Skip to content
AI proxy & router

The request layer is where the truth is

Two front doors, one pipeline. An on-device agent and a cloud proxy see the same AI traffic, tie every request to a person and a project, and apply policy before the request reaches the provider. Invoices tell you what you spent. The wire tells you why.

The problem

The invoice arrives a month after the decisions were made

Every IC is now a capital allocator. They pick the model, they set the context, they run the loop. By the time the provider invoice lands, thousands of those decisions have already been made and none of them carry a project code.

The only place spend, identity and intent exist together is the request itself. That is where Flowstate sits.

Capture

Two front doors, one pipeline

Both produce identical telemetry, so the dashboard never cares which door the traffic came through.

On-device agent

For the machines you manage

Runs on macOS today. Rolled out zero-touch through your MDM, so the developer is never prompted and never configures anything. It watches AI traffic on the machine and nothing else.

Cloud proxy

For everything else

Point tools, CI jobs and servers at proxy.flowstate.inc and the same pipeline picks them up. Same attribution, same policy, same telemetry.

Only AI traffic. Both doors funnel traffic against an explicit server-driven allowlist of AI hostnames, polled hourly. This is never a general internet proxy, and new providers ship without a new binary.

Attribution

Every request carries a name and lands on a project

The person rides standard proxy auth on every request, so identity is not inferred, it is carried. Sessions resolve to a project through the repo, branch and ticket references in the work itself, and from project on to initiative and cost center.

That is the difference between a bill and a ledger. The invoice says $98,000 went to one provider. The wire says which projects it built.

One request, resolved
Request POST /v1/messages
Person Sarah Chen, via proxy auth
Project Payments refresh, from PAY-412
Initiative Checkout modernization
Cost center CC-2140
Enforcement

Policy runs inline, before the provider sees the request

Four outcomes, decided on the wire. Observe-only is the default, enforcement is something you opt into when you trust the data.

Allow

The request passes through untouched. Telemetry is captured either way.

Warn

The request goes through and the person sees why it was flagged.

Coerce

The request is rewritten to comply, most often by routing to an approved model tier.

Block

A structured response links the policy that fired and an exemption request. Never a silent failure.

Smart routing

Route rather than block when routing is honest

Policy can rewrite the model on the wire. Rules are written against tiers, frontier, workhorse and lighter, not against vendor SKUs, so they survive every model release.

The saving is not estimated. It is measured token by token from the price delta between the model requested and the model served, and it shows up on the dashboard as money.

Routed this month
Frontier requested, workhorse served 41,209 requests
Workhorse requested, lighter served 12,877 requests
Measured saving, price delta $31,480
Speed and resilience

Fast enough that nobody notices, safe enough that nobody minds

If Flowstate is ever unreachable, the pipeline fails open. AI tools keep working and telemetry buffers locally until the connection returns.

<10ms
Added latency at p50
<50ms
Added latency at p99
~10s
Capture to dashboard
~5s
Policy change propagation
24h
Local telemetry buffer if offline
Keys

Production keys are reviewed, not assumed

New production keys start in a pending state and require admin review before they carry traffic. When something goes wrong, revocation runs at the right blast radius: one key, one device, or a whole key generation at once.

Privacy

Built to know less over time

Vendor credentials are never stored. They are forwarded and redacted from every capture. Prompt and response capture is configurable by your org, and raw content is discarded on a schedule, 30 days by default and configurable shorter, leaving only signals and aggregates. Sessions marked personal never carry text detail. The category alone is the record.

Put attribution and policy on the wire

Book a demo and watch a request resolve to a person, a project and a cost center in about 10 seconds.