Dev & Engineering persistent-sessionsself-improvingtask-runnerscheduled-jobssubagentsmessaging-integrationsmcp-serverswebhook-automation

Kiro Crew

Persistent development workspace that self-improves and continues beyond a session.

FollowAgents review · FARS-2.1
Not recommended
59/ 100 5-point scale 3.0 / 5
1 2 3 4 5 6
1Trust16 / 29 · 2.8/5

Evidence shows local-by-default, interactive approvals, OS sandbox, sensitive-path checks, credential redaction, deny rules, audit events, and governance profiles, but lacks implementation details or test evidence. Deductions: dependency security only mentions SHA-256 verification without vulnerability scanning or supply-chain measures; rollback only mentions task checkpoints without system-level rollback; source attribution only mentions SLSA attestation without concrete verification methods.

2Reliability8 / 14 · 2.9/5

Documentation is consistent, but no actual runtime evidence. Deductions: dependency availability lacks dependency list or version pinning; failure messages only mention retries and checkpoints without concrete error handling examples.

3Adaptability12 / 18 · 3.3/5

Documentation covers multiple scenarios and deployment modes, but lacks concrete configuration examples or boundary descriptions. Deductions: capability boundaries not explicitly listed; trigger precision lacks specific trigger condition examples; environment fit lacks detailed configuration beyond Windows.

4Convention11 / 18 · 3.1/5

Documentation structure is clear, install notes are detailed, license is explicit, but no version history or changelog. Deductions: naming stability lacks API stability commitment; known limitations not explicitly listed; maintenance responsibility not clearly stated.

5Effectiveness9 / 13 · 3.5/5

Documentation claims persistence, self-learning, self-evolution, but lacks evidence of actual effectiveness. Deductions: output usability lacks concrete output examples; marginal value not compared with existing tools; cost-benefit lacks performance or resource consumption data.

6Verifiability3 / 8 · 1.9/5

Claims in documentation lack traceability, no test results or independent verification. Deductions: claim traceability lacks specific test or verification links; cross-source corroboration lacks third-party verification; fact-inference separation not clearly distinguished.

Evidence confidence: Low Reviewed Aug 11, 2026 Reviewed revision c122ceea5d4e
The upstream repository has new commits since this review. The score still applies to the reviewed revision shown and may not cover the latest changes.
Before you use it
  • Publisher identity is unverified; be cautious about supply-chain risks.
  • Dependency security only mentions SHA-256 verification without vulnerability scanning or dependency pinning.
  • Rollback mechanism only mentions task checkpoints without system-level rollback.
  • Security claims in documentation lack implementation details and test evidence.
Review evidence [1][2][3][4][5][6][7][8]
See the full review method →

What does this agent do, and when should you use it?

Kiro Crew is an open-source development workspace that runs locally or on remote hardware. It provides a persistent, self-learning, and self-evolving agent experience through desktop apps, a web dashboard, CLI, and messaging channels like Slack and Discord. The core is a Gateway, a long-running process that manages sessions, memory, schedules, approvals, and messaging connections. Agent sessions drive kiro-cli over the Agent Client Protocol, supporting parallel sessions, long-running tasks, unattended autonomy, delegation to subagents, and more. Kiro Crew also supports installable apps, MCP servers, and markdown skills for extensibility. Security is enforced via interactive approvals, OS sandboxing, sensitive data guards, and governance policies.

Kiro Crew is a development agent platform that offers persistent sessions, self-learning (corrections and failures become lessons), self-evolving skills, long-running tasks (checkpointed tasks), unattended autonomy (scheduled jobs, webhooks, messaging events), subagent delegation, installable apps, and extensible tools (MCP, skills, hooks). It supports desktop apps, a web dashboard (port 5476), CLI commands (kirocrew chat, run, cron, spawn, security), and messaging channels (Slack, Discord, Telegram, Teams, Webex, WeCom, WeChat). The Gateway manages all agent sessions, scheduling, memory, approvals, and security policy. Agents run via kiro-cli over ACP and can handle multiple concurrent sessions.

  1. Developers who need an agent to continue working after closing the chat, e.g., running long builds or migration tasks.
  2. Teams that want the agent to respond to messages via Slack or Discord without keeping the dashboard open.
  3. Teams needing recurring reports or maintenance tasks, automated via scheduled jobs.
  4. Developers who want the agent to learn from failures and accumulate reusable skills over time.
  5. Scenarios requiring parallel work on multiple sub-tasks (e.g., research three options) and synthesize results.
  6. Independent developers who want to run the agent on their own hardware for data control while accessing via web or CLI.

What are this agent's strengths and limitations?

Pros
  • Persistent execution: continues working even after chat is closed.
  • Self-learning: learns from mistakes and corrections, forming durable lessons.
  • Self-evolving: automatically synthesizes reusable skills from repeated patterns.
  • Multiple interfaces: desktop, web, CLI, Slack, Discord, and more.
  • Extensible: supports MCP servers, markdown skills, and installable apps.
Limitations
  • Requires a Kiro account and relies on kiro-cli for model access, potentially limiting model choices.
  • Linux installation requires a system package manager (apt/dnf/yum) to install Python 3.10+ if not already present.
  • Windows lacks a desktop app and OS sandbox; requires explicit opt-in for unsandboxed execution.
  • Current deployment model keeps all components on a single host, not horizontally scalable.
  • Initial configuration involves managing credentials (e.g., Slack tokens) and setup steps, which can be complex.

How do you install or deploy this agent?

Install Kiro Crew via multiple methods:
- Desktop app: Download the DMG (macOS) or AppImage (Linux) from download.crew.kiro.dev.
- One-line install: Run curl -fsSL https://download.crew.kiro.dev/cli.sh | sh (Stable), or add --channel insider or --channel nightly to switch channels.
- Docker: docker run -d --name kirocrew -p 127.0.0.1:5476:5476 -v kirocrew-home:/home/kirocrew ghcr.io/kirodotdev/kirocrew:stable.
- From source: Requires Python 3.10+, Node.js 22+, npm, and kiro-cli. Clone the repo, run make build, then source .venv/bin/activate, followed by kirocrew setup, kirocrew doctor, and kirocrew gateway.

How do you use this agent?

First, ensure kiro-cli is signed in via the kiro CLI tool (device-code sign-in). Start the Gateway (via desktop app, kirocrew gateway, or Docker). Open http://localhost:5476 in a browser and start a conversation. For CLI use, run kirocrew chat for interactive chat, kirocrew run TASK.md for task runs, kirocrew cron for scheduled jobs, and kirocrew spawn run "task" to spawn subagents. Configure settings via the dashboard or kirocrew config commands. Approvals appear in the dashboard or connected messaging channels.

FAQ

What costs are involved in running Kiro Crew?
Kiro Crew itself is free, but it requires a Kiro account for model access, and model usage may incur costs depending on the provider. Additionally, you need your own hardware to run the instance (local or cloud).
How secure is the agent's tool access?
It provides interactive approvals, OS sandboxing on Linux/macOS, sensitive path guards, credential redaction, 137 deny rules, audit logs, and governance policies (policy files can enforce an enterprise ceiling).
What does the agent remember?
It retains preferences, project context, decaying history summaries, lessons, and skills. Stored state is inspectable and editable. Incognito and temporary session modes are available to opt out of persistence.
Can I run Kiro Crew on Windows?
Yes, via a source install, but there is no desktop app. Agent subprocesses are refused by default on Windows unless you set sandbox_allow_unsandboxed_exec; OS sandboxing is not available.
What telemetry is collected?
By default, it sends an anonymous daily heartbeat with a random instance ID, app version (stripped of build timestamps), Python minor version, install channel, and first-run flag. It never sends prompts, responses, file contents, credentials, or IP addresses.

Compare agents like this one

The same FARS review applied across the shortlist this agent qualifies for.

Related agents