# What Bridle does

Bridle sits between your AI agents and the model providers they call. You change a base URL and add a key. From that point on, every call is measured, capped, logged, and checked, and the risky ones wait for a human.

## Six capabilities

1. Meter every call: model, token counts, real dollar cost, and latency, grouped into runs. Live spend, not an invoice surprise.
2. Cap spend and kill runaways: a dollar budget per run; when it is crossed the run is terminated and further calls get HTTP 402.
3. Hold risky actions for a human: payments, deletes, outbound email pause for named approval, with a Slack ping.
4. Block malicious inputs: scoring for prompt injection, jailbreaks, secret leakage, and exfiltration, including spacing and base64 evasion; in block mode dangerous requests never reach the provider.
5. Keep tamper-evident logs: every event hash chained to the previous one, so edits or deletions are provable.
6. Export compliance evidence: one click per agent, per period, mapped to EU AI Act Articles 12, 14, 15, 26, and 73.

## Works with the agents you already have

Bridle is provider compatible. If your code speaks the OpenAI, Anthropic, or Gemini API, it already speaks to Bridle. No SDK to adopt, no framework to migrate to. Your provider key passes through and is never stored.

```python
client = OpenAI(
  base_url="https://api.agentbridle.com/gateway/openai/v1",
  default_headers={"X-Bridle-Key": "bk_...", "X-Bridle-Run-Id": "nightly-invoices-042"},
)
```

Ten minutes to route your first agent. Free for your first agent during the beta. Join: https://agentbridle.com/#waitlist · contact@agentbridle.com