The CLI is free to use. You bring your own AI API keys (OpenAI, Anthropic, Google). Codev Cloud (Tower IDE in the browser) is also free — no credit card, no trial period.

Codev OS open source

Production quality code.

Built autonomously by AI.

Q: How is this different from Cursor, Copilot, or Devin?

Those tools help you write code faster. Codev puts AI inside a real engineering process — specs, multi-agent review, phase gates, isolated builds. We tested both approaches: Claude Opus 4.6 alone scored 5.7 vs 7.0 with Codev's multi-agent review. Three minds catch what one misses.

Q: Do I need to use all three AI models?

The default uses Gemini, Codex, and Claude for maximum coverage. In our data, Codex caught all 4 security vulnerabilities, Claude caught type safety and doc issues, Gemini caught architecture-level bugs. You can configure which models participate, but we recommend at least two.

Q: Can I use this with my existing codebase?

Yes. Run `codev adopt` in any existing project. It analyzes your codebase structure, sets up context documents, and you're ready to write specs. No migration required.

Q: Is my code sent to third-party AI services?

Yes, code is sent to the AI providers you configure (Google, OpenAI, Anthropic) for review. All communication is encrypted in transit. You can self-host models if your security requirements demand it.

Ship production-quality code with AI that follows real engineering process — specs, autonomous builders, multi-model review, and phase gates. Open source.

What would you like to build?

Talk to Architect...

Get Started

Launch Cloud

Discord

106

PRs merged in 14 days

Builders fully autonomous

Median build time

Pre-merge bugs caught

14 days. 26 spec projects. 801 commits. Output equivalent to a 3–4 person elite engineering team. Read the full analysis →

Security catches — all caught pre-merge

security.log — multi-agent review

CRITICAL

SSRF blocklist bypass Percent-encoded paths bypassed URL blocklist. Fixed with decodeURIComponent + URL normalization. Caught by Codex · Spec 0097

CRITICAL

Path traversal via startsWith /project matches /project-secret. Fixed with startsWith(base + path.sep). Caught by Codex · Spec 0099

HIGH

File permissions not enforced on existing files writeFileSync mode only applies on creation. Fixed with chmodSync. Caught by Codex · Spec 0097

HIGH

Timestamp-based ID collisions Date.now() file IDs collide under rapid operations. Fixed with crypto.randomUUID(). Caught by Codex · Spec 0099

20 pre-merge catches · 3 reviewers · $1.59/PR consultation cost · Full data →

3 Minds Are Better Than 1

Claude Code vs. Codev

Both used Claude Opus 4.6. The difference: Codev adds multi-agent review — Gemini, Codex, and Claude cross-examine every checkpoint. Three AI reviewers scored both blind. Read the full report →

Claude Code — todo app 5.7/10

Todo App

All

Active

Set up project structure

Add NL interface bug: crashes on parse

Write tests 0.26:1 ratio

useSyncExternalStore: infinite re-render loop detected

2 commits0 docs8 bugs

Codev — todo app 7.0/10

Todo Manager

All

Active

By Date

Set up project structure spec-verified

NL interface w/ Gemini 3-way reviewed

Tests + error handling 0.73:1 ratio

All gates passed. PR #1 ready for merge.

15 commits8 docs6 bugs (0 critical)

5.7/10

Claude Code (no methodology)

Bugs8 (1 Critical)

Test:code ratio0.26:1

Git commits2

Documentation0 files

7.0/10

Codev (spec → build → review)

Bugs6 (0 Critical)

Test:code ratio0.73:1

Git commits15

Documentationspec + plan + review + 5 consults

Bugs

4.7

7.3

Quality

6.3

7.7

Maintain

7.3

7.7

Tests

Extend

Reviewed by Claude Opus 4.6, GPT-5.3 Codex, Gemini 3 Pro

“Used agent farm today for the first time, and finished a 12 part piece of work in a few minutes. It managed to consume all my 5 hours worth of Claude Max quota in those few minutes, but that was a very good use of the quota.”

Younes

#codev on Discord

“I'm running multiple agents across multiple projects (solo dev, business owner). Codev's spec-first methodology has been an extension of what I was already doing manually and works great.”

rontxhou

#general on Discord

“I find that codev/spider is really good at getting information from the user to create the spec.”

jayen

#codev on Discord

“I tried the spec in Claude Code yesterday and it worked fantastically. Multi-perspective happened inside Claude Code and also did great.”

jayen

#codev on Discord

“Hey Waleed -- I've been kicking the tires on the platform. Its impressive so far! Nice work!”

bstan78

#general on Discord

Younes

#codev on Discord

“I'm running multiple agents across multiple projects (solo dev, business owner). Codev's spec-first methodology has been an extension of what I was already doing manually and works great.”

rontxhou

#general on Discord

“I find that codev/spider is really good at getting information from the user to create the spec.”

jayen

#codev on Discord

“I tried the spec in Claude Code yesterday and it worked fantastically. Multi-perspective happened inside Claude Code and also did great.”

jayen

#codev on Discord

“Hey Waleed -- I've been kicking the tires on the platform. Its impressive so far! Nice work!”

bstan78

#general on Discord

How It Works

The methodology your AI is missing.

01 — Specify

Every feature starts with a spec.

Structured specs define requirements, edge cases, and success criteria — so the AI knows exactly what "done" looks like. Not a chat message. A document with acceptance gates.

02 — Build

Autonomous builders in isolated worktrees.

An Architect agent manages scope and approves gates. Builder agents run in isolated git worktrees — parallel builds, clean branches, no conflicts. You approve what matters.

03 — Review

Three models. Zero blind spots.

Gemini, Codex, and Claude cross-examine every checkpoint. In our data, Codex alone caught all 4 security vulnerabilities. Each model's blind spot is another's strength.

Codex: 38 blocks Claude: 8 blocks Gemini: 5 blocks

04 — Gate

Protocols, not prompts.

Enforced phase gates ensure every feature follows the same path: specify → plan → implement → review. No skipping steps. No "just ship it." Every decision is auditable in git.

SPIR · TICK · EXPERIMENT · MAINTAIN · LAUNCH

Tower AIDE

See your builders work in real time.

Tower is a web-based AIDE that monitors architect and builder agents. Split-pane view: architect terminal on the left, builder status on the right. Also available free as Codev Cloud — no install needed.

Tower AIDE desktop view — Architect terminal managing 4 builders with live status, PRs, and project backlog

Tower running locally — 4 builders, 3 PRs under review, architect managing gates and consults

Works on mobile too

Codev Cloud runs Tower in the browser. Monitor builders from your phone.

Work View

Builders, PRs, backlog

Builder Terminal

Live consult output

Gate Management

Approve from anywhere

Codev Cloud FREE

Tower AIDE in the browser. Monitor builders remotely. No credit card.

Launch Cloud

Get Started

Open source. Five minutes to production.

terminal

$ npm install -g @cluesmith/codev

Co-develop a Spec

Define requirements, edge cases, and success criteria together with the AI.

codev/specs/0001-feature.md

Spawn a Builder

Builder runs in an isolated worktree. No conflicts.

af spawn -p 0001

Review the Plan

Builder writes a plan. You approve it. Multi-agent review validates it.

porch approve 0001 plan-approval

Walk Away

With a spec and plan, the builder works autonomously for hours — not minutes.

Builder runs 4+ hours unattended

Codev Cloud FREE

Tower IDE in the browser. Monitor builders remotely. No credit card.

Launch Cloud

FAQ

How is this different from Cursor, Copilot, or Devin?

Do I need to use all three AI models?

Can I use this with my existing codebase?

Is my code sent to third-party AI services?

Is it free?

Production quality code.
Built autonomously by AI.

One AI model scored 5.7 alone. Add multi-agent review and it scores 7.0. Three minds catch what one misses.

Built autonomously by AI.

Security catches — all caught pre-merge

3 Minds Are Better Than 1

Claude Code vs. Codev

How It Works

The methodology your AI is missing.

Every feature starts with a spec.

Autonomous builders in isolated worktrees.

Three models. Zero blind spots.

Protocols, not prompts.

Tower AIDE

See your builders work in real time.

Works on mobile too

Get Started

Open source. Five minutes to production.

Co-develop a Spec

Spawn a Builder

Review the Plan

Walk Away

FAQ

How is this different from Cursor, Copilot, or Devin?

Do I need to use all three AI models?

Can I use this with my existing codebase?

Is my code sent to third-party AI services?

Is it free?

Stay in the Loop

Get updates on what we're building.

Production quality code. Built autonomously by AI.

Production quality code.
Built autonomously by AI.