Human approvals for AI agents and automation.

Human approvals for AI agents and automation.

Connect your software to the hosted ActionBox API. Review requests in a human inbox, record a decision, and let your workflow continue.

Create a free Source Try the live demo

Free for developer CLI use · No credit card required

Production deployment

Action open · act\_9a3c

Needs review

Action required

Approve the production schema migration?

Version 2.18.0 passed CI. Review the affected rows, backup state, and rollback plan before the workflow continues.

Affected rows14,208

BackupVerified

RollbackReady

Approve Reject

Audit trail

Fingerprint · sha256:7cf4…a219

1. 12:04 CI passed 482 checks · 0 failures 2. 12:07 Action created Workflow paused safely 3. Now Waiting for you 28 minutes remaining

Terminal CLI

$ actionbox ask "Deploy v2.18?" --wait

Paused for a human response

Hosted web inbox

Review the current context, choose a typed response, and keep the event history.

CI/CD workflow

POST /v1/actions

Gate waiting safely

Actionbox MCP

ask\_human → get\_action

Waiting for a reviewer

Action Controls

Give people safe ways to fix what software cannot.

Approval is one answer. Controls let a person request a bounded operation while the Action stays open and the Source remains responsible for execution.

Action open

Import stalled

Needs attention

72,831 / 184,221

records

Failed validation24 records

Last safe checkpoint72,800

Resume from checkpointRetry failed batchStop importSee the control lifecycle

One Action lifecycle

From silent software to a verified outcome.

ActionBox keeps the interruption, the human response, and what happened afterward in the same server-authoritative record.

1. 01

Observe

Ask

Decide or control

Verify

Typed decisions

Boolean, choice, text, number, rating, and structured form responses.

Silent-work detection

Heartbeat Watches turn missed progress into an Action that needs attention.

Human input for agents

MCP-capable clients can ask, wait, and resume from the authoritative result.

“

BS

Bhupal S

Founder of Berkeley Computer

“Before ActionBox, our agents could sit blocked for hours because nobody knew they were waiting for input. Now we’ve cut median approval time from 38 minutes to 6 minutes, saving more than 50 engineering hours a month while keeping control over what reaches production.”

Integrations

Connect your workflows.

Bring a person into the loop when your workflow needs a decision. Start with a guide for the tools you already use.

GitHub ActionsAdd a human approval step to your pipeline.Read the guiden8nAsk for a decision and branch your workflow on the answer.Read the guideLangGraphConnect graph interrupts to a human decision.Read the guideOpenAI Agents SDKReview tool requests before resuming your agent.Read the guide

zsh — actionbox ask

Copy

$ actionbox ask "Deploy to production?" \\ \--option approve=Approve \\ \--option reject=Reject \\ \--wait --timeout 30m

Action act\_7f9c created · Waiting for human decision...

Action published to the hosted inbox.

✓ Decision received: approveexit code: 0

Command line interface

Pause software directly from the shell.

\`input()\` for scripts and automation that cannot wait on a local tty. The command blocks using bounded long-poll requests, waiting on the human decision before continuing.

See the complete CLI and pipeline guide

• Low-overhead wait
Bounded long-poll requests avoid a tight local polling loop.
• Machine-readable results
Print the selected option or a complete JSON result for scripts and CI pipelines.
• CI/CD ready
Use the hosted API from GitHub Actions, or call the CLI from your pipeline.

Model Context Protocol

Give AI agents a person to ask.

Connect an MCP-capable client to Actionbox's hosted human-input tools. The agent creates a durable question, a person answers in the inbox, and the client reads the typed result.

• Protected local bridge
The public CLI keeps the Source key out of JSON-RPC messages and chat history.
• Source-scoped identity
Use a separate Source for each runner so credentials can be rotated independently.
• Durable typed decisions
Human input remains in the hosted Action lifecycle when the client disconnects or restarts.

MCP client · Actionbox

Drop table users\_legacy on prod cluster to clean up disk space.

actionbox · ask\_humanWAITING FOR HUMAN

toolpostgres\_execute\_sqlstatusopen

Actionbox request sent to team inbox. Waiting for approval before dropping table...

Approved by reviewer · Decision recorded

Available where work happens

Five surfaces. One authoritative Action.

Review from the web, mobile inbox, or Slack, or let software wait through the CLI and MCP. Every surface reads the same hosted state.

• 01CLIPause scripts and CI jobs with a bounded wait.macOS · Linux · Windows
• 02MCPGive compatible agents durable human input.Streamable HTTP
• 03WebReview context, history, assignments, and outcomes.Modern browsers
• 04MobileReceive optional push alerts and respond on the go.iPhone · iPad
• 05Slack (Beta)Review requests, use controls, and follow the decision history.Team · Assisted setup
Choose an integration path

Your inbox, wherever you work.

MobileSlack (Beta)Web

ActionBox for iPhone

The decision inbox that can reach you.

Get a push alert when an Action needs attention, open the latest context, and respond from your phone. The API remains the source of truth, so a delayed notification never changes the workflow state.

• Review typed requests
• Receive optional push alerts
• Snooze work that can wait
• Read context and event history

See how mobile review works

Requires iOS 15 or later. Sign in with your ActionBox account.



The live ActionBox inbox for iPhone.

Slack (Beta)

Review the request. Stay in Slack.

Read the context, approve or reject, and see who responded. The decision and its history stay with the Action.

• Key facts up front, full evidence in the details
• Action Controls and comments from Slack
• Replies that show who responded and what they decided
Set up Slack

Available on Team. Beta installation is assisted.



Actual Slack interface with a demonstration request. Select the image to view full size.

ActionBox on the web

Room for the full picture.

Review requests, inspect rich context, and follow the decision history from your browser.

Open the web inbox

ActionBox · Web inbox

Inbox

Needs response

History

NEEDS RESPONSE · DEMO

Approve the simulated $2,400 refund?

The amount exceeds the $1,000 automatic approval limit. Review the order before responding.

Order#8842

Amount$2,400

ApproveReject

Demonstration only. No refund is executed.

Built for high-stakes automations.

Supervise autonomous software when mistakes are irreversible, expensive, or destructive. Put a human in the loop without building custom dashboard infrastructure.

• ### AI Agent Guardrails
Intercept external API calls, bash executions, and tool mutations. Give autonomous agents a safe human escape hatch when uncertainty is high. LangGraphClaude MCPOpenAI Agents
• ### Deployment Gates
Pause GitHub Actions, GitLab CI, and Terraform runs. Review rollout plans, diffs, and cluster health checks before production traffic switches. GitHub ActionsTerraformKubernetes
• ### Data & Financial Mutations
Safeguard customer refunds, user data purges, and destructive database migrations with human sign-off and decision history retained according to your plan. Stripe RefundsSchema MigrationsGDPR Erasures

Supervises cron jobs, microservices, internal scripts, and AI workflows with the same unified API.

Production field guides

Design the control system around the approval.

Practical, referenced guides for the identity, policy, state, and evidence boundaries that make human review reliable.

Browse all guides

Policy and evidence

AI agent governance

Define which operations run automatically, which need review, and what evidence remains afterward.

Read the guide : AI agent governance

Production architecture

AI agent orchestration

Coordinate durable state, tool policy, human decisions, retries, and execution outcomes.

Read the guide : AI agent orchestration

Operational visibility

AI agent observability

Trace proposals, decisions, tool calls, and outcomes without treating raw prompts as an audit trail.

Read the guide : AI agent observability

Predictable pricing for developers and engineering teams.

Plans include monthly Actions and members. Add capacity as you grow. Choose monthly billing or save on an annual plan.

Two months free annually

MonthlyYearly

Free forever on Developer·14-day money-back guarantee on first paid purchase

Developer

For individuals automating personal scripts, cron, and side projects.

$0

Forever free

• 250 Actions / month included
• 3 active Sources
• 3 heartbeat Watches
• 7-day history & payload retention
• Hosted web inbox and API playground
• Standard webhook retries (24h window)

Get started free

Pro

Most popular

For power engineers, solo founders, and automated personal toolchains.

$10/month

$120 billed annually · Save 17% · Updating local price…

Everything in Developer, plus:

• 1,000 Actions / month included
• Additional Actions: $8 per 1,000-Action pack
• Packs expire monthly; no automatic usage charges
• Unlimited Sources & Unlimited Devices
• 20 heartbeat Watches
• 30-day history retention & data export
• Rich context blocks (Diffs, JSON, Logs)
• Action Controls for safe follow-up operations (rolling out)
• Synchronous AI agent bridge (--wait)

Start with Pro

Team

For engineering teams, CI/CD pipelines, and shared automation workflows.

$49.17/month

$590 billed annually · Save 17% · Updating local price…

Everything in Pro, plus:

• 5,000 Actions / month included
• Additional Actions: $5 per 1,000-Action pack
• Packs expire monthly; no automatic usage charges
• 90-day history retention & audit logs
• 5 members included; $8/month per additional member
• 50 heartbeat Watches
• Shared inbox and role-based access
• High rate limits (100 RPS burst / 50k calls/hr)

Get started with Team

Scale

For teams that need more Actions and members. Available by arrangement.

Contact us

Confirm availability and capacity with us before purchase.

Everything in Team, plus:

• Discuss 25,000 Actions and 15 members
• Target pricing: $199/month or $1,990/year
• 90-day history retention
• Additional capacity priced in your agreement

Discuss Scale

Compare all plan quotas and details

Enterprise

Fit ActionBox into your operating model.

Work directly with us to evaluate identity, access, retention, integration, support, and service expectations for high-impact autonomous software.

Contact Enterprise See plan quotas

• Identity requirement review
Review roles, account access, and organization security requirements.
• Governance planning
Review team workflows, audit retention, billing, and shared inbox requirements.
• Integration and service planning
Review workflow connectivity, delivery expectations, rollout needs, and support targets.

Frequently asked questions.

Have questions about integrating with your stack? Email info@actionbox.cloud or explore the interactive documentation.

• What is a human approval system for autonomous software?
It is infrastructure that pauses a risky or ambiguous software operation, presents the right typed context to a human operator, records a structured decision, and lets the workflow resume safely on a signed callback. Actionbox extends that pattern with run-health monitoring, heartbeat watches, and post-execution outcome tracking.
• Can I review ActionBox requests in Slack?
• How does the CLI wait without burning CPU or polling?
• Does Actionbox work with non-AI automations?
• Are human decisions cryptographically verifiable?

`action.control_requested`Source running

`action.resolved`signed outcome

Give your software a way to ask—and a safe way to recover.

Create one free Source. Ask for a decision, attach bounded controls, and read the signed result from your existing workflow.

Create a free Source Read the 5-minute guide