REALITY KERNEL BETA
Execution-layer security
Every plan runs the same engine, returns the same signed verdicts, and writes to the same tamper-evident ledger. Tiers differ in volume, session intelligence, and where the sensor plane lives.
For engineers evaluating execution-layer controls in a single agent or framework.
POST /v1/check/v1/pubkeyFor developers testing agent security against real traffic before a production rollout.
For teams shipping agents to production who need sequence-aware detection and policy enforcement.
allowed_tools, allowed_egress enforced pre-simulationPOST /v1/tokenPOST /v1/scan, up to 50 commands, gate on policy_passFor regulated environments that require kernel-level attestation and evidence that survives an audit.
rk-sensor — on-premises eBPF kernel sensor (Rust · Linux ≥ 5.8)The engine decides the cost from the shape of the command, not from your plan. Read-only verbs and allow-listed binaries never pay the simulation price. Every response returns the exact charge in credits_consumed and your balance in credits_remaining.
| Operation | Cost | Trigger |
|---|---|---|
| Fast-path check | 1 credit | Command matches the read-only fast path (ls, cat, git status, …). Returns in under a millisecond. |
| Shadow-world simulation | 5 credits | Command has side effects — filesystem writes, egress, privilege changes, interpreters. Projected across 5 parallel trajectories. |
| Policy violation | 1 credit | Least-agency policy rejects the command before simulation. Signed BLOCK at confidence 1.0. |
| Batch scan entry | 1 or 5 | Each entry in POST /v1/scan is priced exactly as an individual check. |
| Operator override | 0 credits | POST /v1/override appends a signed WARN_APPROVED / WARN_REJECTED row. Never billed. |
| Idempotent replay | 0 credits | Repeat a request with the same Idempotency-Key within 5 minutes and the cached verdict is returned free. |
| Public key · audit read · token mint | 0 credits | GET /v1/pubkey, GET /v1/audit, POST /v1/token, and GET /v1/me are metadata operations. |
X-RK-Credits-Low: true and a human-readable X-RK-Credits-Warning header so your client can alert before a 402 ever occurs.Roughly 500 fast-path checks, 100 full shadow-world simulations, or any mix. In a typical agent workload where ~95% of tool calls are read-only, 500 credits covers several hundred real agent actions — enough to wire up a framework, run your test suite, and verify signatures end-to-end. Credits do not expire while the key is active.
A tenant key signs into a shared audit ledger and consumes real compute. Automated domain verification lets us issue keys to engineering teams within hours while keeping the ledger free of throwaway identities. Requests from free-mail domains are reviewed manually and may take slightly longer.
No — it is 100,000 checks at the blended fast-path/simulation mix observed across production tenants, which we provision as a credit pool with headroom. If your workload is unusually simulation-heavy, the console shows your live burn rate and we right-size the pool with you rather than surprise you with overage.
The API returns 402 Insufficient credits and writes nothing to the ledger. It does not fail open. Your integration should treat a 402 the same way it treats a BLOCK — halt and escalate — which is why the SDK reference clients do exactly that by default.
The audit row stores the command and stated intent so that auditors can reconstruct the decision, with per-tenant retention you control from the console (retention_days). Enterprise deployments in a dedicated VPC keep the ledger entirely inside your boundary.
Platform vendors embedding Reality Kernel into their own agent runtime receive co-engineering time, roadmap input on the sensor plane, white-labelled verifier tooling, and preferential terms. It is scoped per engagement — book an architecture review to start the conversation.
Wire up your first agent tonight. Move to Starter or Professional when you're ready for sequence detection and CI gates.