Deployment & sovereignty proof
Every other page on this site makes a sovereignty claim. This is the page that proves it. CyberCLI is a single binary that runs entirely inside your network — no daemon, no agent fleet, no control plane we operate. Below is the whole data path, the complete list of what can ever cross your boundary, and how the audit chain, credentials, and licensing actually work.
Signals flow left to right. Everything inside the dashed boundary runs on hardware you control — air-gappable if you run the Local lane.
Other SOC tools queue alerts behind their AI. If the AI is slow, your operator waits — and the dashboard sits stuck at the elevated posture for hours after the threat passed. CyberCLI flips that: the operator is the highest-trust authority on the box, and a human verdict overrules AI math the moment it lands.
One click. The signal drops out of CYCON math immediately. A matching suppression rule auto-fires so the same pattern won't re-elevate. The raw archive still has the line for forensics.
The signal stays in your view + in CYCON math, but no escalation. Useful for "I see it, I'm aware, no Case needed yet." No AI tokens burned past the initial triage.
Opens a Case, locks the signal's CYCON contribution until the case closes. The Trust Ladder takes over from here — the operator can authorize action on the case the moment they decide.
Every Ignore / Monitor / Investigate writes a TriageDecision row with actor=operator, trust_level_used=4. The v1.5 local Warden trains on those rows — your operator's calibration becomes your model's calibration. No vendor sees the labels. The longer you run, the more your local AI matches how YOU read the threat picture.
Most SOCs run a single shared ingest queue. A chatty Suricata sensor under a port scan can bury a Wazuh auth alert for hours. CyberCLI is built so a noisy connector saturates its own queue (and sheds bounded to dead-letter) — never starves another connector's path.
| Stage | Bound | When it saturates |
|---|---|---|
| Per-connector queue | 10K frames each | Newest frame sheds to dead-letter (Vector raw archive still has it). Other connector queues unaffected. Round-robin drain across connectors guarantees fairness. |
| Warden admission gate | 3 in-flight (Pro) | New arrivals route to passthrough lane immediately — no 14-min queue join. The signal still lands in your dashboard at raw severity; only the AI narrative is skipped. |
| Herald severity floor | high+ default | Skips the layman-narrative pass on medium events. Warden's structured digest still runs. Operators with AI budget can lower via CYBERCLI_HERALD_MIN_SEVERITY. |
| Cgroup memory ceiling | 3G / 4G | Kernel throttles past MemoryHigh; OOM-kills + clean restart in ~5s past MemoryMax. StartLimitBurst=3 surfaces silent failures instead of hiding restart loops. |
Live state: GET /v1/system/ingest-health returns per-connector queue depth, drop counts, Warden bypass count, and a derived pressure badge (ok · throttled · critical · failed). The dashboard health pill reads this directly.
The honest version. Two things never leave; the rest only leave because you wired them, to destinations you own.
| Data | Crosses your boundary? | To where |
|---|---|---|
| Your logs & events | Never | Never. Parsed, scored, and stored on your box. |
| Your data — logs, events, tools, config | Never | Never. None of it leaves your box. Zero usage telemetry, zero diagnostics, zero phone-home of your SOC. |
| License check to CyberCLI | Only if you connect it | A periodic signed check — your license id, a nonce, a timestamp. Nothing about your data or usage. It's how renewals and tier upgrades reach the box automatically. |
| Vendor cloud APIs (Pro) | Only if you connect it | Outbound read only, and only for the Pro connectors you choose to connect (M365, Okta, EDR…). Your call. |
| Frontier AI (optional) | Only if you connect it | Off by default. Only if you bring your own key — then it goes to your provider on your bill, never through us. |
| Operator notifications | Only if you connect it | To the channels you configure (Slack, email, webhook). You own the destinations. |
| License check | Only if you connect it | A periodic signed check, offline-capable with a grace window. No usage data attached. |
"No CyberCLI cloud" means exactly this: we operate nothing your data passes through. Connecting a Pro cloud connector is your decision and an outbound read to that vendor — not a route through us.
The one outbound call CyberCLI ever makes on its own. Optional, monthly, soft. Sends a license id and a random nonce. Nothing else. Receives back a refreshed license JWS if your subscription state changed since the last check. If the call fails, your install keeps running on the token it has until it expires. You can turn it off entirely.
POST https://cybercli.com/api/license/heartbeat
Content-Type: application/jose+json
{
"alg": "EdDSA",
"kid": "<your license id>",
"payload": {
"license_id": "lic_a4b9...",
"nonce": "<32 bytes random>",
"ts": "<UTC ISO timestamp>"
},
"signature": "<EdDSA over payload>"
} That's the whole request. License id + nonce + timestamp, signed with the license's own private key (which never leaves the box). The signature proves you have the license; the nonce stops replay. No event count, no endpoint count, no host names, no IPs, no installed-version, no anything else.
We see your license id. We know you exist as a paying customer. That is the entirety of what the heartbeat reveals.
expires_at. No countdown, no nag, no degraded mode.license_heartbeat = "off" in $CYBERCLI_CONFIG_DIR/config.tomlcybercli license install <path> whenever you want to refresh. Standard air-gap pattern; a dedicated cybercli license heartbeat off subcommand ships in v1.x.The contract
We do not now and will not later add anything else to the heartbeat payload. Adding a new field is a breaking change to the contract on this page — it ships as a major version bump, a public diff, and an opt-in toggle. Any usage telemetry would be a separate, explicitly named feature you turn on. The default install never sends it.
The Guild's reasoning runs in one of three lanes, set per deployment. Sovereignty and cost move together; you choose where to sit.
Ollama on your hardware, auto-detected. Abliterated open models with no provider guardrails — the SOC reasons about real attacker tradecraft without a content filter refusing. Fully sovereign: nothing leaves the box.
Default lane. Free tier runs here.
Bring your own Claude / GPT / Gemini key for the hardest triage. The request goes straight from your box to your provider — CyberCLI is never in the path and never sees your key beyond the local vault.
Opt-in. BYOK, off by default.
Local models do the high-volume triage; only the ambiguous cases escalate to a frontier model. Best cost-to-quality for teams that want frontier judgment without a frontier-sized bill.
Recommended for most paid deployments.
| Lane | Case payload to a frontier provider | Telemetry to CyberCLI | License heartbeat (1×/mo) |
|---|---|---|---|
| Local | Never | Never | Optional · license id + nonce only |
| Hybrid | Only ambiguous cases · routed straight to your provider, on your key | Never | Optional · same payload as Local |
| Frontier | Every case · routed straight to your provider, on your key | Never | Optional · same payload as Local |
Sovereignty is only real if you can prove what happened and trust where your keys live. These hold regardless of which AI lane or Trust Level you run.
Every action — proposed, approved, executed, rolled back — appends a row to a hash-linked log. Change any row and the chain root changes. You can prove what happened to a forensic investigator without trusting CyberCLI.
Connector secrets live in a local vault, scoped per connector, never written to logs and never transmitted to us. A compromised connector can't read another's keys.
The chain is append-only and replayable: re-run the decision history against a case and get the same result. No hidden state, no silent mutation.
Every audit row records which connector emitted the signal and which executed any action — including community connectors. Forensic isolation is a property of the chain, not the dashboard.
The full leash — eight calibrated levels of AI autonomy per action — lives on the Trust page.
src/cybercli/audit/entry.py {
"schema_version": "1.0.0",
"seq": 41827,
"prev_hash": "9b1e3a…c712",
"entry_hash": "f4d9c0…2a83",
"decision_id": "dec_01HX9...",
"decision_packet_ref": "case://2026-05-31/117/decision.json",
"idempotency_key": "isolate_host:srv-id-01:dec_01HX9",
"initiator": {
"type": "ai",
"agent_role": "warden",
"model_route": "local"
},
"action": {
"type": "triage",
"action_template_id": "warden.triage.v1",
"scope": "single",
"parameters": {"signal_id": "sig_01HX9", "severity": "high"}
},
"outcome": "approved_by_human",
"trust_level": 3,
"cycon": 3,
"ts": "2026-05-31T14:28:11.412Z"
} entry_hash = sha256(canonical_json(this row with entry_hash blanked)). The next row's prev_hash equals this row's entry_hash. Genesis = 64 zeros.
# the chain lives at $CYBERCLI_DATA_DIR/audit.db. # Three ways to verify: # # 1. CLI walker (read-only, exits 0/1/2): # cybercli audit verify # cybercli audit verify --from 1 --to $(cybercli audit head) # # 2. Dashboard Audit tab: live tail + # row-level entry_hash recompute on view. # # 3. Roll your own — every entry's payload is # canonical JSON; sha256 it (with entry_hash # blanked) and compare. We publish the # canonicalization spec on this page.
"outcome": "approved_by_human" + initiator.type always identify who acted.entry_hash; row deletes break the next row's prev_hash.cybercli check injects a canned scenario through the live engine, watches CYCON elevate, asserts the AuthorityBridge caps over-ceiling actions, confirms no test event woke a real channel, and re-verifies the chain across the session. Generates a sharable check.html receipt — no marketing trust required.cybercli check — engine smoke test, ~3 seconds.
Injects a canned scenario through the live pipeline, walks six engine layers, and prints PASS/FAIL with the audit-row range for each. Every test signal is flagged is_test=True so the AuthorityBridge refuses real connector actions and the DeliveryPolicy returns an empty plan (no Slack/Pushover/PagerDuty fires). The audit-chain rows ARE written — those are your verifiable receipts.
$ cybercli check CyberCLI System Check · 2026-06-15T14:22:11Z L0 Beacons (CYCON elevation) ✓ PASS 1.3s CYCON 3 → 2 (5 audit rows · wazuh.suspicious_login) L1 DigestPipeline routing ✓ PASS 0.0s warden (triage), herald (notify) · tier=pro L2 AI provider lane ⚠ WARN 0.0s no AI provider configured — open the Dashboard's AI Engine page L3 AuthorityBridge gate ✓ PASS 0.1s 2/2 packets gated correctly · ceiling=L5 · above-ceiling capped 7→5 L4 Delivery suppression ✓ PASS 0.0s 25/25 plan() returned [] (is_test=True) L5 Audit chain verify ✓ PASS 0.3s 18 rows, all hashes match ━━━ PASS WITH WARNINGS ━━━ 18 test audit rows written. Run `cybercli audit verify --from 41828` to re-verify. (Free-tier first install shown above — L2 WARNs because no AI provider lane is selected yet. Pro buyers with Ollama or a frontier key see ✓ PASS.)
For sharing — MSP commissioning artifact, CISO follow-up, or a friend confirming the install works — generate the self-contained HTML report: cybercli check --report html -o check.html. Single file, no JS, no external assets.
An endpoint is a distinct host or device emitting telemetry. Free covers 10; Founder Pro / Pro cover 50; Business 51–500. You see your current count in the dashboard.
The license verifies with a periodic signed check and keeps working through a grace window if the box is air-gapped. No always-on connection required.
Buy now, activate later — the license is tied to your account, not a machine. Move it between boxes; re-activation is a sign-in away.
We ship breadth first, then depth on the connectors design partners actually run. Here's the honest line as of this build.
cybercli onboard <id> each.Wazuh — and Suricata, and your firewall — see. They produce alerts. CyberCLI is the layer that decides and acts: it correlates every source into one posture (CYCON), triages and verifies with the Guild, and executes containment only inside the leash you set (the Trust Ladder) — with a receipt for every step. Keep your detectors. CyberCLI orchestrates them into a SOC instead of a wall of alerts.
Your box. Your keys. Your audit chain.
We fight for the users.