Govern every agent action — priced by blast-radius
Agents take real actions. ARK gates each one before it runs and emits a tamper-evident record of the whole trajectory. You pay per step, by how much each step can break — not a flat fee on every token.
Same prepaid credits as chat. 1 credit = $0.001. No new wallet, no new SKU. Point your agent at /v1/agent.
Tier 0 · Record
1 credit / step
Runs on every step — plans, observations, reads, clean writes.
- Blast-radius classification (read · write · irreversible)
- Hash-chained, tamper-evident audit write
- 0 model calls — fully deterministic
- This is the layer a regulator re-walks to verify the run
See it run →
Tier 1 · Grounding check
14 credits / step
Runs on a flagged reversible write whose arguments aren't grounded.
- One grounding check before the write commits
- Catches ungrounded write actions reversibly — not after
- 1 cheap model call (priced at our standard margin floor)
- Triggers only where it matters — most steps stay Tier 0
Start free
Tier 2 · Full suite
15–25 credits / step
Runs on irreversible actions and the final answer — non-waivable.
- 15 credits Standard · 25 Strict
- Claim-tiering + calibration + sycophancy/advisory + trust score
- Blocks an irreversible, ungrounded action before it executes
- Irreversible + final are always Tier 2 — no parameter downgrades them
Talk to us
Start in audit, not enforce. Run mode="audit" for the first week — ARK records and classifies every step and shows the "would-have-blocked" rate, but blocks nothing. Flip to mode="enforce" once your false-block rate is ~0. Auditing is metered the same; only the veto is held back, so you earn trust before anything is ever stopped.
How per-step metering works
Most steps are near-free. The expensive governance fires only on the steps that can actually cause harm — flagged writes, irreversible actions, and the final answer. The tier is assigned on our servers from the action's blast-radius; it is never set by your client.
| Step | Tier | What runs | Price |
|---|
| Plan · observation · read · clean write | Tier 0 | Classify + hash-chain audit (deterministic) | 1 cr |
| Flagged reversible write (ungrounded) | Tier 1 | One grounding check | 14 cr |
| Irreversible action · final answer | Tier 2 | Full governance suite + trust score | 15 / 25 cr |
Worked example — a 20-step Strict run (1 plan, 12 tool calls of which 2 irreversible, 6 observations, 1 final answer, 3 flagged writes):
Tier 0 (20 × 1) + Tier 1 (3 × 14) + Tier 2 (3 × 25) = 137 credits ≈ $0.137.
A $199 / 275,000-credit Pro plan covers roughly 2,000 runs like this a month. Running the full suite on every step would cost ~500 credits — risk-tiering is about 3.6× cheaper and margin-safe.
Bring your own LLM
Lower per step
Your key pays the model tokens; ARK bills a governance-service fee only.
- Highest margin for you — you already own the inference
- Governance fee per Tier 1 / Tier 2 step, not full credits
- Same gate, same tamper-evident audit
Contact sales
ARK Sovereign
$0 / step
Self-host the gateway on your own infrastructure — flat annual licence.
- No per-step metering — licence only
- Data never leaves your region (zero-egress) — for regulated PHI
- Bring your own LLM; unlimited runs
Contact sales
Design-partner pilot
Free · 30 days
Wire ARK into one real agent. Audit-first, then enforce. Metering on, billing waived.
- ~15 min to wire into a LangGraph / create_agent agent
- Prove it on your own workflow before you pay
- You see the exact credits each run would cost — live
Apply for a pilot