Skip to main content

Hosted IDE Agents on AgentCore

Already using Claude Code, Kiro, or Codex on your laptop? Run them on AgentCore instead. Your session survives a closed lid, credentials stay off local machines, and every agent gets its own isolated microVM.

Why Move Off Your Laptop​

The problem: coding agents need uninterrupted sessions, a persistent filesystem, and access to credentials. A laptop provides all three, but poorly. Close the lid and the agent dies. Credentials sit in ~/.aws/, ~/.ssh/, and .env. Port collisions and shared environments introduce flakiness.

On AgentCore Runtime:

  • Isolated microVM per session — no shared state, no port collisions
  • Persistent workspace — session storage at /mnt/workspace survives disconnects and resume
  • Managed credentials — Token Vault or IAM roles, nothing on local disk
  • Connect from anywhere — reconnect via agentcore exec from any terminal

Which Agents Can You Host?​

AgentCore Runtime runs any process that fits in a container, it is not limited to a fixed list of agents. Any coding agent that can run in a Linux environment (CLI-based, headless) can be deployed. The table below shows validated examples, not the full set:

AgentAuth ModelModel
Claude CodeIAM role (Bedrock native)global.anthropic.claude-opus-4-6-v1
OpenCodeIAM role (Bedrock native)global.anthropic.claude-sonnet-4-6
HermesIAM role (Bedrock native)Bedrock (Llama, Claude)
KiroToken Vault + IdentityBedrock (auto model router)
CodexToken VaultOpenAI models via Bedrock
CursorToken VaultBedrock

The runtime provides a containerized microVM with persistent storage, networking, and managed credentials. Agents that use Bedrock natively (Claude Code, OpenCode, Hermes) need only an IAM execution role with bedrock:InvokeModel. For agents that call external APIs (Codex, Cursor, or your own custom tools), store keys in Token Vault and access them via environment variables at runtime. For Claude Code specifically, set CLAUDE_CODE_USE_BEDROCK=1 in the container environment.

How It Works​

Your Terminal→agentcore exec (WebSocket)→MicroVM (agent + /mnt/workspace)↔Gateway (GitHub MCP)

Each agent runs in its own microVM with a persistent workspace. You connect via interactive shell (WebSocket) or fire one-shot commands (HTTP/2). The session stays alive even when you disconnect.

Connect to Your Agent​

# Drop into the agent's VM (interactive shell)
agentcore exec --it --runtime my-coding-agent --session-id sess-jane-1234

# Reconnect later — your session survived closing the laptop
agentcore exec --it --session-id sess-jane-1234 --shell-id shell-789

# One-shot command — no LLM, no token cost
agentcore exec --runtime my-coding-agent --session-id sess-jane-1234 \
  "cd /mnt/workspace && npm test"

Interactive shells support colors, tab completion, and Ctrl+C. On brief network drops the session reconnects automatically; longer drops print a resume command with the --shell-id.

Wire Up Tools via MCP​

Give your agent access to GitHub, Jira, or internal APIs through a shared MCP Gateway:

# Claude Code
claude mcp add agentcore \
  https://<gateway-id>.gateway.bedrock-agentcore.us-west-2.amazonaws.com/mcp \
  --transport http

Sub-agents inherit the MCP configuration. The Gateway aggregates all targets into a single endpoint with semantic tool search.

Example: Deploy Claude Code on AgentCore​

The core deployment step is creating a runtime with your container image and a shared filesystem:

import boto3

client = boto3.client("bedrock-agentcore", region_name="us-west-2")

client.create_agent_runtime(
    agentRuntimeName="claude-code-agent",
    agentRuntimeArtifact={
        "containerConfiguration": {
            "containerUri": "123456789012.dkr.ecr.us-west-2.amazonaws.com/claude-code:latest"
        }
    },
    networkConfiguration={"networkMode": "PUBLIC"},
    protocolConfiguration={"serverProtocol": "HTTP"},
    filesystemConfigurations=[
        {"sessionStorage": {"mountPath": "/mnt/workspace"}},
        {"s3FilesAccessPoint": {
            "accessPointArn": "arn:aws:s3files:us-west-2:123456789012:access-point/skills-ap",
            "mountPath": "/mnt/s3files"
        }},
    ],
    roleArn="arn:aws:iam::123456789012:role/AgentCoreExecutionRole",
)

The container image includes the Claude Code CLI, Node.js, and an entrypoint that sets CLAUDE_CODE_USE_BEDROCK=1. S3 Files at /mnt/s3files provides shared skills and MCP proxy scripts across all sessions; session storage at /mnt/workspace gives each session its own persistent project directory.

Runtime microVM sessions are capped at 2 vCPU/8 GB and a 2 GB container image, both fixed limits. For coding agents with heavier toolchains, larger dependency sets, or repository clones that exceed these ceilings, use Runtime Instances for AWS-managed EC2 compute with configurable instance sizing.

Full Deployment Walkthrough​

Each agent follows the same pattern: setup.sh (build arm64 image, push to ECR) then deploy.py (create IAM role + AgentCore Runtime).

For the full infrastructure walkthroughs: