TACK CLOUD
Tack Cloud is the paid tier of Tack — runtime risk scoring for AI agent tool calls. tack-guard scores every tool call locally, in under 1ms; Tack Cloud gives it a baseline that survives restarts and correlates behavior across sessions and across agents, catching the attacks a single, in-memory session structurally cannot see.
ONE LINE
You do not re-architect anything. Keep your existing tack-guard
setup and add guard.connect(). The
scoring still runs locally; events sync to Tack Cloud after the
verdict, so baselines now persist and correlate. The API is already
final, so this is a no-redesign upgrade.
import { createGuard } from '@tacksec/guard'
const guard = createGuard()
guard.connect('tk_live_xxx') // the only new line
// Everything else is identical. Scoring still runs
// locally, in under 1ms. Baselines now survive restarts
// and correlate across sessions and across agents.
const result = guard.evaluate({
action: 'write',
target: 'user_db.credentials',
})WHAT CLOUD ADDS
tack-guard is single-session by design: fast, private, zero dependencies. Tack Cloud adds exactly the three things that need memory across processes and across agents.
Your agent's per-agent baseline is written to Tack Cloud, not held in process memory. Restart, redeploy, or scale to ten replicas and the baseline is still there. No cold start, no relearning window where attacks slip through unscored.
A patient attack escalates privileges over hours or days, one quiet step per run. A fresh process forgets the earlier steps; Tack Cloud remembers them. Detection windows span 24h, 72h, and 7d, so slow credential-creep is scored against the whole history, not one session.
An agent's ID is set by the caller, so an attacker can spread malicious steps across many agent IDs to stay under any single session's radar. Tack Cloud makes your account the trust boundary and correlates behavior across every agent ID, so splitting the attack no longer hides it. This is the one detection a local, in-memory guard cannot do.
HOW IT FITS
The verdict is computed locally and returned before anything is sent anywhere. Tack Cloud receives events after the fact, so it adds memory without adding latency, and an outage cannot block your agent.
ON THE ROADMAP
Persistent, cross-agent baselines ship first because they are the part nothing else can give you. A dashboard, alerts, and an LLM analyst follow, driven by what early users actually ask for.
See every agent, its threat score over time, and the events behind a quarantine.
Push a detection to email, a webhook, or Slack the moment an agent drifts.
LLM-assisted forensics that explains a correlated incident in plain language, not just a score.
PRICING
tack-guard stays free and MIT licensed, with zero runtime dependencies, for as long as it exists. Tack Cloud is a simple paid tier layered on top for teams that need baselines to persist and correlate. Final pricing lands at launch.
FAQ