One control plane for every agent your team ships
A hook in each teammate's coding agent sends every tool call to one workspace. There it is checked against rules you set once, recorded with who ran it, and priced. Claude Code, Codex, Cursor and the rest land in the same rows. The workspace starts in audit mode: it records first and blocks nothing until you turn rules on.
What changes when it is a team
One engineer with one agent can watch it. A team with eleven agents on four harnesses cannot. Three things go wrong:
Every laptop has its own rules
Each teammate sets up auto mode, allowlists and a sandbox their own way. The team's safety floor is whichever setup is loosest, and nobody can see which one that is.
Nobody can answer “what did our agents do today?”
Logs sit per repo, per harness, per laptop. When a branch is force-pushed or a table is dropped, finding out which agent did it, for whom, means reading transcripts one machine at a time.
The bill has no names on it
Model spend arrives as one number. Which agent? Whose run? Which loop ran all Tuesday night? Nobody knows, so nobody fixes it.
The questions an operator asks first
Why not auto mode plus a sandbox?
Keep both. A sandbox bounds where code runs. It cannot tell a wanted git push --force from an unwanted one, because both use your credentials. Auto mode lives on one laptop and is set by one person. ACP adds the piece a team lacks: one set of rules for every teammate's agent, and one record of what ran. Why they compose →
Does it cover my other harness?
One installer detects Claude Code, Codex, Cursor, Copilot, OpenCode and nine more, plus CrewAI, LangGraph and Deep Agents in code. Coverage is not identical: only Claude Code shows a held call inline. On every other harness the call is blocked, you approve it on the console or from the email, then the agent retries. Each integration page lists what that harness does and does not enforce.
What leaves my machine?
By default the audit record is metadata: tool name, command key, who ran it, the decision, timing, token counts and cost. Command and output text is not stored unless an admin turns on previews (capped at 1,500 characters, PII-redacted where policy says so). A held call keeps up to 500 characters of its arguments so the approver can read them. --local mode sends nothing at all. What is stored →
What if ACP is down?
In Claude Code an interactive session keeps going with a loud UNGOVERNED banner; unattended tiers (subagent, background, API) stay blocked until ACP answers. The shared hook used by the other harnesses keeps going unless you set ACP_FAIL_MODE=closed or ~/.acp/failmode on that machine. Pick the posture you want before the first overnight run.
What your team gets
Bash.git.push to Never for background agents once and it holds on every teammate's machine.curl | sh, or a recursive delete outside the project is held for approval. rm -rf of home or root, mkfs, disk wipes and fork bombs are always blocked. What the floor does and does not catch →claude-acp, codex-acp) and model calls are metered at the gateway, not estimated.This is the actual console
Real views from a live workspace, not mockups. Click any to expand.
How teams roll it out
Sign in to create the workspace and invite the team before anyone installs. A teammate who installs first gets their own personal workspace, and your org is forked. Invite by email, by link, or claim your company domain so new signups join on their own.
The installer finds the harnesses on the machine and adds the hook. Each person gets their own key, so every call arrives under their name. The first tool call writes the first audit row. Everything starts in audit mode: recorded, not blocked.
The setup that holds overnight: on Policies → Agents, set Bash.rm, Bash.git.push, Bash.curl and your database clients to Never for background and subagent. Leave interactive sessions on Ask first. Keep production credentials out of the agent's environment. Then /acp-enforce. The six recipes →
Free covers up to 5 initiating agents (an identity that starts work; its subagents are free), with unlimited calls and 12-month audit retention. Team ($100/mo) covers up to 25 and turns on PII redaction enforcement and custom output patterns. Retention is 12 months on every plan. Team setup, step by step →
Putting a vendor in your call path is a trust decision
So we answer the hard questions in writing. What we store: metadata by default, previews opt-in and redacted. What is open: the harness hooks, plugins and governance SDK modules are MIT-licensed, and with --local they run with no server at all. What is not: the hosted gateway and console (policy engine, approvals, cost X-ray) are not open source; a self-hosted or VPC gateway is available on Enterprise. And our security posture, including what we do not have yet.
Get your team's agents on one panel this week
Start free: up to 5 initiating agents, unlimited calls, subagents free. Or book 30 minutes and we will wire the first agent in with you.