[ security · local-first ]

Security by architecture, not policy.

The strongest privacy guarantee is the absence of a server in the middle. AgentiqFlow runs entirely on your machine, holds your keys in a sealed vault, and only touches the network when you call a model provider. Here's exactly how that's enforced.

[ network ] 01/06

Loopback by default

The sidecar binds to 127.0.0.1 — loopback only. The dashboard runs in your browser at localhost:8858. It is not reachable from your LAN unless you deliberately set HOST=0.0.0.0, and LAN exposure is an explicit opt-in, not the default.

127.0.0.1 · localhost:8858 · opt-in LAN
[ vault ] 02/06

Sealed AES-GCM vault

Provider keys and OAuth tokens are sealed with a passphrase you set — AES-GCM encryption, PBKDF2 at 200k iterations. Stored on your disk under ~/.agentiqflow. Lock and unlock from the UI. Secrets never leak to logs.

AES-GCM · PBKDF2 200k · lock/unlock
[ egress ] 03/06

Egress only on your call

Every layer runs on your machine. Outbound network traffic happens only when you call a model provider — and you choose which calls leave the box. There is no telemetry and no background analytics.

opt-in egress · no telemetry
[ supply chain ] 04/06

Signed installers & updates

Updates are verified against a minisign signature over the release manifest, plus a sha256 integrity check on every artifact. A build that doesn't match its signature is rejected before it's applied.

minisign · sha256 manifest
[ workflow http ] 05/06

SSRF-guarded outbound HTTP

HTTP calls made from workflows and integrations pass through an SSRF guard that blocks loopback, private, and link-local targets — including the cloud-metadata address at 169.254.169.254 — so a crafted URL can't be pointed at your internal network.

blocks private · link-local · metadata
[ keys ] 06/06

Your keys, no markup

You bring your own provider keys. The runtime calls the provider directly, at the provider's price — nothing is relayed through our servers and there is no token markup. Prefer no third party at all? Point an agent at a local Ollama model.

BYO keys · direct calls · local Ollama
[ control plane ]
  • State-changing control-plane requests are gated on Origin and Host — loopback is always allowed; extra hostnames (e.g. a Tailscale name) are opt-in via AGENTIQFLOW_ALLOWED_ORIGINS.
  • Optional multi-user mode uses scrypt password hashing, 7-day session tokens, and per-user audit trails.
  • An append-only JSONL audit log records every workflow run, git commit, login, and admin action — filterable, searchable, exportable.
  • Webhook receivers verify HMAC signatures in constant time before a payload is processed.

Want the full picture of what runs where? Read the self-host guide or see how the layers fit together on the capabilities overview.