Your agents never sleep. Control them so you can.
Every tool call and model call your agents make passes through ACP — logged, priced, and with your policies set deterministically. See what they did, stop what you’d never allow, and know why it cost what it did. Works wherever you run agents across every common agent framework and harness.
Prefer fully on-device? Add --local — no account, nothing leaves your machine. what each mode writes →
ACP does three things well.
Autonomous or interactive: allow, ask, or deny on every single action, by rule. The same call gets the same decision every time — no model in the loop deciding your safety.
Watch what your agent actually does, then make it policy. Shadow mode replays your last week of real calls and shows what would have been denied — before anything blocks.
Every run traced action-by-action: each call’s cost, latency, and the decision that gated it. When a run costs $50, you can see exactly why.
Claude Code, Codex, Cursor, opencode, OpenClaw, CrewAI, LangGraph — each ships its own permission system, with its own rules in its own place. ACP is one set of rules, enforced across all of them. See every integration →
You shipped an agent. Now it's a black box.
You wrote checks to keep the agent from doing something dumb. They hold while you babysit it — and the whole point of handing off is that you stop watching. If any of these is you:
Same task, wildly different cost — and you find out from the invoice, not the run.
It reads your .env, pulls customer PII, pipes secrets straight into a prompt — and you can't see what left the building.
Every team ships in a different framework — and there's no one place to see or control them all.
The same task cost 0.9¢ — and $501.
Averages lie. That’s one real coding agent on the same kind of task, across 72 runs — 0.9¢ at the median, $501 in the tail. Cost tracks how much context the orchestration loop re-reads to decide its next move, and that scales with how far the run wanders. One number can't tell you whether an agent is cheap or a runaway waiting to happen.
- Every run as one dot — median, middle-80% band, and the expensive tail
- Failed runs called out — you pay for the ones that never finish, too
- Loop vs leaf: the orchestration loop re-reading context is usually most of the bill
- Budget caps that halt the run, deterministically — not just a warning
What did it actually do all day?
When a run costs $50 or does something you didn't expect, you need to see what actually happened. ACP records every action in order — each model call and tool call with its latency, tokens, cost, and allow-or-deny decision, plus the real identity behind it. Click any event and see exactly what it did.
- The full timeline — loop, leaf, and tool calls, as they happened
- Per-event cost, latency, model, and the policy decision that gated it
- Real identity and scopes carried through every delegation hop
Your agent walked in holding 76 tools.
Every request a coding agent makes declares its full tool catalog — the model can't call what it can't see. A real Claude Code session declares 76 tools: the coding loop, yes — and tools that send messages, publish public web pages, schedule their own future runs, and drive your logged-in browser. Most were never invoked. All of them are standing open.
ACP captures the declared surface on the agent's first call — before anything runs — and turns it into a control table: every tool a click to allow, flag, or deny. When the surface drifts mid-session (we've watched one gain 21 tools in an afternoon), you hear about it.
- The full catalog, visible before first invocation
- One click per tool: allow · flag · deny · ask
- Drift detection — know when the surface grows
Pick the agent you run. The rules come written.
Five ready policies at the level you already think — “an unattended agent that can’t touch the shell.” Assign one in shadow: it replays your last week of real calls and shows what it would have denied, and nothing blocks until you click enforce.
Runs unattended. Works freely — can’t rewrite its own memory, touch the shell, schedule itself, delegate, or read credentials.
Blocks nothing. Records everything — every call in the trail, PII redacted.
Codes beside you. Pauses before file deletes and credential reads; a secret headed out is denied.
Reads the world. Writes only inside its workspace — anything past that waits for your OK.
Ships things. Every deploy and infra change waits for a yes.
What each persona allows, pauses, and denies — attended and unattended →
Your first audit row in thirty seconds.
A coding agent in your IDE, a CrewAI pipeline, a LangGraph service, the OpenAI SDK in a script — pick the method that fits your stack. One install, no code changes, and every tool call is controlled. Free for individuals, forever. $15/seat for teams.
One install per stack — the same control plane behind all of them. See every integration →
Open core, hosted control plane. The enforcement modules are six MIT-licensed npm packages you can read and self-host — the hosted control plane is how you run them in production.
Free for individuals, forever. Per-seat for teams.
Every feature works free for one person — unlimited calls, agents, and clients. Seats start when you add a second person: one price per human, covering every AI tool and agent they run.
- Unlimited calls, agents & clients
- See, control & price every call
- Full audit, cost & policy
- Hardline floor & approvals
- 30-day audit retention
- Everything in Individual, per person
- Unlimited calls & agents per seat
- Shared policies, approvals & roles
- Org-wide audit & cost rollup
- 1-year audit retention
- Volume seat pricing
- SSO / SAML · SCIM
- VPC / on-prem / self-host
- Unlimited audit retention
- SOC 2 evidence exports · DPA / BAA
See full pricing details, FAQ, and examples at /pricing →
Control every tool call your agents make.
Free for individuals, forever — no credit card, no call caps. See your first controlled tool call in about thirty seconds.
Prefer fully on-device? Add --local — no account, nothing leaves your machine.
Already installed? Open your console →
Rolling agents out across a team? See ACP for teams →
Or book 30 minutes — I’ll wire it into your agent with you, live.
— David Crowe, founder
