by LockedIn Labs
DocumentationStart here

Documentation

The trusted path

ARC1 is the control plane for enterprise AI. It decides what AI use is allowed, puts that decision into every enforcement point a company runs, and keeps a signed record of what each one did.

A path you can name

Every model request and every agent action an enterprise permits should travel a path it can name in advance and prove afterwards. That is the trusted path: an approved policy, the enforcement points that apply it, and a chain of signed records showing what each of them decided.

ARC1 owns that path without asking a company to replace what already carries its traffic. It has two parts, deployed separately.

The hub and the gateway

The hub
The control plane. It holds policy, approvals, budgets, spend attribution and evidence; compiles approved policy into each target’s own format; and reads the result back. It stays outside model traffic: no prompt, reply or provider key passes through it.
The ARC1 Gateway
The optional data plane: an LLM proxy and an MCP gateway that run inline in the customer’s environment, once they are explicitly installed, enrolled and routed. It handles request content at its own boundary and sends the hub signed receipts that carry metadata only.
The targets
The gateways and agents a company already runs. ARC1 compiles the same approved policy into Kong, Azure API Management, LiteLLM and other gateways, and into managed settings for Claude Code and Codex.

Installing the hub neither activates the gateway nor replaces an existing one. A company can govern the gateways it has from the hub and decide later which routes, if any, run through the ARC1 Gateway.

Where a request goes

An application or coding agent calls a gateway, and the gateway calls the model. When that gateway is Kong, Azure API Management or LiteLLM, ARC1 has already compiled the approved policy into its configuration, and it reads the configuration back on a schedule and reports drift. Readback proves what the gateway holds. It does not prove that every request was enforced.

When a route runs through the ARC1 Gateway, the gateway applies the policy itself and signs a receipt for each decision: who asked, which model and connection served it, what the policy decided and how many tokens it cost. The hub verifies and keeps those receipts beside its own; The receipt describes both.

On managed machines, Claude Code and Codex read a signed policy the hub distributes, and the managed gate on each machine checks sensitive data before a prompt or tool call leaves it. Those checks consult policy; they do not send model traffic through the hub.

One deployment, in your environment

ARC1 is deployed once per organization, in the customer’s own environment, with the customer’s identity provider and PostgreSQL row-level security on every table of organization data. Nothing requires a connection to LockedIn Labs.

The hub is designed to hold metadata: counts, model identifiers, timestamps, salted hashes, policy versions and receipts. It does hold directory identity, work ownership and financial records, because attribution needs them. Opaque identifiers and hashes are not anonymization, and a metadata-only design does not by itself establish HIPAA compliance.

Spend keeps its basis. Provider-billed, gateway-measured and estimated figures are never added together, every total names its basis and window, and synthetic demonstration records carry the DEMO label and are never combined with measured data.

How we label evidence

Each page of this documentation ends with a table of where its parts stand, in one of three words: LIVE, LAB or NEXT. Each table defines them, and so does the documentation index.

Kong Gateway OSS 3.9.3 and LiteLLM Proxy 1.103.1 are lab-verified: apply, readback and drift recovery ran against those versions with synthetic traffic. Contract-tested means a connector was tested against the vendor’s documented interface without a live service. Customer-verified would mean acceptance in a customer’s own deployment, and none is recorded yet.

Where it stands

Where each part stands
CapabilityStateEvidence
Hub: policy, two-person approval, readback, drift, spend attribution and signed evidenceLIVEOn our hosted production deployment, with synthetic data.
Compile, apply and read back on Kong Gateway OSS 3.9.3 and LiteLLM Proxy 1.103.1LABLab-verified on 28 September and 13 October 2026 with synthetic traffic. The Kong run did not include model inference.
Azure API Management, Apigee, Amazon Bedrock Guardrails, Cloudflare AI Gateway and PortkeyLABContract-tested against each vendor’s documented interface.
ARC1 Gateway inline: the LLM proxy and the MCP gatewayLABBuilt and tested against a stand-in provider; no live provider acceptance is recorded.
Installation in a customer’s own Azure tenantLABBuilt, not yet run in a customer tenant.
Moving chosen routes onto the ARC1 GatewayNEXTThe direction of our next phase; existing gateway paths stay.

LIVE: in the released product and running on our hosted production deployment, with synthetic data. LAB: built or being built, and run only in our lab or against a vendor’s documented interface. NEXT: designed, not yet built. How we use these words.