For AI agents with commitment issues

Your LLM is stuck. Make it ask for help.

Stop burning tokens on work that needs external context, credentials, approval, or judgment. Agent Blocked lets your AI tell you when human help is the fastest path forward.

Start free See setup

Product walkthroughSee how an agent asks for help

criticalprod-agent needs credentials

AWS deploy role expired before release verification could continue.

email sentsms sentvoice queued

setupConnect browser-agent

Copy environment, then run the matching CLI command in the project repo.

Create profile > connect agent > test delivery > receive real alerts

CALL

Email

Quiet, searchable incident recaps for low-drama events, postmortems, and async review.

CALL

Text message

Fast pings for on-call humans when an agent needs a nudge but not a full phone call.

CALL

AI voice call

Agent Blocked writes a short script, places the call, and reads it aloud.

Less token smoke

When the blocker is outside the model, more thinking is just expensive waiting.

Missing credentials, unclear goals, approvals, and production decisions are not reasoning problems. Agent Blocked turns those moments into a clean human handoff before your agent loops, guesses, or burns budget on work that cannot succeed yet.

Customizable triggers

Choose exactly when an agent should ask for help.

Hard blocked

Credentials missing, OAuth expired, or a human approval gate is stopping the run.

Needs direction

The agent is still capable, but the goal is ambiguous or the plan needs a human nudge.

Tool failure loop

Repeated API, browser, shell, or retrieval failures are burning tokens without progress.

Context limit

The conversation is near its context or token budget and the agent may lose important state.

Check-in request

A low-severity check-in when a long-running agent needs human reassurance or context.

Approval required

The agent wants permission before sending, buying, deploying, deleting, or escalating.

Low confidence

The model can continue, but confidence has dropped below your comfort threshold.

Built for real agent failures

Escalation rules for practical blockers and low-confidence moments.

Credential gates

Wake the right person when a deploy agent needs AWS, Supabase, Stripe, or Vercel access.

Goal drift

Ask for direction when the agent is technically working but wandering away from the outcome.

Approval moments

Interrupt before irreversible actions: sending email, charging cards, deleting data, or shipping code.

External limits

Alert when a run hits GitHub Actions, API, billing, provider, or account limits.

Context limits

Notify before a long-running session compacts or loses the details needed to finish cleanly.

Quiet hours

Suppress non-critical interruptions overnight while still letting critical events break glass.

Connected agents

Attach one profile to many named agents, then view each agent's setup commands and recent events.

After setup

See which agent needed help, why, and who was notified.

Each profile keeps a clear incident trail with delivery attempts, skipped channels, failed provider sends, and the setup commands for every connected agent.

prod-agenttrialing / monthly

needs_credentials

AWS deploy role expired before the release agent could continue.

critical

email: sentsms: sentvoice: sent

lonely_agent

The browser agent needs a quick confidence check before continuing.

low

email: sentvoice: skipped

Human intervention loop

A simple support desk for agent handoffs.

Connect

Add the people who should be interrupted and choose one notification channel per profile: email, text, or AI voice.

Attach agents

Add one or more agents to the profile, then copy the exact setup commands shown for each tool.

Intervene

Humans get a useful message with the agent name, severity, what happened, and what the agent needs next.

Published npm CLI

Install the agent adapters from npm.

Create an alert profile, add an agent, then save the generated scoped token and install command in the project where that agent works. The installer adds project instructions, a local reporter command, and a recent-notification check for duplicate alerts.

View integration examples

Save local config

npx agent-blocked@latest configure

Install and verify

npx agent-blocked@latest install --tool=all
npx agent-blocked@latest doctor --tool=all

Report important events

npx agent-blocked@latest report \
  --event=needs_credentials \
  --severity=critical \
  --reason="Missing deploy credentials"

npx agent-blocked@latest codex -- codex "ship the fix"

Coding CLI installs

Wire it into the agent tools teams already use.

Create a profile, add an agent, then run the matching installer in the project where that agent works. Claude Code and Codex can fire hooks automatically for supported events. Gemini CLI and Aider get project instructions that tell the agent when to notify the user.

Claude Code

Native hook config for permission prompts, idle prompts, permission denials, and stop failures.

npx agent-blocked@latest install --tool=claude

Codex

Project hooks and AGENTS.md instructions let Codex report approval requests and tool failures.

npx agent-blocked@latest install --tool=codex

Gemini CLI

GEMINI.md instructions give Gemini CLI workflows a consistent blocked-agent escalation command.

npx agent-blocked@latest install --tool=gemini

Aider

CONVENTIONS.md instructions route approval, credential, and direction blockers to humans.

npx agent-blocked@latest install --tool=aider

Open full install docs

10-day free trial

Start with 10 days free.

Choose the plan that starts after the trial. Stripe stores the payment method now, and you can cancel from the billing portal before the trial ends.

Default

Monthly

per month after trial

10 days free first

Best for trying Agent Blocked with one team or project.

Start 10-day trial

Save 17%

Yearly

$90

per year after trial

10 days free first

Two months free for teams that know they want it running.

Start 10-day trial

Reasonable objections

For teams that need automation with clear human oversight.

Is this just another alerting tool?

No. It is intentionally shaped around AI agent failure modes: missing credentials, vague goals, approval gates, low confidence, and repeated tool loops.

Will this leak secrets to the browser?

No global webhook secret is shown in the UI. Setup returns a scoped agent token, and sensitive provider keys stay server-side in Vercel environment variables.

What if delivery fails?

Delivery attempts are logged per channel so you can see sent, failed, or skipped notifications from the manage page.

When do real alerts start?

After Stripe Checkout starts the 10-day trial. A payment method is required before profiles, tokens, setup commands, or test alerts unlock.

Why not let the model keep trying?

If the blocker is outside the model, more tokens rarely solve it. The fastest path is a structured handoff to a human with the right context.

Need help wiring an agent?

Email support@agentblocked.com or visit the support page for setup, billing, delivery, and webhook guidance.

Contact support

Agent reports

When an agent gets stuck, it sends a clean handoff.

Agent Blocked turns blocker details from your agent into a readable incident, applies the profile rules, and alerts the right human through the selected channels.

See npm CLI setup

POST /api/agent-blocked
Authorization: Bearer $AGENT_BLOCKED_AGENT_TOKEN

{
  "agentId": "research-agent-7",
  "runId": "run_123",
  "provider": "openai",
  "model": "gpt-5",
  "eventType": "needs_credentials",
  "severity": "critical",
  "confidence": 23,
  "reason": "The agent needs AWS credentials before it can continue.",
  "details": "Credential prompt appeared in the deploy workflow."
}