aws-solutions/konductor

---

Why Konductor?

Most AI coding assistants are one agent in one conversation. You end up as the project manager: writing the requirements, asking for a design, pasting it back for implementation, then remembering to ask for tests.

Konductor installs a team that already knows how to work together. Ask the konductor orchestrator for something in plain language. It breaks the request down, sends each piece to the right specialist, checks what comes back, and sends it back for fixes when it falls short.

• 🎼 One request, the whole lifecycle. Requirements, design, code, tests and docs, without hand-carrying artifacts between chats.
• ✅ Every deliverable gets checked. Work passes a maker-checker gate, and failed checks go back for up to two fix cycles.
• 🧠 Expertise built in. 82 skills and 19 step-by-step workflows (SOPs) cover threat models, IAM policies, DynamoDB design, code review, E2E testing and more.
• 🪶 No infrastructure. Agents are configuration (prompts, skills, SOPs) that your Kiro or Claude Code runtime runs directly. No servers, no extra accounts.

flowchart LR
    You(["💬 You"]) --> K["🎼 konductororchestrator"]
    K --> PM["📋 k-product-managerrequirements"]
    K --> AR["🏛️ k-architectdesign"]
    K --> DEV["🛠️ k-developercode"]
    K --> QA["🧪 k-quality-assurancetests"]
    K --> MORE["🔎 researcher · 📊 tpm🌐 browser · 🖼️ media"]
    PM & AR & DEV & QA & MORE -.->|"checked before handoff"| K
    classDef orch fill:#7C5CFF,color:#ffffff,stroke:#7C5CFF
    classDef agent fill:#0B0A12,color:#ffffff,stroke:#3DDC97
    class K orch
    class PM,AR,DEV,QA,MORE agent

[!TIP] Learn it hands-on. A-SDLC: Conducting Agents with Konductor on Kiro is a self-paced workshop of about 5 hours. You drive the agents through ambiguous, problem-based scenarios and learn to judge when the work is actually done. It runs on your own machine with Kiro CLI. You don’t need an AWS account, but the agents use your own Kiro credits.

---

Quick start

1. Install (Linux x86_64/aarch64 or Apple Silicon macOS):

curl -fsSL https://raw.githubusercontent.com/aws-solutions/konductor/refs/heads/main/scripts/konductor-bootstrap.sh | bash

That installs for Kiro CLI. For Claude Code, run the same command with | KONDUCTOR_HARNESS=claude bash at the end. Piping a script to bash runs code you haven’t read; to check it first, see what the install script does below. Using the Claude Code plugin marketplace instead? Run /plugin marketplace add aws-solutions/konductor, then /plugin install konductor@konductor (details). Using Kiro IDE? Install via a Kiro Power.

2. Start the orchestrator:

kiro-cli chat --agent konductor     # Kiro CLI
claude --agent konductor            # Claude Code

3. Ask for something real:

Design and implement a service that ingests IoT sensor events and alerts on anomalies.

The orchestrator sends the design to k-architect, the implementation to k-developer and the test plan to k-quality-assurance, and verifies each handoff before moving on.

---

Meet the team

konductor is where you start. It never writes code itself: it delegates, verifies, and re-delegates when a quality gate fails (up to two fix cycles).

Every agent runs claude-sonnet-5.5. Full details, including each agent’s tools, are in the agents reference.

---

See it in action

Review your branch before opening a PR (Kiro CLI):

kiro-cli chat --agent konductor
> Run the k-code-review-workflow SOP against my current branch.

Get a threat model in one shot (Claude Code):

claude --agent konductor -p "Create a threat model for a public REST API backed by DynamoDB."

Run the whole lifecycle with the k-full-sdlc SOP, from codebase analysis through documentation. Artifacts land in .konductor/.

# Kiro CLI
kiro-cli chat --agent konductor
> /agent-sop:k-full-sdlc project_description: "JSON parser" max_fix_cycles: 5

# Claude Code
claude --agent konductor -p "/sop-k-full-sdlc project_description: 'a CLI tool that tails a log file and alerts on error spikes'"

More walkthroughs: understand a codebase, design a system, plan and verify work and review code and tests.

---

What’s inside

🧠 82 skills

Skills are packaged expertise the agents load when they need it. For Claude Code, every skill in an agent’s skills: frontmatter is injected in full at session start; other skills load on demand (how loading works).

Browse them all in the skills catalog.

📜 19 SOPs

SOPs are repeatable, multi-step workflows. Run them as /agent-sop: in Kiro CLI or /sop- in Claude Code.

Plus about-konductor for onboarding and k-delegate, which the orchestrator uses internally. Each one is explained, with a flowchart, in SOP workflows.

💾 Persistent memory

Agents with the persistent-memory skill keep a small scratchpad across sessions, with no setup. The directory is created on first write and added to .gitignore.

Tune the size limits and an allowlist of external domains with an optional config. A validator rejects any write that breaks the rules and tells you why; nothing is dropped silently.

mkdir -p .konductor && cp skills/persistent-memory/memory-config.json.template .konductor/memory-config.json

🔌 Integrations

AWS MCP comes pre-wired for k-architect and k-developer on Claude Code, launched through mcp-proxy-for-aws-cli==1.7.0. You need uvx and AWS credentials. Both agents get the full server, including AWS API tools, so scope agent-initiated calls with the aws:ViaAWSMCPService and aws:CalledViaAWSMCP IAM condition keys. Kiro CLI users add aws-mcp to their own MCP config, where knowledge tools are auto-approved and API tools ask first. See Install for Claude Code or Install for Kiro CLI. The AWS Agent Toolkit also offers an OAuth connection, an AWS-managed plugin and standalone skills.

Everything else is bring-your-own:

• Browser automation: k-browser uses a @playwright/mcp server. Install Chromium after configuring it.
• Slack search: k-researcher with a Slack MCP server (guide).
• Asana: k-product-manager with an Asana MCP server (guide).

---

Project layout

konductor/
├── agents/        11 agent specs (*.agent-spec.json)
├── agent-sops/    19 SOPs (*.sop.md)
├── skills/        82 skills (skills//SKILL.md)
├── context/       Context loaded at agent startup (core rules for every agent, orchestrator routing rules)
├── cli/           The konductor CLI (Rust)
├── mcp/           The skill-lookup MCP server
├── shared/        Shared Rust crates (telemetry)
├── docs/          User guide, integration guides, GitHub Pages site
├── scripts/       Install, release and plugin tooling
└── tests/         Benchmark scenarios and script tests

Contributing

Bug reports, ideas and pull requests are welcome. Start with CONTRIBUTING.md, and see the Code of Conduct. To change agent behavior or build the CLI from source, read contributing and customizing.

Security

See SECURITY.md to report a vulnerability. Please don’t open a public issue.

License

Apache License 2.0. See LICENSE.txt; third-party attributions are in NOTICE.txt.

Data Collection

Konductor sends usage metrics to AWS. They are not anonymous: every event carries an identifier that stays the same across all projects on your machine.

What’s sent

• Seven event types: agent_invocation, subagent_invocation, mcp_tool_call, cli_error, package_installed, package_version_updated and package_uninstalled.
• Agent events cover only Konductor’s own agents (the orchestrators and k-* specialists), never your agents or a runtime’s built-in ones. They carry the agent’s name and, for a delegation, the delegating Konductor agent’s name when the runtime exposes it.
• Other events carry the failed CLI subcommand (for errors), the runtime (kiro-cli-v2, kiro-v3 or claude), the CLI version and, for install and update events, the installed content version.
• Session IDs, when the runtime exposes one, are hashed with a per-install identifier first, so the raw value is never sent and the same session hashes differently on another install.

What’s never sent: project paths, directory names or file contents. The exact schema is in docs/telemetry-schema.json.

The machine identifier. The first time an event is reported on a machine, Konductor creates a UUID and stores it in $HOME/.konductor/telemetry.json. Every project on that machine sends the same UUID, so events from different projects can be linked to one machine, though not to which projects.

Claude Code plugin. The marketplace plugin sends agent_invocation and subagent_invocation the same way, through a plugin hook, with claude-plugin as the runtime. It shares the machine identifier in $HOME/.konductor/telemetry.json with the CLI. See Install via the Claude Code plugin marketplace.

Opting out

• --no-telemetry also removes an opt-in record left by an earlier install, and stops a global install’s hooks from reporting in that project.
• telemetry.json exists only after the first event is reported; a --no-telemetry install never creates it. Konductor never resets a consent value you set.
• Reporting needs both the project and the machine setting to allow it. konductor doctor shows which one is in effect.
• Keep telemetry.json valid. It needs all four fields (schema_version, UUID, created_at, telemetry_consent) and a UUID of exactly 64 lowercase hex characters. Konductor never repairs, replaces or deletes the file, so if it’s malformed, events fall back to an untraceable placeholder identifier and your consent setting is ignored until you fix or delete it by hand.

This solution sends operational metrics to AWS (the “Data”) about the use of this solution. We use this Data to better understand how customers use this solution and related services and products. AWS’s collection of this Data is subject to the AWS Privacy Notice.

---

Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the “License”); you may not use this file except in compliance with the License. You may obtain a copy of the License at

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an “AS IS” BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.