Open-source transaction infrastructure

Make agent actions
reversible.

Safely execute multi-step agent workflows across APIs, databases, SaaS, and infrastructure—with recovery built in.

FRAMEWORK-NEUTRAL · MODEL-NEUTRAL · SELF-HOSTABLE

Customer Onboarding Agent
tx_98f31ac
Execution pathINTERRUPTED
SalesforceCreate customer record
Completed
StripeCreate subscription
Completed
NetSuiteCreate billing account
Completed
!
Provisioning APICreate workspace
Failed
!
Transaction interruptedAgentTX generated a safe recovery plan.
Recovery plan5 ACTIONS
Salesforce recordRollback available
Stripe subscriptionCompensate cancellation
NetSuite accountRollback available
Workspace provisioningSafe to retry
Email notificationNot executed
StatusRECOVERING
Total actions4 / 5
Recovery time~12 sec
Irreversible0
Human approvalNo

↑ Try it — execute the recovery plan

BUILT FOR THE AGENT STACK YOU ALREADY USE

OpenAIAnthropicLangGraphCrewAIPydanticAIMastraMCP
01 / DEVELOPER EXPERIENCE

Transaction safety
in a few lines.

Wrap high-impact workflows in a transaction. AgentTX records every action, prevents duplicate execution, and safely resumes after failure.

01

Simple defaultsStart safe, customize when needed.

02

Explicit recoveryDefine rollback and compensation inline.

03

Complete historyTrace every side effect and decision.

Read the quickstart
onboard_customer.py
from agenttx import transaction

@transaction()
async def onboard_customer(customer):
    crm = await tx.step(
        salesforce.create_customer(customer),
        rollback=salesforce.delete_customer
    )

    subscription = await tx.step(
        stripe.create_subscription(customer),
        compensate=stripe.cancel_subscription
    )

    await tx.step(
        send_welcome_email(customer),
        irreversible=True,
        approval="before_execute"
    )
$ agenttx run onboard_customer ✓ transaction created
THE PARTIAL FAILURE PROBLEM

An agent can fail after changing
five production systems.

Reasoning frameworks decide what to do. AgentTX makes sure the real-world side effects are safe when something goes wrong.

01CRM recordCreated
02SubscriptionStarted
03BillingConfigured
04WorkspaceFailed
05EmailBlocked
02 / TRANSACTION MODEL

Every action has a
recovery strategy.

AgentTX understands that not every real-world action behaves the same way. Make the semantics explicit.

REVERSIBLE

Roll it back.

Directly undo actions in reverse dependency order.

Create record → Delete record
COMPENSATABLE

Offset the effect.

Run a new action that restores the business state.

Charge customer → Issue refund
RETRYABLE

Try it safely.

Use idempotency keys to prevent duplicate effects.

Timeout → Reuse operation result
IRREVERSIBLE

Require approval.

Put a human checkpoint before the point of no return.

Send email → Approval required
03 / BUILT FOR REAL-WORLD ACTIONS

Where agents act,
AgentTX protects.

SUPPORT

Prevent duplicate refunds

Reconcile payment state before retries and compensate failed downstream actions.

Explore use case →
CODING

Make deployments recoverable

Checkpoint changes, protect production pushes, and reverse failed releases.

Explore use case →
INFRASTRUCTURE

Provision without orphaned resources

Roll back partial environments and keep cloud state consistent.

Explore use case →
OPEN SOURCE / APACHE 2.0

Built in the open.
Ready for production.

Run AgentTX anywhere. Inspect every line. Extend the transaction layer for your own tools and workflows.

View on GitHub ↗Read the docs →
agenttx / agenttxPublic

Transaction and recovery infrastructure for AI agents.

Python☆ Star⑂ Fork
$ pip install agenttx
04 / PRICING

Start open source.
Scale with confidence.

Pay for transaction actions—not tokens. Self-host for free, or let AgentTX Cloud run the infrastructure.

OPEN SOURCE

Free

For developers and self-hosted deployments.

  • Transaction runtime
  • Rollback & compensation
  • Retry engine
  • Local replay
View on GitHub
TEAM

$499/ month + usage

Controls and visibility for growing teams.

  • Team workspaces
  • 30-day history
  • Approval workflows
  • Shared policies
Start building

Let agents take action.
Keep control.

Retry safely. Roll back failures. Compensate side effects.
Protect irreversible actions.

OPEN SOURCE · FRAMEWORK-NEUTRAL · BUILT FOR PRODUCTION