Paladin

Deterministic policy enforcement for AI agents. Early access, v0.1.

Identity tells you who an agent is. Paladin is the other half: what it may do, right now, decided before it acts.

Book a call

Run it now

Paladin installs as one command and enforces from the first request. It runs on your machine. Nothing leaves your network.

pip install paladin-ai
paladin up
paladin install-hooks --agent claude

After the third command every Claude Code tool call is checked against your policy before it runs.

Works with Claude Code and OpenAI Codex (every tool call), Cursor (shell commands today), browser agents (navigation and page content), and any other agent behind the forward proxy. Three ways in: hooks where the agent has them, a wrapper launch where it does not, and a forward proxy for anything that cannot cooperate. Your agent's code never changes. Third-party names identify compatibility only and belong to their respective owners.

• See the use cases
• Read the docs
• See the benchmarks
• See pricing

A rule, not a judge

The safeguard built into your agent's harness is a second model with an opinion about each action. Paladin is a policy with an answer. Same request, same decision, every time, with the rule that made it in the log.

• Nothing guesses. There is no model in the decision path. A rule matches or it does not, and the log records which rule decided, so an auditor can check the call.
• Nothing is added to the agent. An allow adds nothing to the agent's context. Zero tokens, no prompt, no second model to pay for or to wait on.
• Nothing gives up. Every action that reaches Paladin is checked, the ten-thousandth like the first. There is no fatigue threshold that hands control back to a person after a run of blocks.

Why a rule and not a judge

Paladin is in early access. It installs today and enforces from the first request. We are working with a small number of teams on real runtime policy problems.

Book a call

Cortexa Labs, Inc.