SafeAct — exactly-once for AI agent actions

The whole change

One decorator. A caller-supplied key.

Your side effect stays exactly as you wrote it.

Add @safe and pass a key that names the logical action. The same key seen again returns the stored result instead of re-running — and it holds across process restarts, because the record lives in SQLite, not memory.

Exactly-once, even across crashes & retries
Honest in-doubt — never a blind retry
Sync & async · one file · no server

refund_agent.py

from safeact import safe

@safe(store="agent.db")
def charge_refund(key, customer, amount):
    stripe.charge(customer, amount)   # real side effect
    return {"charged": amount}

# retry-safe: runs once, then replays the result
charge_refund("refund-8842", cust, 50)
charge_refund("refund-8842", cust, 50)  # no double-charge

What the open-source core gives you

Four guarantees.

Enforced in code, underneath any agent framework — and proven in a durable record.

Exactly once

Crash, restart, retry — the action happens once, and holds across process restarts.

Honest in-doubt

If the gap is truly unknowable, it stops and surfaces for a human — never a blind retry.

Audited

A durable record of every attempt and its outcome — your whole audit trail.

Composable

Saga rollback for multi-step flows; drops into MCP & LangChain in one line.

Verify, then execute

Two checkpoints every agent action should pass.

An agent can hurt you two ways: a wrong answer, or a double action. wobbly catches one. SafeAct stops the other.

01 · Verifytest-time

wobbly

Is the agent's judgment sound?

Catches wrong, brittle, and biased outputs before any action fires — with no labels and no answer key. Runs in CI, on your laptop, zero risk.

535 receipts · flagged outputs 2.7× more likely to be wrong

02 · Executerun-time

SafeAct

Did it happen — once, on the record?

Runs the action exactly once through crashes, restarts and retries — honest in-doubt when it genuinely can't know, with a durable record of every attempt.

exactly-once · mutation-tested · one SQLite file

One refund · two failures · two tools

A support agent decides to refund $50. The $50 could be hallucinated — it should have been $5. Or the refund could fire twice, because the run crashed and retried. wobbly flags the shaky refund in testing. SafeAct guarantees the good one fires exactly once in production.

100% open source · no paid tier

Four libraries. One trust thesis.

The whole thesis, shipped as four small packages — one pip install each, every line open source. No server, no account, no upsell.

Apache-2.0 licensed 0 external deps in core Runs offline Read the source on GitHub

safeactcore

Exactly-once execution, honest in-doubt, and saga rollback — one SQLite file, no server.

pip install safeact

safeact-mcpmcp

Make any MCP tool exactly-once and auditable, in one decorator.

pip install safeact-mcp

safeact-langchainlangchain

Exactly-once LangChain / LangGraph tools, keyed on the tool_call_id.

pip install safeact-langchain

wobblyeval

Catch wrong LLM outputs with no labels. On 535 real receipts, the outputs it flagged were 2.7× more likely to be wrong.

pip install wobbly

Let your AI act. Keep the proof.

SafeAct comes out of running real, action-taking agents in production — where we learned exactly how they break. Free and open source.

Get started — free Explore wobbly