What Smartflow does, and who else is in the room

Runs inside your cloud or fully disconnected. FedRAMP / IL5 / GAIA-X SKUs, not a shared multi-tenant SaaS.

Sovereignty tab shows where data sits and what's stripped before it moves. Posture you can hand an examiner.

Evidence lands in your S3 / Azure / GCS / SnapLock with object-lock retention. Residency of the audit trail, not just live logs in our DB.

Pull enterprise content without routing file bodies through a third SaaS middleman.

Route sensitive traffic to weights on your own hardware so inference never has to leave the box.

Provider secrets never leave the gateway. Rotate or revoke without touching app code.

AD group maps to team, budget, and policy. Access changes flow from the directory, not a spreadsheet.

Force verified-identity re-auth on high-risk actions and seal it into the audit chain.

Establishes who the agent is, not just the human who launched it — carried on egress.

Ties a request to the laptop and the coding agent at the intercept point.

Every call carries user, model, cost, and risk tier. One record, not a metric and a log to reconcile.

Records are hash-chained, and an open-source verifier lets a third party check the chain independently.

Trusted-time proof plus write-once storage. Examiner-grade, not "we keep logs indefinitely."

1.10 stamps the control and policy that fired onto the record, so a finding maps straight to a rule.

The same record exports to Splunk, OTel, and ServiceNow without re-deriving it per tool.

PII, secrets, and regulated patterns caught on the request path — not flagged after the fact.

HIPAA, GDPR, PCI and more attach by directory group, so the right rules apply to the right team automatically.

Opt-in tokenization keeps workflows usable so teams don't route around the gateway to get their data back.

Paste an acceptable-use doc and get discrete, checkable rules out — not a PDF nobody enforces.

Article-by-article evidence binder for a Notified Body — concrete conformity, not "audit-ready" blog copy.

Bind policy to a virtual key, AD group, team, or model. Enforced on the call, not documented in a wiki.

Hold risky actions for human sign-off, or flip to read-only / halt when something's wrong.

Scores your own traffic and suggests rules. Nothing leaves the box to a third-party test service.

Finding flows into ServiceNow / Archer / OneTrust and the closure flows back — not a one-way alert.

Unsigned or unapproved MCP servers don't get a tool call. Supply-chain control for agent tools.

Cost by provider, model, user, and group from the record the gateway already enforces. No second pipeline.

Forecast the overrun and stage a policy in response — not an email that lands after the money's gone.

Surfaces ungoverned spend — traffic that bypassed the gateway — by comparing provider bills to what we metered.

A spend spike turns into a proposed rule you can review and enforce, closing the loop on runaway cost.

Cache and token savings shown against actual spend — enforcement plus visibility, not invoice-only.

Point the base URL at the gateway. No app rewrite for either endpoint style.

Retryable vs non-retryable failures handled distinctly, so a provider blip doesn't take the app down.

Send group X to a cheaper model, or a regulated team to a compliant one — as a rule, not a code branch.

Self-hosted and cloud targets are first-class, so sovereignty routing is a config choice.

Scope which models a key can reach, so access control and routing are the same control.

Easy turns go cheap, hard turns stay on the model you asked for. Same scorer in Smartflow and Halo. Install notes.

Semantic hits without a sidecar vector DB to run and secure. One less dependency in the boundary.

Compress context on the way through, with an optional faithfulness check so you don't cache a hallucination.

Uses the provider's own prompt-cache when it's cheaper, so you don't pay twice for the same prefix.

Caches tool discovery and repeat calls, and reports the measured percentage of calls avoided.

Reuses results across repeated multi-step agent tasks so the same plan doesn't re-run from scratch.