Getting started

This page is for you if you have just installed Prismor, or are about to, and want to know what it does and how to use it day to day. You do not need to be a security expert.

What Prismor does

AI coding agents such as Claude Code, Codex, Cursor and Copilot run commands, read and write files, and reach out to the internet for you. Prismor sits between the agent and your computer. It checks each of those actions against a set of rules before it runs, and keeps a record of everything the agent did.

Depending on how you set it up, Prismor either just records risky actions or stops them. Everything it records stays on your machine unless you connect it to an organization yourself (see FAQ).

Install

The quickest way is with pip:

pip install prismor

On macOS and Linux you can also use the install script, which picks the right method for your system:

curl -sSL https://prismor.dev/install | sh

On Windows, use pip. Other options (installing from a git checkout, or letting your agent install it through SKILL.md) are covered in Installation.

First run: prismor setup

Run the setup wizard from inside a project folder:

prismor setup

The wizard asks you a few questions:

  • Mode. Whether Prismor should only watch (observe) or also block (enforce). See Observe vs enforce below.
  • What blocks. If you pick enforce, you choose which rules are allowed to block. Nothing is selected at first. Press a to take the recommended set.
  • Agents. Which coding agents on your machine Prismor should watch.
  • Secret cloaking. Whether to hide your API keys and passwords from the agent (see cloak).
  • Unlock password. An optional password you type when you want to let the agent change Prismor's own settings for a few minutes.

When it finishes, check that everything is wired up:

prismor status

status shows whether the hooks are installed, which mode you are in, and the one thing you should do next.

Using Codex? Codex will not run Prismor until you trust its hooks once. Open codex in the project and accept the hook-trust prompt. Until you do, nothing is checked. prismor status tells you if this step is still missing.

Asked your agent to install it? When an agent runs prismor setup, there is no screen for the wizard. Setup prints the questions instead, the agent asks you, then runs the matching command.

Open the dashboard

prismor dashboard

This starts a small web page on your own computer at http://127.0.0.1:7070 and opens it in your browser. Only your machine can reach it. Stop it with Ctrl+C in the terminal.

If the dashboard says no workspaces are registered, Prismor has not seen any agent activity yet. Install hooks in a project (with prismor setup), use your agent for a bit, then click Refresh.

To fill the dashboard with what your agents did before you installed Prismor, replay their saved history:

prismor analyze-history --since 90d

These past sessions are marked as reconstructed. A finding there means a rule would have matched, not that anything was blocked at the time.

A tour of the dashboard

The pages are listed in the sidebar on the left.

  • Overview. The summary screen: what needs attention, how busy your agents have been, which rules fired, where agents connected to on the internet, and where they keep failing. Start here.
  • Sessions. Every agent run, one row each. Open one to see what the agent did step by step. For a live session you can pause Prismor or clear its session scope from here.
  • Files. Every file an agent wrote, and the session that changed it. Use it to answer "which agent edited this, and why?"
  • Agents. Each coding agent Prismor has seen. You can switch one off entirely, or give it its own mode (observe or enforce) instead of the global one.
  • Policy Control. The rules, and whether each one blocks or only reports. You can turn rules on and off and edit the list of websites agents may contact. Some rules are pinned and cannot be turned off.
  • MCP Servers. The MCP servers (add-on tool servers) your agents use, where each is configured, and whether Prismor is checking its calls.
  • Extensions. Skills, plugins, hooks, MCP servers, secrets and packages your agents used. New or changed items are flagged until someone reviews them.
  • Fleet. A preview of the organization-wide view, which is part of Prismor Enterprise. On a personal install it shows no real data.
  • Docs. These documentation pages, searchable, without leaving the dashboard. They work offline.

More detail on each page is in Dashboard.

Observe vs enforce

Prismor has two basic settings.

  • Observe means "watch and write it down." Every action is checked and recorded, and anything risky shows up as a finding. Nothing is stopped. This is a good way to start, because it never gets in your way.
  • Enforce means "watch, write it down, and stop the risky ones." Rules you chose to block will refuse the action before it runs, and the agent is told why.

A good path is to start in observe, look at the dashboard for a few days, then switch to enforce once you know what would be blocked. You can preview a stricter setup without blocking anything:

prismor mode apply dev-safe --observe

Modes explains the three ready-made postures (dev-safe, trusted-workspace, regulated-airgap) and what each one does and does not protect.

Where to go next