AgentCore

AgentCore Skill

General skill for Amazon Bedrock AgentCore: deploy and operate harnesses and runtime agents, and migrate AI assistant configurations from 15 coding assistants (GitHub Copilot, Claude Code, Cursor, Cline, Codex, Windsurf, Antigravity, Gemini CLI, Deep Agents, Dexto, Firebender, Kimi Code CLI, OpenCode, Warp, agentic-core).

Harness-first: agents that are persona + skills + standard tools deploy as declarative harnesses (no code). Only agents with custom logic get generated Strands code. Verified against agentcore CLI 0.22.0 (July 2026).

The plugin also bundles companion AWS skills — aws-cloudformation, aws-expert, aws-serverless, aws-solution-architect — so a coding agent has broader AWS context when building or migrating agents.

Quick Start

Prerequisites

python3 scripts/preflight_check.py

Needs: Python ≥3.10, Node ≥20, uv, AWS CLI + credentials, agentcore CLI ≥0.22.

Migration

# 1. Scan a repository for AI assistant configs
python3 scripts/scan_configs.py --repo-root <path> --format json

# 2. Generate the AgentCore project (harnesses + code runtimes)
python3 scripts/generate_project.py \
  --inventory migration-inventory.json \
  --output-dir ./agentcore-project \
  --region us-east-1

# 3. Deploy (or run the generated agentcore-commands.sh)
cd agentcore-project
./agentcore-commands.sh

Structure

This repo is a Claude Code plugin that bundles the agentcore skill plus a set of companion AWS skills (all auto-discovered from skills/).

├── .claude-plugin/
│   ├── plugin.json              # Plugin manifest
│   └── marketplace.json         # Marketplace manifest (installable via /plugin)
├── .mcp.json                    # Bundled MCP server (Bedrock AgentCore) — auto-connects when the plugin is enabled
├── skills/
│   ├── aws-cloudformation/      # CloudFormation authoring
│   ├── aws-expert/              # General AWS guidance
│   ├── aws-serverless/          # Serverless (Lambda/API GW) patterns
│   ├── aws-solution-architect/  # Architecture guidance
│   └── agentcore/               # The AgentCore migration skill
│       ├── SKILL.md             # Entry point
│       ├── scripts/             # preflight_check, scan_configs, generate_project, invoke_harness
│       ├── references/          # harness.md, migration.md, security-iam.md, deployment-checklist.md, ...
│       ├── evals/evals.json     # Regression scenarios (self-test)
│       └── assets/              # poc-cve-verify/ (+ VERIFIED.md), iam-policies/
└── LICENSE                      # MIT

Installing

As a plugin (Claude Code marketplace):

/plugin marketplace add LeonMelamud/agentcore-toolkit
/plugin install agentcore@agentcore-skills

Hosted (public) at github.com/LeonMelamud/agentcore-toolkit. Also installable via the skills.sh CLI: npx skills add LeonMelamud/agentcore-toolkit.

As a personal or project skill (copy the bundle):

cp -r skills/agentcore ~/.claude/skills/agentcore        # personal
cp -r skills/agentcore .claude/skills/agentcore          # project (or .github/skills/ for Copilot)

Connecting to AWS (required)

Nothing here works without AWS credentials. Every skill drives the AWS CLI / boto3 / the agentcore CLI, and the bundled MCP server (below) makes real Bedrock AgentCore API calls — all of them fail with no configured AWS identity. Installing the plugin does not connect you to AWS; do this once after install:

1. Configure credentials (any one):

aws configure sso                 # recommended for orgs / SSO
aws configure                     # static access keys
export AWS_ACCESS_KEY_ID=...  AWS_SECRET_ACCESS_KEY=...  AWS_REGION=us-east-1
# …or run on an EC2/ECS/Lambda role — the SDK picks it up automatically

2. Verify the whole toolchain is reachable:

python3 skills/agentcore/scripts/preflight_check.py

Checks Python ≥3.10, Node ≥20, uv, AWS CLI + live credentials (prints your account id), the agentcore CLI ≥0.22, and Bedrock AgentCore access. Fix anything it flags before running a workflow.

3. Enable Bedrock model access in the target account/region (Bedrock console → Model access). Without it, invokes fail with ModelNotAccessibleException; a 0 daily-token quota surfaces as ThrottlingException (see the skill's error table).

Bundled MCP server

The plugin ships .mcp.json declaring the official Amazon Bedrock AgentCore MCP server (awslabs.amazon-bedrock-agentcore-mcp-server). When the plugin is enabled, Claude Code launches it via uvx and it shows up as bedrock-agentcore-mcp MCP · connected, giving the agent direct AgentCore runtime / memory / gateway / identity API tools alongside the skills.