Governed AI workflows
This now lives under Govern. See the pillar page →

Governed AI Workflows

AI agents that can call real tools — email, calendar, databases, APIs — need a runtime where governance is the runtime, not a bolt-on. Aarmos is that runtime: you build and run agents on-device, and governed tool calls are policy-gated before they fire. Proof falls out the other side as a signed receipt.

Governance isn't a stage — it's the runtime

Governed paths have no unpolicied execution route: every call the agent proposes on them passes through the same on-device gate, with rules you own and a receipt you can read. The mechanisms below are what the runtime is made of.

  • Policy you write, not a vendor default. Set per-tool rules — allow, ask, or deny — for every connector. Import an admin-signed governance policy to pin allowed tools and OAuth scopes across a fleet.
  • Consent before execution. The gateway surfaces the exact tool call in plain words before it fires. Approve once, for the session, or for 24 hours. Defaults are conservative — writes, sends, deletes, and payments require explicit approval.
  • Signed Receipts, not log files. Every decision is appended to an on-device chain where each Receipt links to the one before it. On Evaluation and above, full run traces are signed per device and exportable as JSONL — the evidence you can hand an auditor.
  • Break-glass kill switch. One tap suspends every outbound tool call and locks the vault. In-flight sessions abort, and the disabled state persists across reloads until an operator or a fresh signed governance policy explicitly resumes.

How it works in practice

1

Agent proposes a tool call

The LLM picks a tool — send_email, update_calendar, query_database — and the gateway intercepts it before any packet leaves the device.

2

Gateway checks policy

The call is matched against your rules, any active signed org governance policy, and the tool's capability class from the 8-verb model (read, write, delete, execute, communicate, delegate, transact, control).

3

Operator sees the call

You get the tool name, arguments, and destination in plain words. One tap to allow, ask, or deny. Session overrides are time-boxed and logged.

4

Receipt is written

The decision and the full trace are written to the on-device ledger as a chained Receipt — signed for off-device verification.

Local-first governance

The policy engine, vault, audit ledger, and kill switch all run in your browser. Your rules are enforced on-device, so governance does not depend on network reachability or a vendor's uptime. Admin-signed governance policies are verified locally with WebCrypto and cached with a 7-day grace period so offline devices keep enforcing the last-valid rules.

For teams, a centralised audit relay aggregates evidence packages from enrolled devices so your SIEM sees a unified stream. Device signing keys stay on the device — the relay stores and forwards signed bundles; it does not sign them.

Honest scope

Aarmos governs tool calls routed through its gateway. It does not audit the internals of third-party MCP servers or LLM providers, and it cannot protect against a compromised device or malicious browser extension. Read the full scope on the threat model.