defract vs kiro

kiro runs their recipes.
defract runs your process.

kiro CLI 3.0 ships cloud sessions (run in a managed sandbox, disconnect and resume), persistent memory via /memories, and retires the standalone desktop app. the workflow recipes (feature-pipeline, investigate, publish-pr) are still vendor-authored templates on Bedrock billing - not yours to define. defract's lifecycle is different in kind: user-defined stages, your Anthropic key, a visual design stage before any agent writes code, and specialist review agents before anything ships.

macos + windows

both are structured tools. kiro and defract are in a different category than raw parallel runners like Conductor or the native Agent Teams - both believe structure should precede code. kiro CLI 3.0 (Nov 2026) adds cloud sessions, persistent memory, and spec mode in the CLI - real advancements. the workflow recipes (vendor-authored, Bedrock billing) are still the same three templates. the distinction is architectural: kiro gives you a fixed pipeline on managed cloud infrastructure. defract's lifecycle is yours to define, running on your own Anthropic subscription, local-first. one gives you recipes; the other gives you a process.

where they diverge

where the two tools
take different roads.

01

your process, not three vendor recipes

kiro 1.2 Workflows ships three named pipeline recipes: feature-pipeline, investigate, and publish-pr. each is a multi-step sequence with nested steps, approval gates on pause-on-input, and background cloud execution. that is a real advancement - structured pipelines are better than none.

but the recipes are vendor-authored and fixed. the feature-pipeline recipe is kiro's version of a story-to-PR workflow - you can run it, not define it. defract's lifecycle is different: you author the stages, the stage prompts, and the approval conditions. the process belongs to your team, not to kiro's template library. and defract's stages run on your Anthropic key, not on Bedrock credits, so the economics are yours too.

02

a visual design stage, not just a design.md

kiro’s spec pipeline includes a design artifact - a design.md describing the intended UI and behavior in prose. that is a real checkpoint. defract’s design stage produces something different.

before any implementation agent starts, defract produces a visual design artifact - a rendered UI reference the agent actually works from, not ambiguous prose. the gap between “a dashboard with an activity feed” and something an agent can implement without inventing layout decisions is where output quality breaks down. a visual reference closes that gap in a way a markdown doc does not. this is the stage that surprised us most in how much it reduced drift.

03

a specialist review panel, not a single verdict

kiro’s feature-pipeline gates code on an approved review verdict, and that is a real gate. defract’s review differs in kind: a panel of specialist agents judging the work against the original story, run locally on your own key.

specialist review agents

a security reviewer (OWASP-anchored), a refactor reviewer (dead code, naming, over-engineering), and a generic reviewer each evaluate the work - not just the diff, but in the context of the original story.

gated before it ships

code only progresses once all three agents sign off and you approve. kiro’s feature-pipeline runs a reviewer loop too (capped at three attempts); defract’s difference is the specialist panel, and that it runs on your machine, your key.

04

local-first, bring your own key

kiro routes your code context through aws bedrock for inference - a credit-based model, with content stored on aws infrastructure. defract runs the opposite model.

defract is local-first: PTY transcripts and your codebase context never leave your machine. you pay anthropic directly for whatever inference you use - no intermediary credits, no interaction caps, no cloud storage of your proprietary code. for teams on anything sensitive, that is a meaningful operational difference, not a footnote. it also means your costs scale with what you actually build, not a per-seat credit tier.

side by side

the honest comparison.

defract kiro
unit of work a story shipped to production a spec-driven implementation session
workflow type user-authored: define your own stages, prompts, and approval conditions vendor-preset recipes: feature-pipeline, investigate, publish-pr (not user-defined)
lifecycle stages story → design (visual) → architecture → implementation → review → release requirements.md → design.md → tasks.md → implementation (Workflows adds named multi-step pipelines)
design stage visual design artifact - rendered UI reference before any agent starts design.md - a text description of planned behavior
code review specialist agents (security · refactor · generic); gated merge on sign-off feature-pipeline gates code on an approved review verdict (reviewer loop, 3-attempt cap); single pass, not a specialist panel
memory / context persistent codebase memory - compounds with every shipped story steering files (spec mode); CLI 3.0 adds persistent session memory via /memories
data handling local-first - PTY transcripts never leave your machine cloud-routed - code context through aws bedrock; content stored on aws infrastructure
economics bring your own anthropic key - pay anthropic directly, no interaction caps credit-based - free (limited) · pro $20/user/mo (1,000 credits) · pro+ $40/user/mo (2,000 credits + premium models)
parallel worktrees ›yes no parallel worktree support
environment dedicated desktop app - not an IDE CLI + VS Code extension; standalone desktop app retired in CLI 3.0
platform macos · windows macos · windows · linux · web
availability available now generally available

where kiro is ahead today

we’ll be direct: kiro is cross-platform - macos, windows, linux - while defract runs macos and windows. if your team is on linux, that still matters. kiro is generally available with enterprise-grade features already shipping: IAM and SSO authentication, usage dashboards, team credit pooling, and cost controls. CLI 3.0 (Nov 2026) added real capabilities: cloud sessions (run work in a managed sandbox, disconnect and resume without losing state), persistent memory via /memories (genuine cross-session learning), and spec mode in the CLI. the Workflows recipes (feature-pipeline, investigate, publish-pr) give you a structured pipeline out of the box if you don’t want to define one yourself. the spec artifacts (requirements.md, design.md, tasks.md) land in your repo as readable markdown. note: CLI 3.0 retired the standalone Classic desktop app and inline autocomplete - kiro is now primarily the CLI and VS Code extension.

the tradeoff kiro makes is ownership: the workflow is theirs, not yours, and inference routes through Bedrock on a credit model. defract’s lifecycle is yours to define - stages, prompts, approval conditions - running on your own Anthropic key, local-first. the visual design stage, the specialist review gate, the local-first data model, and the user-authored process are where defract runs a fundamentally different system. if kiro’s preset recipes cover your workflow, it is worth evaluating on its own terms.

your process, not a vendor template.

download defract and run a story through the full pipeline - visual design, implementation, review, ship - on your own Anthropic key.