Amazon Bedrock AgentCore enables you to deploy and operate AI agents securely at scale using any framework and model. AgentCore provides tools and capabilities to make agents more effective, purpose-built infrastructure to securely scale agents, and controls to operate trustworthy agents. This CLI helps you create, develop locally, and deploy agents to AgentCore with minimal configuration.

• Node.js 20.x or later
• uv for Python agents (install)

Upgrading from the Bedrock AgentCore Starter Toolkit? If the old Python CLI is still installed, you'll see a warning after install asking you to uninstall it. Both CLIs use the agentcore command name, so having both can cause confusion. Uninstall the old one using whichever tool you originally used:

pip uninstall bedrock-agentcore-starter-toolkit    # if installed via pip
pipx uninstall bedrock-agentcore-starter-toolkit   # if installed via pipx
uv tool uninstall bedrock-agentcore-starter-toolkit # if installed via uv

npm install -g @aws/agentcore

Use the terminal UI to walk through all commands interactively, or run each command individually:

# Launch terminal UI
agentcore

# Create a new project (wizard guides you through agent setup)
agentcore create
cd my-project

# Test locally
agentcore dev

# Deploy to AWS
agentcore deploy

# Test deployed agent
agentcore invoke

Note: Run agentcore deploy after add or remove to update resources in AWS.

A harness bundles a runtime, model, tools, skills, memory, and observability into one declarative config. Use it when you want infra without writing agent code.

After export harness, read app/<agentName>/EXPORT_NOTES.md before running deploy — it lists any manual follow-up the exporter could not automate.

Create agents with --with-config-bundle to auto-wire config bundle support into the generated template.

See Knowledge Bases for ingestion, vectorization, and retrieval setup.

Failure-pattern analysis across agent sessions. Insights configs run continuously alongside online evals and surface clusters of bad outcomes.

Policy engines apply Cedar-based pre/post-call policies to agent invocations — including Bedrock content filters (VIOLENCE, HATE, SEXUAL, MISCONDUCT, INSULTS), prompt-attack detection, and sensitive-information redaction.

Pay-per-call agent transactions via the x402 protocol. When a tool call returns 402 Payment Required, the payments system signs and submits payment then retries automatically.

See Payments for the full setup including instrument creation and tool allowlists.

Curated session datasets for batch evaluation and recommendation runs.

Add a managed web-search target to a gateway:

agentcore add gateway-target --type connector --connector web-search \
  --gateway <gateway-name> --name <target-name>
# Optional: --exclude-domains "example.com,foo.org"

See Gateway for full target setup including Lambda, MCP, OpenAPI, Smithy, and API Gateway.

my-project/
├── agentcore/
│   ├── .env.local          # API keys (gitignored)
│   ├── agentcore.json      # Resource specifications
│   ├── aws-targets.json    # Deployment targets
│   └── cdk/                # CDK infrastructure
├── app/                    # Application code

├── app/                    # Application code
│   └── <AgentName>/        # Agent directory
│       ├── main.py         # Agent entry point
│       ├── pyproject.toml  # Python dependencies
│       └── model/          # Model configuration

Projects use JSON schema files in the agentcore/ directory:

• agentcore.json - Project resources (agents, memory, credentials, gateways, evaluators, online evals/insights, knowledge bases, harnesses, policy engines and policies, payment managers and connectors, config bundles, datasets, runtime endpoints)
• deployed-state.json - Runtime state in agentcore/.cli/ (auto-managed)
• aws-targets.json - Deployment targets (account, region)

• Harness - Declarative agent: bundle runtime + tools + skills + memory + observability without writing agent code
• Runtime - Managed execution environment for deployed agents
• Memory - Semantic, summarization, user-preference, and episodic strategies
• Credentials - Secure API key + OAuth credential management via Secrets Manager
• Gateways - MCP gateways with Lambda / MCP server / OpenAPI / Smithy / API Gateway / web-search / knowledge-base targets
• Evaluations - LLM-as-a-Judge for on-demand, batch, and continuous agent quality monitoring
• Recommendations - Auto-optimize system prompts and tool descriptions from real session traces
• A/B Tests - Traffic-split between config-bundle or target-based variants and promote the winner
• Insights [preview] - Failure-pattern analysis and clustering across agent sessions
• Knowledge Bases - Managed Bedrock Knowledge Bases auto-wired to gateways
• Policies & Guardrails - Cedar pre/post-call policies including Bedrock content filters, prompt-attack detection, and sensitive-information redaction
• Payments - x402-protocol microtransactions for pay-per-call tools and APIs
• Config Bundles - Versioned runtime configurations as a separately-deployable resource

Reference

• CLI Commands Reference - Full command reference for scripting and CI/CD
• Configuration - Schema reference for config files
• Frameworks - Supported frameworks and model providers
• PERMISSIONS - IAM permissions required to deploy

Resources & features

• Memory - Memory strategies and sharing
• Gateway - Gateway setup, targets, and authentication
• Knowledge Bases - Managed Bedrock Knowledge Bases wired to gateways
• Payments - x402-protocol microtransactions for paid tools/APIs
• Config Bundles - Versioned runtime configurations
• Container Builds - Container build types and Dockerfile setup

Evaluation & quality

• Evaluations - Evaluators, on-demand evals, and online monitoring
• Batch Evaluation - Run evaluators across sessions at scale
• Recommendations - Optimize prompts and tool descriptions
• A/B Tests - Split traffic between variants and promote the winner

Operations

• Local Development - Dev server and debugging
• Transaction Search - Trace + log search across agent invocations
• Telemetry - CLI usage telemetry — what's collected and how to opt out
• TUI Harness - Programmatic TUI driver for testing
• Testing - Unit, integration, and e2e test infrastructure
• Feedback - Submit feedback from your terminal

See the AgentCore Samples repository for end-to-end examples using the CLI, including multi-agent workflows, MCP gateway targets, and framework integrations.

Have a quick comment or suggestion? Send it from your terminal:

agentcore feedback "your message" --screenshot path/to/screenshot.png

The CLI will display the AWS Customer Agreement and prompt for consent before submitting. See docs/feedback.md for usage details.

For bugs, regressions, or feature requests that need discussion, open an issue on GitHub instead.

See SECURITY for reporting vulnerabilities and security information.

This project is licensed under the Apache-2.0 License.