Agent Stream inspects and governs every LLM call your apps and agents make: blocking prompt injection and data leaks, capping what each team can spend, and routing to the right model, without changing a line of agent code.
OWASP Prompt injection is the #1 risk on the OWASP Top 10 for LLM apps, in both the 2023–24 and 2025 editions. Agent Stream is built to shut it down.
Most gateways bolt on “PII sanitisation and toxicity filtering”. Agent Stream runs a layered, fail-closed safety pipeline: Prompt Guard plus a Judge, Jury and expert-witness council, on every call.
Point your existing OpenAI SDK at Agent Stream instead of the provider. No SDK swap, no agent rewrite. It behaves exactly as before, just governed.
Guardrails, budgets, routing and audit live in one layer, so there’s no separate safety tool wired to a separate cost tool wired to a separate proxy.
Watch every model call ride the governed pipeline, track spend across the estate, follow calls live, and see exactly why each one was blocked. The whole story, in one place.
From shutting down the OWASP top risks and capping team spend to reliability, editor-native models and document-grounded chat.
Apply one auditable guardrail across every model and team, fail-closed by default.
A defensible, provable AI safety posture.
Cap and allocate AI spend across every team, member and provider.
Predictable spend, no bill surprises.
Let model calls degrade gracefully with no per-service work.
Graceful degradation, on your terms.
Roll AI coding assistants out to the whole org through corporate SSO.
A paved road to AI in the editor.
Ground answers on sensitive knowledge without handing it to a model.
Expose the chat, not the corpus.
Give security and finance one audit trail for every AI call across the org.
Answers when compliance comes asking.
Agent Stream sits between your apps, agents and editors and the models they call, so every request is inspected, governed and recorded on the way through.
Point your OpenAI SDK at a surface route instead of the provider. Safety, cost and resilience apply centrally, and your app and agents stay exactly as they are.
Affinidi hosted by default, with the same governance for every call. Self-hosted? Talk to us
Change a policy, budget or provider and it takes effect with no restart and no dropped requests. Governance you can tune live.
Keep production AI behaviour measurable, reviewable and under control.
Defence in depth, fail-closed by default, aligned to the OWASP LLM Top 10. A deterministic guard on both seams, wrapped around a review council that gates every call.
Deterministic regex and PII inspection on both the request and response seams, with an optional model-backed PII pass. Redact-and-continue or block outright, so sensitive data never reaches a provider.
OWASP LLM02 · Sensitive Information DisclosureAn Expert Witness gates both seams, a Judge pre-checks each request before the model runs, and a Jury votes on every response. Fail-closed by default, so a guardrail outage blocks rather than leaks.
Every block is classified (prompt injection, PII, toxicity, policy violation) and charted per team, member and surface. Show exactly what was caught, and why.
Every call is metered per stage and attributed to a team and a member, so AI spend is capped, allocated and predictable.
Monthly USD and token ceilings per team, member or surface. Hit the cap, and requests stop.
See who spent what, on which model, and export a chargeback CSV straight to finance.
A router sends cheap requests to a small model and hard ones to a frontier model, automatically. Right-size every call.
Balance across a pool of providers by lowest cost or lowest latency. No code changes.
Cost-spike, token and error-rate alerts fire before the monthly bill can surprise you.
Beyond safety and cost, every model call carries identity, resilience and full visibility.
Tie each request to a person in Microsoft Entra or Okta: per-user budgets, group policy and entitlement-filtered model menus. Identity, not just an API key.
Cross-provider failover, weighted / lowest-latency / lowest-cost load balancing, retries and circuit breakers. Model calls degrade gracefully, per surface.
Members and teams are full virtual keys: model allow-lists, per-minute rate limits and monthly caps. Spend and blocks roll up across every surface for chargeback.
Live spend, token, latency and throughput dashboards, rejection analytics, and native export to Langfuse, OpenTelemetry and Prometheus.
The Agent Stream for VS Code extension brings a governed, SSO-backed model menu right into the editor, alongside GitHub Copilot Chat and any OpenAI-compatible agent. Only the models a developer is entitled to, no key in the editor.
Front Microsoft 365 Copilot from inside the Agent Stream for VS Code extension: pull in your own emails and documents mid-conversation, grounded under your Entra ACLs. Only the answer crosses back, the agent never sees the corpus.
OpenAI, Anthropic, Gemini, Bedrock, and Ollama through a single endpoint. Switch providers or split traffic from the dashboard, no redeployment needed.
Fine-grained, context-aware allow/deny through an embedded Rego engine, evaluated in-process at the global and per-surface level. No sidecar.
Chat, embeddings, images, audio and batch all ride the same pipeline, so safety, cost and attribution apply to your whole model footprint, not just chat.
Gate audio input, transcription and realtime voice by model capability and policy — the same enforcement that covers chat now reaches every voice interaction.
Document-grounded chat requires a per-user Microsoft 365 Copilot licence and Entra admin setup, is text-only in this release, and is built on Microsoft’s beta Copilot API.
Agent Stream pairs with two companion products that extend the same identity, policy and audit model end to end.
A unified control layer that governs AI usage across every agent interaction.
Coming soon Where AI transactions are governed, verified and settled in real time.
Bring safety, cost control and full attribution to every LLM call. Deploy on the Affinidi Portal.