM client:
from openai import OpenAI
client = OpenAI()
```
That's the whole integration. Every call `client` makes is now checked before it runs and
logged after — you'll see it in the dashboard.
`firewall="dry_run"` is the default and the right place to start: TokenPolice runs the
full check and records what it *would* do, but never blocks. Flip it to `"enforce"` once
your dashboard rules look right. See [Dry-run vs enforce](/docs/concepts/dry-run-vs-enforce).
## 3. Attach identity (recommended) [#3-attach-identity-recommended]
To budget per user or per session, wrap your entrypoint with `@tp.workflow`. It tags every
call inside with one identity and threads them into a single trace:
```python
@tp.workflow(name="support_agent")
def run_agent(user_id: str, paid_plan: str, session_id: str, query: str):
return client.chat.completions.create(
model="gpt-4o-mini",
messages=[{"role": "user", "content": query}],
)
```
By default `@tp.workflow` reads `user_id`, `paid_plan`, and `session_id` straight from the
wrapped function's arguments — so calling `run_agent(user_id="u_123", paid_plan="pro",
session_id="chat_42", query=...)` attaches them automatically. These identities are what
budgets and rules match on. More in [Identity](/docs/concepts/identity).
## 4. Run it [#4-run-it]
Run your app as usual and make an LLM call. Then open
[app.tokenpolice.ai](https://app.tokenpolice.ai) — the call shows up with its token counts
and cost. Only token metadata is sent; your prompts and completions never leave your process
([what's in that record](/docs/concepts/data-privacy)).
## Then finish the job [#then-finish-the-job]
The three steps above get you metering. A production integration also names its steps, wraps
its tool calls, flushes before a serverless function freezes, and handles a blocked call —
and then proves all of it works. The
[integration checklist](/docs/get-started/integration-checklist) walks through the rest in
order.
## Next steps [#next-steps]