Operate the agents you already shipped.
See every agent. Fix failures, control costs, and make changes without redeploying.
Sound: off
acp / control plane
Demo
1 running · 1 idle · 1 warning · 1 error · 1 paused
Your app exchanges telemetry with ACP, which sends commands to connected agents. checkout-recovery: Error support-router: Running invoice-parser: Idle lead-enricher: Warning digest-writer: Paused
Projected daily cost
$416.00
agent checkout-recovery
status error
model gpt-4.1-mini
prompt v12
Observe
Pause
Swap model
Push prompt
From the first signal to the next safe run.
Find the failing agent, intervene, and confirm the result.
Incident / cost anomaly
03:07:04
01
02
03
04
Spend accelerating
checkout-recovery · agt_84c1
anomaly
Retries
21
Cause
unresolved
Next step
inspect
03:07:49 POST /api/agents/:id/commands · pause sent
03:08:04 webhook acknowledged · x-acp-signature ok ok
03:08:34 next heartbeat · status paused verified
01
Detect
See the spike.
Runs, token use, cost, errors, and health put the problem on one screen.
Observability tools already do this part well.
02
Diagnose
Find the agent.
Separate the misbehaving agent from the rest of the app before changing anything.
No codebase hunt. No guessing which deploy is live.
03
Intervene
Pause it now.
Send the command to the integration already running inside the target app.
The local wrapper checks state before the next LLM call.
04
Verify
Watch the line flatten.
The next heartbeat confirms the new state. Production stays up; the model calls stop.
No prompt edit. No build queue. No redeploy.
Your app. Your agents. One connection.
Create an app, add the SDK, and receive the first heartbeat.
01
Create the app Receive a one-time API key and webhook secret.
02
Install the SDK One file, one route, one wrapper around existing calls.
03
See it connect The first heartbeat marks the app connected.
Connection lab / production API
credentials ready
Generated environment variables
Copy
ACP_URLhttps://agentcontrolpanel.si public
ACP_API_KEYacp_•••••••••••••••••••••••• one time
ACP_AGENT_IDagt_84c1 per agent
ACP_WEBHOOK_SECRET•••••••••••••••••••••••••••• secret
Keys are generated by ACP. Store them in the target app's environment.
acp.ts · sdk
long form
Copy
First heartbeat / live connection
Copy
POST /api/heartbeat
Content-Type: application/json
{
"api_key" : "<ACP_API_KEY>" ,
"agents" : [{
"agent_id" : "<agent-id>" ,
"status" : "idle" ,
"tokens_since_last" : 0,
"runs_since_last" : 0,
"errors_since_last" : 0
}]
}
Heartbeat received · app connected
Two directions, one wire.
Your app reports activity. ACP sends signed commands. Model traffic stays in your app.
Upward · your app reports
Heartbeat every 30s
POST /api/heartbeat
POST /api/report
GET /api/prompts/:agentId
Status, tokens, runs and errors since the last beat. Run reports carry model used, latency and prompt version.
The control plane
ACP holds the state
commands pause · resume
commands stop · restart
commands update_model
commands update_prompt
Six commands, one audit trail. Prompt pushes are approval-gated before they reach the app.
Downward · signed webhook
HMAC, timestamp, nonce
header x-acp-signature
header x-acp-timestamp
header x-acp-nonce
Signature over ts.nonce.body, a 300-second freshness window, and a replay cache that rejects a reused nonce.
See what happened. Decide what happens next.
Five surfaces, one control plane: send commands, read failing runs, account for tokens, score responses, and turn a repeated failure into a deployed skill.
Remote command
Interactive simulation
checkout-recovery agt_84c1 · production
running
Pause agent signed webhook v2
command pause
target agt_84c1
deploy not required
Changes apply in the target app.
Pause agent
✓
Ready to send
The acknowledgement will appear here. This demo does not contact a real agent.
app.acp / runs 1,284 runs today · 11 failed
checkout-recovery 03:07:41 · 9.1s
18.4k
support-router 03:06:12 · 2.2s
3.1k
invoice-parser 03:05:58 · 3.6s
1.1k
checkout-recovery 03:04:20 · 2.4s
1.8k
run_8f21c4 checkout-recovery · gpt-4.1-mini · 6 tool calls
failed
Error
ToolTimeout: stripe.refund did not return within 30000ms
Why it failed
A Stripe timeout triggered six retries and used 18.4k tokens.
How to fix it
01 Raise the tool timeout to 60s and send an idempotency key so a retry cannot double-refund.
02 Cap in-run retries at two, then fail fast instead of re-planning.
03 Push prompt v13, which tells the agent to stop after a tool error and hand off.
Retry run
Open payload
Send to Skills Hub
app.acp / token usage 7 days 30 days 90 days
Input vs output per day
input output
Estimated cost
$284.10
Agent App Model
checkout-recovery orchestrator · 41%
2.6M
support-router router · 24%
1.5M
invoice-parser tool · 19%
1.2M
digest-writer agent · 11%
0.7M
lead-enricher worker · 5%
0.4M
checkout-suite 14 agents · 52%
3.3M
support-desk 9 agents · 26%
1.7M
billing-ops 6 agents · 15%
0.9M
growth-loops 2 agents · 7%
0.5M
gpt-4.1-mini 63% of tokens
4.0M
claude-sonnet-4-6 29% of tokens
1.9M
gpt-4.1 8% of tokens
0.5M
app.acp / feedback run_8f21c4 · awaiting review
Agent response
“I was unable to process the refund. Please contact support during business hours and reference order 41822.”
Score the run, or let the reviewer model score it. Low scores collect against the agent and become the evidence for a new skill.
Reviewer model score
auto
Resolution 34
Policy 48
Tone 88
Overall 57
3 of the last 20 runs scored under 60 on resolution.
Draft a skill
app.acp / skills hub · testing lab skill_refund_timeout draft
01 Failure detected 11 failed runs, same tool timeout
02 Skill drafted retry policy + handoff wording
03 Tested in the lab 40 replayed runs, agent A/B
04 Deployed on pass only if the overall score clears the gate
Before / after · 40 replayed runs
Run the lab test
Tokens per run 18.4k → 4.2k
Review the results before applyingmeasured: pending
not deployed
01 Surface 01 of 05
Control
Pause, restart, swap the model or push a prompt. The command lands in the integration inside your app.
ACP never routes your LLM traffic. Local state is checked before the next call, so nothing waits on a deploy. Pick a command to send it in the panel.
Pause or resume
Stop model calls without stopping the target app.
Swap the model
Apply a local override before the next request.
Push a system prompt
Approval-gated, then fetched and cached without a deploy.
Stop or restart
Change one agent's operating state remotely.
02 Surface 02 of 05
Runs
Open the failing run, read the reason it failed, and get the fix in the same view.
Every run report carries model, latency, tokens, tool calls and the error. Failures are grouped so a repeated cause is obvious instead of buried in a log.
Filter to failures, one agent, or one app.Read the error, the explanation of what the agent did, and the steps to fix it.Act from the run: retry it, push a prompt, or send it to the Skills Hub.
03 Surface 03 of 05
Token usage
Account for spend per agent, per app, per model, and per day.
Input and output are counted separately, because the fix for a long prompt is not the fix for a long answer. Compare the last 7, 30 or 90 days.
Split input against output on every agent and every day.Group by agent, app or model to find the one line that moved.Compare 7, 30 and 90 day windows against the run count behind them.
04 Surface 04 of 05
Feedback
Score a response by hand, or let a reviewer model score it. Both teach the agent.
Feedback attaches to the run, not to a spreadsheet. When scores for one agent keep landing low, the same evidence becomes the case for a new skill.
Rate any run yourself, with a note the next reviewer can read.Auto-score runs with a reviewer model on resolution, policy and tone.Escalate a pattern of low scores straight into a skill draft.
05 Surface 05 of 05
Skills Hub
When an agent keeps failing, it gets taught a skill instead of a patch.
Draft a skill from failed runs. Replay it in the lab and deploy only if it passes.
Draft a skill from the failures that prove it is needed.Test it in the lab on replayed runs, before and after, side by side.Deploy on pass only. A skill that does not clear the gate stays a draft.
Observe
Runs, tokens, cost, errors, health.
Every report puts operating context beside the agent that produced it.
checkout-recovery 1,842 tok
support-router 624 tok
invoice-parser 1,104 tok
Local state
Control without a proxy.
The wrapper checks paused or stopped state before each LLM call.
traffic route unchanged
agent state remote
Model override
Change the next run.
The integration resolves the override locally before the request.
before gpt-4.1-mini
after claude-sonnet-4-6
Prompt cache
Fetch, cache, fall back.
Successful fetches are cached; failures fall back to cache, then hardcoded prompt.
current v12
incoming v13 · cached
Failure path
Your app keeps running.
Heartbeat and report failures are caught so ACP stays outside the critical path.
ACP unavailable caught
agent run continues
The screens you operate from.
Dashboard
Live monitor
Agents
app.acp / dashboard
Demo
Workspace
Dashboard
Live Monitor
Agents
Runs
Apps
Audit
Settings
Dashboard
320 agents across 32 apps
Money
Cost anomaly · 184% above 7-day median
checkout-recovery
Explain Rollback Pause
21 retries in 12 minutes
checkout-recovery
Explain Rollback Pause
No heartbeat for 4 minutes
digest-writer
Explain Rollback Pause
Activity
Recent activity
12s checkout-recovery success 1.8k 2.4s
41s checkout-recovery error — 9.1s
1m support-router success 624 1.1s
2m invoice-parser success 1.1k 3.6s
4m lead-enricher idle — —
Live Monitor
Live
Refresh:5s
1 agent needs attention
Search agents…
All
Health (worst)
checkout-recovery
checkout-suite
18 gpt-4.1-mini
12s ago 184.2k tokens
support-router
support-desk
96 gpt-4.1-mini
1m ago 42.6k tokens
invoice-parser
billing-ops
91 claude-sonnet-4-6
2m ago 61.0k tokens
digest-writer
growth-loops
insufficient data
No activity —
Agents
Search agents…
All Apps
All Types
List Hierarchy
Runs Agent Result Tokens Duration checkout-recovery Error — 9.1s support-router Success 624 1.1s invoice-parser Success 1.1k 3.6s
Apps checkout-suite 14 agents
Connected support-desk 9 agents
Connected billing-ops 6 agents
Warning Audit Action Target Record Pause requested checkout-recovery Command recorded Model updated support-router Acknowledged Prompt approved invoice-parser Version v12
Settings Workspace Demo workspace
Read-only preview Agent configuration Models and prompts are managed per agent.
Changes require an explicit action Command security Signed webhooks with timestamp and replay protection.
See the Security & Trust page for details
Understand every run. Act on what you find.
Compare run diagnostics, evaluations, and agent controls.
Capability
ACP
Langfuse
LangSmith
Helicone
Braintrust
Datadog
Run diagnostics and evaluations
Run diagnostics · feedback scoring · replay evaluations
Scores & experiments
Online & offline evals
Tracing
Scorers & experiments
APM
Retrofit an already-shipped app
Generated integration
Instrumentation
Instrumentation
Gateway or SDK
Instrumentation
Instrumentation
Pause one hardcoded agent remotely
Yes
Not its job
Not its job
Not its job
Not its job
Not its job
Swap model without redeploying
Local override
App-managed
App-managed
Via gateway
App-managed
App-managed
Push prompt without redeploying
Versioned fetch + webhook
Prompt management
Prompt Hub
Via gateway / SDK
Prompt environments
App-managed
Hosted maturity and ecosystem
Early
Established
Established
Established
Established
Enterprise-grade
Swipe table to compare →
Signed commands
HMAC-SHA256 over ts.nonce.body
Compared with a timing-safe check, and verified against the current and previous secret so rotation never drops a command.
Replay defence
300-second window, nonce cache
A stale timestamp is rejected with 403. A nonce that has already been seen is rejected with 403.
Fails open
Reporting errors are caught
Heartbeat and report failures log and return. A prompt fetch failure falls back to cache, then to the hardcoded prompt.
Not in the path
No proxy, no key custody
Your provider keys stay in your app. ACP never sits between your agent and the model.
The questions a skeptical developer asks.
Answered against what the code actually does today.
Do I have to adopt a framework?+
No. Add the generated integration to your existing app. Your agents stay where they are.
What happens if ACP goes down?+
Your agents keep running. Reporting failures are caught, and prompts fall back to the cache or your app defaults.
Does my LLM traffic route through you?+
No. Your app calls the model provider directly and keeps its provider keys.
How fast does a pause actually take effect?+
The next wrapped run stops before calling the model. Confirmation follows on the next heartbeat.
Should I drop my observability tool?+
No. Keep your existing tools for analysis and use ACP to operate your agents.
How many external teams run this today?+
Zero external teams are connected. The hosted product is in early access.
320+
agents operated
32+
applications
6 months
in daily production
Who built it
Built from the operator’s seat.
Built from six months of operating 320+ agents across 32+ applications.
That history proves the operating model. ACP currently has zero external connected apps; the public product is at the beginning of that validation.
The whole evaluation
Connect one app. Operate one real agent.
If you can see a heartbeat and pause a live agent without a deploy, you understand the product.
1
Create the app Request access. Once approved, create the app and receive an API key and webhook secret. No sales call.
2
Paste the integration Hand the generated file to your coding agent, or drop it in yourself.
3
Take control Watch the first heartbeat land, then pause the agent from the dashboard.
your connected apps
checkout-suite 14 agents · beat 12s ago connected
support-desk 9 agents · beat 41s ago connected
your app waiting for the first heartbeat pending
connect new +