Sovereign AI

Your models. Your data. Verifiable records.

When a team moves from a hosted AI API to open-source models on its own hardware, provider dashboards and usage logs may no longer cover those calls. AgentGuard runs inside the caller's perimeter and, when signing keys are configured, records the declared model, weights origin, and hosting jurisdiction in a signed receipt.

Govern a local model See a signed receipt
The sovereignty gap

Self-hosting gives you the weights. It takes away the controls.

A self-hosted endpoint does not automatically add the per-scope caps, declared capability checks, provenance fields, or signed decision records implemented by AgentGuard. Those controls run in the caller's process and remain under the operator's configuration.

The answer

A control plane that lives in your runtime, not ours.

AgentGuard Spend is an in-process TypeScript SDK. Governed provider calls do not pass through an AgentGuard inference gateway or proxy. Install it next to your agent, configure a supported provider binding and policy, and the binding evaluates that policy before dispatch. Signing is an additional configured path.

$ npm install @agentguard-run/spend
# Configure the OpenAI-compatible binding with your endpoint URL.

New to self-hosting? Start with the local serving and governance guide. Hardware selection and prices must be validated against your model, serving stack, and current vendor listings.

01

No AgentGuard inference proxy

For supported in-process bindings, governed-call prompts, completions, and provider credentials travel directly between your application and selected provider. Hosted account, license, telemetry, consent, workflow, and managed-access surfaces are separate.

02

Hard enforcement, locally

Spend caps, declared capability checks, workflow envelopes, and the opt-in circuit breaker run in-process before provider dispatch.

03

Model-agnostic by design

Hosted frontier APIs, open-source models on your GPUs, or both in one fleet. Swap models freely; the governance rail and the receipt format do not change.

04

Tamper-evident records

When signing keys and a decision-log store are configured, decision metadata is Ed25519-signed and hash-chained. A holder of the record and trusted public key can detect a signed-field edit offline.

Provenance, recorded

The receipt preserves what the policy recorded.

AgentGuard provenance records model identity, a registry-inferred weights-origin country, provider-route and jurisdiction metadata, and a configured or default retention posture. Known GLM, DeepSeek, Qwen, Kimi, and related model IDs are flagged as foreign-origin and can be refused under the applicable posture. Hosting, retention, BAA, and residency fields can include operator-supplied attestations; the signature protects the record from later editing but does not independently prove those real-world facts.

actionissue_credit · within cap
model_recordglm-4 · configured call
weights_originCN · registry inferred
jurisdictionself-hosted · default or configured
data_retention0 days · default or configured
decisionALLOWED · signed when configured
verifyindependently · off-platform

An independent reviewer can verify the integrity and chain position of a retained signed record. That check does not establish that the configured hosting or retention assertion was true in the outside world. Receipts remain content-free: they record decision and provenance metadata, not prompts or completions.

Who this is for

Built for the teams the frontier labs cannot serve neutrally.

Enterprises running open-weight models on-prem or in a VPC

AgentGuard adds configured caps, declared capability checks, provenance metadata, and optional signed decision records around supported calls.

Regulated industries reviewing residency and logging obligations

Financial services, healthcare, legal, and public-sector teams can retain jurisdiction and retention fields for internal or counsel-led review. A signed record is not a compliance determination.

Teams that keep governed content on their chosen provider path

Supported in-process bindings send governed provider traffic directly from the application to the selected endpoint without an AgentGuard inference proxy.

"You cannot rent intelligence from the same place that rents it to your competitor, and you cannot rent trust from the vendor whose model you are auditing. Own the weights. Own the data. Own the proof."