Brokk Observe

Every agent session, on the record.

Brokk Observe is an audit trail for AI coding agents. It records every Claude Code and OpenAI Codex session in your organization, reviews each one for security problems, and traces bugs back to the session that introduced them.

Works with Claude Code and OpenAI Codex, as a plugin. Secrets are redacted before anything is stored.

Eli Novak in acme/infra
Claude Code session, 39 minutes, 28 tool calls
Risk 100
  1. PromptBump the prod RDS instance class and plan it
  2. Readenvs/prod/main.tfok
  3. Editmodules/vpc/outputs.tfok
  4. Bashterraform validateok
  5. Bashgit push --force origin maindenied
Critical Force-push to a protected branch

A policy denied the call before it ran. The rule bash.force_push_protected filed it as a finding, pinned to the event that tried it.

A session from the demo data, as the dashboard replays it.

What you get

One dashboard for what your agents did, what went wrong, and what it cost you later.

Every session, replayable

Every prompt, tool call, permission decision, subagent, commit and pull request, in order and with its diffs. Secrets are redacted before anything is stored.

A session page: who ran it, its cost and risk, the pull requests and issues it is linked to, and the timeline of tool calls. A session page: who ran it, its cost and risk, the pull requests and issues it is linked to, and the timeline of tool calls.

Rules for the known problems, an LLM review for the rest

About 35 detectors catch secrets reaching the model, curl | sh, force-pushes and prompt injection in tool output. A Claude-powered review then reads the whole session for what rules cannot see.

The findings queue: open findings by severity, each with the rule or review that raised it, its evidence and triage actions. The findings queue: open findings by severity, each with the rule or review that raised it, its evidence and triage actions.

Stop a tool call before it runs

Deny, ask or warn rules, set centrally and enforced by the plugin before the call happens. Try a rule against a command before you switch it on.

The policies page: deny, ask and warn rules with the patterns they match, and a form for adding one. The policies page: deny, ask and warn rules with the patterns they match, and a form for adding one.

Output, with the downside in the next column

People, agents and models ranked on bugs fixed and pull requests merged and reviewed. Bugs introduced and change-failure rate sit right beside them.

The leaderboard: people ranked by bugs fixed and pull requests merged, with bugs introduced and change-failure rate alongside. The leaderboard: people ranked by bugs fixed and pull requests merged, with bugs introduced and change-failure rate alongside.

Bugs traced to the session that introduced them

When a bug is fixed, Observe blames the fix back to the commit that introduced it, and to the session that wrote that commit. You get a change-failure rate for AI-written code that you can check line by line.

The bugs page: how many bugs trace to an agent session, and each bug with the sessions blamed for it and the evidence. The bugs page: how many bugs trace to an agent session, and each bug with the sessions blamed for it and the evidence.

How it works

  1. Install the plugin

    Two commands in Claude Code or Codex, or roll it out to the whole organization with managed settings so it cannot be switched off.

    /plugin marketplace add BrokkAi/brokk-observe-plugin
    /plugin install brokk-observe@brokk
  2. Sessions stream in as they happen

    The plugin sends each prompt, tool call and commit as it happens, and keeps them in a local queue while a machine is offline. Secrets are redacted before anything is stored.

  3. Review them in the dashboard

    Triage findings, replay any session, and see which sessions shipped pull requests and which introduced bugs. Critical findings can go to Slack as they happen.

Get started

Brokk Observe is in early access. Leave your work email and we will get in touch about getting your team set up.

The plugin is public: read the docs and the source on GitHub. Already have an account? Sign in.