Cohesive cognitive layer
Bundle persona prompts, context memory, and linguistic preferences into a single mind that can be reused across products.
AgentOS powers Frame.dev's voice-first assistants. Now the modular runtime—complete with GMIs, workflow engine, guardrail APIs, and streaming orchestration—is available for teams who want to ship adaptive agents under Apache-friendly terms.
Realtime session
LivePersona
“I spotted a tool loop. Escalating to code interpreter with a capped energy budget and memoising your last three diagrams.”
Streaming cost
$0.021
Session budget 8%
Toolchain
3 calls
Planner + Designer + Search
GMIs bundle persona knowledge, memory, tools, and guardrails into composable building blocks. Instead of wiring ad-hoc prompts, you assemble agencies from reusable minds that understand context, entitlements, language, and policy before a request ever reaches the model. Input, translation, and output preferences are negotiated automatically so the same GMI can reason fluently across conversations, locales, and channels.

Bundle persona prompts, context memory, and linguistic preferences into a single mind that can be reused across products.
Every GMI carries guardrail hooks, subscription entitlements, and rate budgets—enforcing what the mind is allowed to say or do.
Drop GMIs into multi-role workflows. They collaborate, hand off tasks, and stream telemetry without bespoke glue code.
Auto-detect the active language from conversation context, pivot content when needed, and respond in the persona’s preferred voice—all before the request ever reaches the model.
Overlay new behaviours, capture lineage, and export minds as JSON so you can audit, rollback, or run them in other environments.
AgentOS brings the GMI manager, conversation memory, tool orchestrator, guardrail dispatcher, streaming manager, and workflow engine together behind a single facade. Configure providers, packs, and guardrails with TypeScript instead of glue code.

AgentOSConfig at a glance
const config: AgentOSConfig = {
gmiManagerConfig: { personaLoaderConfig: { loaderType: "file_system" } },
toolOrchestratorConfig: { orchestratorId: "default" },
guardrailService,
languageConfig: { defaultLanguage: "en", autoDetect: true },
workflowEngineConfig: { maxConcurrentWorkflows: 32 },
extensionManifest: { packs: ["@framers/agentos-pack-defaults"] }
};Operational telemetry
AgentOS is moving from MIT to Apache 2.0 so you get explicit patent grants and clear attribution requirements. You control your personas, tool packs, and marketplace listings—export them as JSON or load them into your own deployments at any time.
Yes. Export personas, prompt packs, and tool registries as versioned JSON. Import them into another AgentOS deployment—or any Apache-compatible fork—without losing history.
Core packages will be relicensed to Apache 2.0 during the public mirror rollout. Existing MIT releases remain available; new versions will carry Apache terms and upgraded NOTICE files.
Yes. Use the vca.chat marketplace to monetise personas or packs, or self-host your own catalog. Apache licensing ensures your commercial rights remain intact.
Pair the runtime with the Voice Chat Assistant demo, AgentOS client workbench, and marketplace-ready personas to understand how GMIs stream decisions, artifacts, and guardrail outcomes in real products.
AgentOS in Action
Media coming soon
Experience real-time streaming with adaptive personas that understand context
Register tools, guardrails, and workflows through extension manifests. Permission tags, rate budgets, and retries are built in.
Auto language detection, persona-aware response preferences, and pluggable translation providers keep GMIs fluent across every surface.
Swap between Postgres, better-sqlite3, sql.js, or custom stores. Use the reference server template to deploy anywhere.
Run the full AgentOS runtime inside the browser with SQL-backed persistence, workflow telemetry, and marketplace personas—perfect for offline prototyping.
Create agents, set your price, and earn from every sale. Join thousands of developers building the future of AI.
Discover pre-built agents ready to deploy, or create your own and start earning.
by Frame.dev
Enterprise-grade system design and code review with deep technical expertise
by Community
AI-powered creative writing and ideation assistant for content creators
by DataCraft
Advanced data analysis, visualization, and insights generation
by EduTech Labs
Adaptive language learning with conversation practice in 20+ languages
by SalesForce AI
CRM integration, lead qualification, and automated follow-ups
by HealthTech
Personalized health, fitness, and mental wellness guidance
The workflow engine spins up Agencies composed of Generalised Mind Instances, applies guardrail policies, streams progress, and hands off to humans when needed. Use it to run research pods, code reviewers, or customer onboarding flows.
Available today
Async generators stream AgentOSResponse chunks with rich metadata for text, tool calls, artifacts, and guardrail outcomes—ready for SSE, WebSocket, or in-memory bridges.
Developer preview
Persona overlays, workflow definitions, and Agency registries let specialised GMIs collaborate on shared goals with human-in-the-loop checkpoints.
Roadmap
Managed streaming, observability, and billing integrations for production teams that want Frame's infrastructure without the ops burden.
“AgentOS transformed how we build conversational AI. The streaming architecture and persona system are game-changers.”
Sarah Chen
Senior AI Engineer at TechCorp
“We deployed 50+ custom agents in just 2 weeks. The tool orchestration saved us months of development.”
Marcus Rodriguez
CTO at AI Startup
“The guardrail system and subscription-aware limits gave us confidence to go to production quickly.”
Emily Watson
Product Manager at Enterprise Co
Browse the TypeDoc site, explore integration guides, or regenerate the API reference locally. Every exported interface ships with rich TSDoc annotations covering streaming invariants, guardrail semantics, and error contracts.
pnpm add @framers/agentos
import { AgentOS } from "@framers/agentos";
const agentos = new AgentOS();
await agentos.initialize(config);
for await (const chunk of agentos.processRequest(input)) {
// handle AgentOSResponseChunkType.* events
}Join the community newsletter for release notes, roadmap updates, and deep dives on multi-GMI workflows.
AgentOS is open source and MIT licensed. Join the early access list to collaborate with the Frame team, migrate your existing assistant, and shape the roadmap.
By subscribing you agree to receive product updates from Frame.dev. No spam, pinky promise.
A product by Frame.dev