Do it now
Reversible, well-understood repairs you have pre-approved: replay a buffered delivery, refresh a token, re-map a renamed field, back off from a rate ceiling. The agent acts and files the narrative.
Product
Watch. Decide. Heal. Report.
One loop, running without pause across every API, webhook and sync between your systems. Here is exactly what happens at each turn of it — and where the loop stops and asks you.
Watching is free Billed per heal Billed per flow-run
01 · Watch
It knows what normal looks like.
Every connection gets a continuous health probe and a memory. The agent learns your baselines instead of waiting for a threshold someone guessed at during setup.
- A schema fingerprint per endpoint, taken every time data moves
- Latency and error-rate baselines learned per hour of the week
- Delivery cadence, so a silence that should be traffic becomes an event
- An expiry horizon on every credential, key and token in the path
Endpoints
every API, webhook and sync
Probed on a schedule you can tighten per connection. A probe costs you nothing, so they run often.
Payload shape
fingerprint, not a copy
The agent records the structure of what passes through — field names, types, nesting — so a rename is visible the moment it happens. Payload contents stay inside your boundary.
Credentials
expiry horizon
Tokens and keys are tracked to their expiry date, so the refresh happens before the failure, not after it.
Volume
expected cadence
Nothing arriving is a symptom too. A queue that goes quiet at the wrong hour raises the same flag a 500 would.
02 · Decide
Classified first, acted on second.
An anomaly is named before anything is touched — auth failure, schema drift, rate ceiling, silent data gap, upstream outage — and the name decides what happens next.
Ask first
Anything that changes data or touches money waits for a yes. You get the diagnosis, the proposed action and the blast radius in one message, and one tap either way.
Stop and page
When the safe move is no move — an upstream outage, an unknown failure mode, a policy with no rule for this — the agent holds the queue in order and hands you the whole picture.
What stays in your hands
non-negotiable
You write the policy and you can change it in a sentence. Nothing irreversible ever runs without a rule you approved, deletions are never automatic, and every action — taken or refused — lands in an audit narrative you can read without a manual.
03 · Heal
The repairs it can safely make itself.
Five moves cover the overwhelming majority of integration failures. All five are reversible, ordered, and logged in words.
Replay from the buffer
silent gaps · failed deliveries
Every inbound event is held briefly whether it succeeded or not. When deliveries go missing, the agent replays them oldest first and confirms each one downstream.
Back off and find another path
rate ceilings
When an endpoint says slow down, the agent slows down, keeps ordering intact, moves the urgent lane onto a second path where one exists, and drains the rest as capacity returns. A failover, not a flood.
Re-map from schema history
drift
A renamed or re-typed field is matched against the fingerprint history and mapped back to what your systems expect. A genuinely new field is surfaced, not guessed at.
Refresh the credential
auth expiry
Token refresh flows run on the horizon rather than at the failure. Writes queue in order while the handshake completes, then go through.
Quarantine the bad record
poison payloads
One malformed record does not get to fail a batch or corrupt a destination. It goes to a review queue with the reason attached; the other 3,880 rows land on time.
Try a failure · compact
Same engine as the homepage demo
All connections healthy. Pick something to break.
Detect
Diagnose
Heal
Report
Orchestration
The flows in between.
Repair is half of it. The other half is running the data flows themselves: declare what should happen when something occurs in one system, and the agent carries it to the others with per-hop transforms, retries and ordering guarantees.
Flow inspector · live
Hover a hop to hold it · click a system to mute it and watch flows heal around it
Hop detail
Payload in motion Hop held for inspection
Per-hop transforms
Field mapping, type coercion and enrichment happen on the hop that needs them, so one system's rename never forces a rewrite everywhere else.
Ordering guarantees
Per-key ordering by default: a later edit never lands before the earlier one it depends on, even after a retry or a pause.
Honest delivery semantics
At-least-once with deduplication, or exactly-once per key where the destination supports it. Whichever a flow uses is written on the flow, not buried in a footnote.
04 · Report
Written for the person who has to explain it.
The warehouse endpoint stopped answering at 02:41. We held 212 fulfilment requests in order, moved the four same-day orders onto the backup path, and drained the rest at 03:06 when it came back. Two addresses were quarantined for review. 0 records lost.
Auto-healed under policy “upstream outage · hold and drain” · 1 heal billed
Every Monday the same voice writes a one-page brief: what broke, what healed itself, what is still waiting on you, and what the week cost. No log-diving, no query language, no new dashboard.
Part of the Auth Alien platform