Preface¶

On August 4, 2026, AWS officially launched Kiro Crew—a persistent Agent workspace for developers, open-sourced under the Apache 2.0 license. Unlike Kiro’s original Autonomous Mode, Crew upgrades AI coding Agents from “session-only assistants” to “24/7 on-call engineering teammates”: multi-Agent and sub-Agent orchestration, cross-session persistent memory, scheduled tasks, and heartbeat monitoring can all be controlled remotely via desktop applications, web dashboards, or chat tools like Slack, Telegram, and Discord.

According to InfoWorld, Kiro Crew was initially incubated internally at Amazon under the name MeshClaw, and adopted by more than 39,000 Amazon developers within six months. Now open-sourced, its goal is to bring this orchestration layer to a broader range of engineering teams. This article outlines Kiro Crew’s positioning, core capabilities, and onboarding process based on AWS’s official GitHub repository and authoritative media reports.

What is Kiro Crew?¶

Kiro Crew is a persistent development workspace that runs on the developer’s own hardware (local Mac, container, or remote Linux host). It relies on kiro-cli under the hood, and drives Agent sessions via the Agent Client Protocol (ACP).

AWS Developer Advocate Darko Mesaros summarized it as: “A persistent, open-source development workspace for large tasks that cannot fit in a single session—turning AI coding Agents into always-on, self-learning, autonomously collaborating teammates.”

Compared to traditional Agent experiences where “you stare at the chat window and the session ends when the task finishes”, Kiro Crew emphasizes three key points:
1. Persistent: Sessions, memories, schedules, and task checkpoints can be restored after Gateway restarts.
2. Self-learning: Corrections and failure experiences are沉淀 into reusable lessons that influence future behavior.
3. Self-evolving: Repeatable patterns can be synthesized into skills, and both memories and skills are visible and editable in the dashboard.

No AWS account is required for deployment, and it does not depend on AWS-managed control planes. The Gateway, Agent runtime, and state all reside on user-controlled machines.

Differences from Kiro Autonomous Mode¶

Kiro is AWS’s Agentic development environment for software engineers, and Autonomous Mode can independently complete a single task within one session. Kiro Crew expands the boundaries of the orchestration layer on this basis:

If you have already configured steering files, skills, and custom Agents in the .kiro directory in Kiro, Crew can read them directly without migration or reconfiguration.

Core Architecture: Gateway and Multiple Entry Points¶

The core of Kiro Crew is the long-running Gateway process, which is responsible for routing messages, persisting sessions, injecting memories and skills, starting scheduled tasks, coordinating sub-Agents, approving tool calls, and displaying activity logs in the dashboard.

Desktop App / Web Dashboard / Slack / Telegram / Discord / CLI
                        ↓
              Gateway (Sessions · Memory · Scheduling · Approval · Apps)
                        ↓
         Agent Sessions (ACP Runtime · kiro-cli · MCP Tools · Models)

An Agent Session is a logically isolated unit: one conversation in the dashboard, a Slack thread, a scheduled task, or a sub-Agent task, each corresponding to an independent session that can run concurrently, with results aggregated back to the parent session.

The main capabilities listed in the official README include:
- Persistent Sessions: Concurrent, isolated multi-channel conversations; recoverable after Gateway restarts; historical sessions can be retrieved.
- Long-Term Tasks: Submit a task specification and leave; the Agent will plan steps, execute, validate, retry on failure, and resume from checkpoints. For example, kirocrew run TASK.md.
- Unattended Operation: Run scheduled Agent work or deterministic scripts without calling the model every time. For example, “summarize pending review work at 9 AM every weekday” can be set as a time-zone-aware recurring job.
- Delegation: kirocrew spawn run "task" starts isolated sub-Agents for parallel research, with results synthesized into the main conversation.
- Extensible Tools: MCP servers, Markdown skills, hooks; built-in kirocrew-core and kirocrew-cron MCPs.
- Defense in Depth: Tool approval, OS sandbox (Linux/macOS), sensitive path interception, credential redaction, 137 built-in deny rules, audit events.

Persistent Memory: Visible and Editable “Engineering Context”¶

Kiro Crew’s persistent memory is one of the most highly anticipated selling points of this launch. It maintains preferences, active project context, decaying historical summaries, and durable lessons; corrections and failures can change future behavior, and repeatable patterns can be elevated to skills. In-process embedding supports semantic retrieval, and stored state can be inspected and edited in the dashboard; temporary modes like Incognito are also provided to prevent sensitive conversations from being persisted.

Amazon emphasized in its launch announcement that memory is visible to developers, who decide which context Crew should carry forward—for example, preferring a certain coding style, or requiring a specific format for an MCP server under certain requests, all of which can be inherited across sessions.

InfoWorld quoted analysts as saying that many enterprises still use Agents like “shadow IT”, with individual developers connecting tools independently and no auditing; Crew’s shared workspace, approval gates, and logs provide a centralized entry point for governance.

Multi-Channel Control: Slack, Discord and Remote Gateway¶

Crew decouples “where the Agent runs” from “where the user operates”. In addition to desktop applications and the web dashboard (default http://localhost:5476), you can connect to Agent status under the same Gateway through the following entry points:

A typical workflow: start a multi-step refactoring task in the dashboard and leave; later add instructions or approve tool calls on Discord; the Agent will ping you via heartbeat or messages when manual review is needed. The messaging channel is an outbound connection, and there is generally no need to expose the dashboard port to the public internet.

Apps and Reference Applications¶

In addition to the core Gateway, Kiro Crew introduces the concept of Apps: custom UI, Agent, skills, scheduling, integration, and backend services combined for specific workflows. AWS simultaneously launched three open-source reference Apps:
1. DevFleets — worktree management
2. Issue Radar — Issue and PR triage
3. Task Runner — long-cycle engineering task execution

The community can also extend dashboard pages, subscribe to Gateway events, and register lifecycle hooks using the App Kit. SiliconANGLE noted that when Crew took off internally at Amazon, the community-driven App ecosystem was already core, and this model will be maintained after the public launch. Crew also supports accessing skills and configurations from open standard platforms like OpenClaw and Hermes, serving as a supplement rather than a complete replacement for existing toolchains.

Quick Start¶

All installation paths run kiro-cli under the hood; if the CLI is not installed on first launch, it will guide you through device-code login.

One-line installation (Stable channel):

curl -fsSL https://download.crew.kiro.dev/cli.sh | sh

After installation, visit http://localhost:5476 to start chatting in the web dashboard. To switch to the insider / nightly channel:

curl -fsSL https://download.crew.kiro.dev/cli.sh | sh -s -- --channel insider

Docker persistent deployment:

docker run -d --name kirocrew \
  -p 127.0.0.1:5476:5476 \
  -v kirocrew-home:/home/kirocrew \
  ghcr.io/kirodotdev/kirocrew:stable

Build from source:

git clone https://github.com/kirodotdev/KiroCrew.git
cd KiroCrew
make build
source .venv/bin/activate
kirocrew setup
kirocrew doctor
kirocrew gateway

Desktop version: macOS offers direct DMG downloads; Linux offers AppImage; there is no desktop build for Windows yet, but you can install the Gateway from source and access the dashboard via a browser. On Linux/macOS, you can use kirocrew service install to register a systemd or launchd service for auto-start on boot.

Configuration for connecting to IM tools like Slack can be found in the repository documentation: docs/guides/slack-setup.md and others.

Openness and Practical Constraints¶

Kiro Crew itself is open-source and free (Apache 2.0), with its GitHub repository at kirodotdev/KiroCrew and documentation portal at kiro.dev/docs/crew/. AWS plans to maintain the project under an open governance model with a public steering committee.

There are important limitations to be aware of:
1. Runtime dependency on Kiro CLI: Although the orchestration layer is based on open standards like ACP and MCP, the current launch version actually runs on AWS’s proprietary kiro-cli (billed by credits). Teams using other Agents like Claude Code, Codex, or Devin may need to build and validate connectors themselves.
2. Operational costs beyond the orchestration layer: Persistent Agents mean a corresponding rise in parallel model calls, CI activity, tool permissions, and manual review volume; enterprises need to formulate least-privilege, approval policies, memory retention, and audit strategies in advance.
3. Agent security is not a silver bullet: Sandboxes and deny rules reduce risk, but high-impact operations still require manual review to avoid pasting keys or sensitive data into conversations.

For teams already using Kiro, Crew is a low-friction incremental capability; for enterprises that have not standardized Agent governance, it is more like a “self-hosted orchestration test bed” rather than a plug-and-play fully automated engineering team.

Conclusion¶

Kiro Crew shifts the competitive focus of 2026’s Agent programming tools from “how much code can be written in a single conversation” to “whether engineering workflows can continue after stepping away from the keyboard”. Its differentiated combination includes multi-Agent orchestration, cross-session memory, scheduled tasks, and Slack/Discord integration; open-source and self-hosting address enterprise demands for keeping code, credentials, and audits within the internal network.

Whether it is suitable for your team depends on whether you already have the Kiro ecosystem, whether you are willing to bear the governance costs of persistent Agents, and the proportion of long-cycle tasks (dependency upgrades, PR triage, migration checkpoints) in your backlog. If you are moving from “pair-programming Copilot” to “asynchronous engineering teammates”, Kiro Crew is worth running kirocrew doctor locally or in a container to verify whether 24/7 orchestration meets expectations using real repository tasks.