Golem documentation | Golem docs

Install Golem, connect your agents to it, decide what they may do, and check the signed record of what they did.

Golem sits in the path of every tool call an AI agent makes. It checks each call against your rules, lets it through, refuses it or holds it for a person, and signs a record of it that nobody can edit afterward, Golem included. These pages cover how to install it, how to connect each agent, how to write the rules, and how to check the record yourself.

How the pieces fit#

The Golem agent is a small background service on each machine. Your coding agents talk to their tools through it: their hooks call it before a tool runs, and their MCP servers sit behind it. It makes the allow, refuse or hold decision on the machine, from rules compiled ahead of time, so nothing waits on a network round trip or a model.

flowchart TD
    A["Coding agent<br/>Claude Code, Cursor, Codex…"] -- "hook or MCP call" --> G["Golem agent<br/>on the same machine"]
    G -- "allowed" --> T["The tool"]
    G -. "refused, or held for a person" .-> A
    G -- "raw payload" --> B[("Your bucket, or<br/>Golem's under your key")]
    G -- "hash and signature" --> L[("Signed record")]
    L --> C["Golem console"]
    B -. "read only if you allow it" .-> C

Each call's full content goes to storage you choose. The record holds its hash, its signature and the facts about it (which agent, which tool, when, and what was decided), chained to the entry before it so that changing or removing any entry shows.

Where to start#

Trying it on one laptop. Install Golem, then open the page for your agent: Claude Code, Cursor, Codex, GitHub Copilot, Gemini CLI, Devin Desktop, Antigravity, OpenClaw, Hermes, or the browser extension for chat sites.
Rolling it out to a team. Roll out to a fleet covers Jamf, Intune and Kandji, and watching before you enforce.
Deciding what agents may do. Rules and CEL, then Block, hold and contain for what happens when one fires. Redaction and the shell jail and sandbox protect what a rule can't.
Investigating. Coverage shows what Golem sees, Session review how its reviewers read sessions, and Cases and evidence export how an investigation ends in a package an auditor can check.
Deciding whether to trust it. The signed record, Where your data goes, People, agents and access, and Limits, which says plainly what Golem does not do.

Words these pages use#

Term	What it means
Golem agent	The service on each machine that sees, decides on and signs every call.
Agent	The AI being watched: Claude Code, Cursor, Codex, a browser chat, and so on.
Workspace	Your organization in Golem. Machines enroll into it with a workspace key.
Session	One conversation or task, with its own chain of signed entries.
Rule	A condition on a call, and what to do when it matches: watch, warn, ask or block.
Flag	Something a rule, a reviewer or a person marked as worth a look.
Case	Flags and evidence gathered into one investigation, with an owner and a clock.

Getting help#

Email info@golem.team. Every page here is also served as Markdown: add .md to the address, or read llms.txt for the whole list.