Blog · Product

Decision Horizon for Claude Code, Cursor, Copilot and every coding agent

Coding agents act through built-in tools that no MCP proxy sees. One command now puts a policy decision, an Ostiary score and a signed record on every one of their tool calls.

Most of the AI agents touching real code today aren't custom agents. They're coding agents: Claude Code, Cursor, GitHub Copilot, Codex, Gemini CLI and a few others, running on developers' machines with the developers' credentials. They read .env files, run shell commands, edit anything in the repository and fetch from the web.

Until today, Decision Horizon could only see part of that. Our MCP proxy records the MCP servers an agent calls, but a coding agent's most powerful tools are built in. Its shell, its file editor and its web fetch never pass through MCP. So the most consequential things these agents do were exactly the things we couldn't record.

One command per agent

Almost every coding agent now has a hook system: a command it runs before and after each tool call, which can allow or block the call. That's where Decision Horizon now sits.

pip install decision-horizon
dh hooks install claude-code
dh hooks install cursor copilot gemini

That writes Decision Horizon into each agent's own hook configuration, keeping whatever hooks you already had. From then on, every tool call the agent makes, built-in or MCP, goes through the same steps as everywhere else in Decision Horizon. We record what the agent had read, apply your policy, let Ostiary score the call against what's normal for that agent, and sign an intent record. Then we answer the agent in its own format: allow, deny, or ask. After the tool runs, the outcome is signed too.

It works with Claude Code, Cursor, GitHub Copilot (the CLI, the cloud coding agent and VS Code agent mode), OpenAI Codex CLI, Gemini CLI, Windsurf, Cline, Kiro, Factory, Continue, Augment, OpenCode and Amp. If you're building your own agent on the Claude Agent SDK, the same engine runs in-process, in Python and TypeScript.

What it catches

Here's a session we run in our tests. The agent reads .env, edits a source file, then tries to curl the API key to a paste site. The edit goes through and is recorded. The curl is paused, because the value it carries came from .env, whatever the command looks like. In Claude Code and Copilot, that pause is the agent's own approval prompt, with our reason in it. In agents without a prompt, it's a deny.

The agent also can't turn us off. Deleting the records, reading the signing key, or editing .claude/settings.json to remove the hook are all denied, and those attempts are recorded as well.

And the secret itself never lands in a record. Arguments are hashed, not stored.

What it doesn't do

A hook runs beside the agent, from configuration its user controls. Someone with full control of the machine can switch it off, and calls made while it's off aren't recorded. Every record says it came from a hook, so nobody mistakes it for something stronger. We also only know what the agent reports to its hook, and a shell command's targets are the paths it spells out. These limits are written into the docs and into the records themselves.

The docs list where each agent's hook goes and what it can do. If your team runs coding agents against anything that matters, we'd like to show you a session record from one of your own repositories.

Want to see a record of one of your own agent workflows? We'll replay one with you.

Book a demo
← Ostiary: a doorkeeper for every agent action