Stop risky AI agent actions before they execute.
Replay the failure. Prevent the next one.
SafeRun sits before tool calls like refunds, emails, database writes, CRM updates, payments, and internal APIs. It can pause, block, or request approval, then capture the decision context so teams can create rules that prevent repeats.
One refund. One pause. One rule.
A support agent is about to refund $1,800 because it misread a cancellation email. SafeRun sees the refund exceeds policy, pauses the action before Stripe is called, captures the agent plan, tool arguments, context, and policy decision, then turns the failure into a rule for future runs.
Built around the action, not just the log.
Control room
See what needs review and what is running safely.
Incident report
Inspect the decision-time context behind one action.
- agent.plan
- tool.args
- policy.match
- decision: paused
Policy catalog
Turn repeated risks into approval or blocking rules.
- refund.amount > $500 → approval
- delete_customer → block
- email.bulk → throttle
Safety report
Package action-safety evidence for design partners.
Not another observability dashboard.
Observability shows what happened after execution. SafeRun decides whether risky agent actions should execute at all.
Observability
- Shows what happened after execution
- Helps debug systems
- Tracks logs, traces, metrics, and evals
SafeRun
- Checks risky actions before execution
- Pauses, blocks, or requests approval
- Captures decision-time context
- Turns failures into rules
They watch the system. SafeRun controls the action.
Use LangSmith, Honeycomb, New Relic, Datadog, or your existing observability stack to understand the broader system. SafeRun sits at the action boundary and decides whether risky tool calls should execute.
Traces, logs, metrics, and evals across the whole system. Answers "what happened?"
Approval before execution at the tool boundary. Answers "should this run?"
Engineering wraps the tool. Ops owns the rules.
Engineers install SafeRun before tool calls. Operations, support, risk, and compliance teams use the review queue, policy catalog, and safety report to manage risky agent actions without filing engineering tickets for every rule change.
Stop hand-writing guardrails around every tool.
Teams start with a few if-statements. Then come thresholds, approvals, retries, policy changes, audit trails, and scattered logic across every agent. SafeRun gives teams one shared action-control layer.
Built for agents with write access.
SafeRun is for engineering teams deploying agents that touch customer, financial, or production systems.
Wrap your tool calls. Keep your stack.
SafeRun works with Python and TypeScript agents, including OpenAI Agents SDK, Anthropic, LangGraph, LangChain, Vercel AI SDK, and MCP tool calls. The framework stays. SafeRun sits in front of the tool call.
from saferun import guard
@guard(agent="support-agent", tool="stripe.refund")
def refund_customer(amount, customer_id, reason):
return stripe.Refund.create(
amount=amount,
customer=customer_id,
reason=reason,
)See the action before it becomes an incident.
A guided walk-through of one paused action and the replay SafeRun captures.
support-agent-v2 attempted stripe.refund for $1,800
- Customer requested refund01 · user message
- Agent selected stripe.refund02 · tool call
- Tool arguments: amount = $1,800, customer_id = cus_928103 · args
- Policy check: refund amount exceeds $500 limit04 · policy
- Decision: paused for human approval05 · decision
- Replay saved with full context06 · audit
- Suggested prevention rule drafted07 · rule
Priced around protected actions, not generic logs.
A protected action is a real tool or API action SafeRun checks before execution: refunds, payments, database writes, emails, CRM updates, internal API calls, deletes, or production changes.
Developers testing SafeRun locally or in test mode.
- Test environment
- Limited protected actions
- Basic replays
- Quickstart
- Python and TypeScript SDKs
Teams deploying agents with real tool access.
- Production protected actions
- Approval requests
- Policy catalog
- Replay history
- Agent Action Safety Report
Teams running multiple agents or higher action volume.
- Higher protected-action volume
- Multiple environments
- Longer replay retention
- Advanced reporting
- Priority onboarding
Regulated or security-sensitive teams.
- SSO
- Audit logs
- Compliance retention
- Private deployment / VPC option
- Custom policies
- Dedicated support
Give production agents a checkpoint before they act.
If your agent can refund, delete, email, update customer data, write to a database, call internal APIs, or touch production systems, SafeRun can help you test the action-control layer before you scale it.