Strands · AgentCore Own your AI platform, on your own cloud.
✱ why this exists
every vendor now agrees: don’t get locked in.
The AI labs are openly competing with their own biggest customers and partners, and even the platform vendors now coach enterprises to keep their AI stack swappable — while each one pitches its own ecosystem as the safe place to do it. We think the advice is right and the conclusion goes further than any vendor will take it: the only harness that’s truly separate from the model is one you can read, fork, and run yourself.
“Keep your harness separate from the model.”
— Satya Nadella, Microsoft CEO, on making any model swappable · TechCrunch, July 2026
The harness is yours.
The chat surface, agents, tools, and governance are open-source code deployed in your AWS account — infrastructure you control, not a subscription you rent.
Models are swappable.
Every provider Strands Agents supports — Bedrock by default, plus Anthropic, OpenAI, Gemini, Ollama, and more — switchable per conversation and per agent. When a better or cheaper model ships, you swap it that afternoon.
Leaving costs nothing.
No per-seat contracts, no proprietary formats, no ecosystem toll. Your data, prompts, and agents live inside a security boundary you already operate.
01 the conversation
chat that ships finished.
Not a demo loop — the whole conversational surface. Token-by-token streaming over SSE, artifacts rendered in a docked panel, voice in and out, and context that manages itself on long conversations instead of falling over.
- Streaming, end to end. Tool calls, interactive tool UIs, and errors all arrive as stream events — the UI never blocks on a spinner.
- Artifacts. Generated documents, code, and pages open in a side panel you can keep working next to.
- Voice mode. Speak to the assistant and hear it answer.
- Automatic compaction. Older turns roll into a summary when context runs long — surfaced in the UI, never silent.
- Conversations that survive anything. Refresh mid-stream, switch tabs, come back tomorrow — the session picks up where it was.
02 agents
build an agent once. your whole org runs it.
Agents bundle instructions, tools, skills, and knowledge bases into something anyone can use. Publish to the marketplace, and teammates pin it in one click — no setup guide, no terminal.
- Marketplace. Publish an agent and control who can find it — a team, a role, or everyone.
- Version snapshots. Every publish is a snapshot; roll back a bad release without touching the editor.
- @-mentions. Pull a second agent into the middle of a conversation when you need its tools.
- Collaborative editing. Share an agent with viewers and editors — ownership is an access gate, not a silo.
- Skills. Reusable instruction packs agents load on demand, governed by the same role system as tools and models.
03 tools & integrations
connect the systems you already run.
Four tool protocols, one registry: direct functions, AWS SDK calls, MCP servers behind AgentCore Gateway, and remote A2A agents. Admins register a server; role-based access decides who gets which tools.
- MCP Apps. Tools can render live, interactive UIs inline in the conversation — not just text results.
- OAuth built in. Tools that need user consent trigger it mid-conversation and resume where they left off.
- Per-tool enablement. Turn individual tools on or off within a server — expose exactly what each role needs.
- Gateway targets. Register Lambda-backed MCP targets from the admin console, secured with SigV4.
- A2A agents. Call out to remote agents running on their own AgentCore Runtimes.
04 knowledge
answers grounded in your documents.
Every agent can carry its own knowledge base. Upload documents, connect sources, and keep them fresh with scheduled re-sync — retrieval runs on Bedrock Knowledge Bases inside your account.
- Per-agent knowledge bases. Each agent retrieves from its own corpus — no cross-contamination between teams.
- Scheduled sync. Connected sources re-index on a schedule, so answers track the documents as they change.
- Citations in the stream. Retrieval results ride the same SSE stream as everything else.
- Export conversations. Save finished conversations out to connected storage.
05 governance & cost
governed by default, down to the token.
Access control and cost visibility aren’t bolt-ons. Roles map to your identity provider’s groups and gate every model, tool, and skill — and every model call is metered, priced, and attributable.
- RBAC from your IdP. Roles bind to identity-provider groups, with inheritance and wildcard grants.
- Delegated admin. Scope admin surfaces to the people who own them — models, tools, users, costs.
- Usage quotas. Per-user token budgets with in-stream warnings before a cutoff, not after.
- Cost analytics. Per-session, per-call cost breakdowns in the admin console.
- Prompt-cache observability. Cache hit/miss status and prefix fingerprints on every call — see exactly which change burned money.
06 your infrastructure
one stack. your AWS account.
Everything deploys as a single CDK stack plus three GitHub Actions workflows. No SaaS control plane, no per-seat licensing — your data never leaves the security boundary you already operate.
$ npx cdk deploy your-prefix-PlatformStack ✓ PlatformStack one stack, every AWS resource $ gh workflow run backend.yml ✓ app-api ECS behind the SPA ✓ inference-api AgentCore Runtime container ✓ rag-ingestion knowledge base pipeline ✓ artifact-render artifact rendering $ gh workflow run frontend-deploy.yml ✓ SPA → S3 + CloudFront Deployed. Your models, your data, your bill.
- Single CDK stack. Every resource lives in one
PlatformStack— no cross-stack ordering to babysit. - Zero-downtime backend deploys. Images ship out-of-band via ECR and rolling service updates.
- Bring your own models. Any provider Strands Agents supports — Bedrock, Anthropic, OpenAI, Gemini, local models, and more — behind an admin-curated catalog.
- Consumption pricing. Pay AWS for what your people actually use — not per seat.
frequently asked questions.
What is the AgentCore Public Stack?
A production-ready, multi-agent conversational AI platform built on AWS Bedrock AgentCore and Strands Agents — an Angular chat frontend, FastAPI backends, and a CDK-defined AWS platform, developed in the open by Boise State University.
Is this an official AWS product?
No — it’s the unofficial frontend for Strands and AgentCore: an independent open-source project that builds on those AWS services and tracks them closely.
Why not just buy Copilot or ChatGPT Enterprise?
Those are excellent products — rented from vendors who are increasingly competing with each other, repricing as they go, and coaching you not to get locked in while selling you their ecosystem. Renting the harness means renting the switching costs. Owning it means model choice, pricing, and roadmap stay decisions you make, not terms you accept.
Which models does it support?
The model layer is Strands Agents, so you’re not tied to any one vendor’s catalog. Amazon Bedrock is the default provider, and Strands also supports Anthropic, OpenAI, Google Gemini, Mistral, Ollama, SageMaker, LiteLLM, and more — plus custom providers for anything it doesn’t ship (see the full provider list). Admins curate the catalog, users switch per conversation, and role-based access controls who can use which model.
Where does my data live?
In your AWS account. Conversations, documents, and knowledge bases stay inside your own security boundary, and models are used for inference only — your data is never used to train third-party models.
What does it cost to run?
There’s no license — you pay AWS for consumption. Built-in quotas, cost analytics, and prompt-cache observability exist precisely so you can see and control that spend per user and per call.
How do I deploy it?
One CDK stack for infrastructure, three GitHub Actions workflows for the application. The deployment guide walks through it end to end.
run your own. tonight.
Clone the repo, deploy the stack, and hand your whole org an AI platform you control.