codev
The operating system for
human-AI codevelopment
You write specs. Agents execute them. Protocols make it reliable. Codev is a methodology and toolchain that turns AI coding from unpredictable into systematic.
The problem isn't the AI — it's the process
AI can write code. You know this. But you also know the code is often wrong — wrong context, wrong assumptions, wrong patterns.
- Generated code that doesn't work in your codebase's context
- Lost context between sessions — chat history as your source of truth
- No way to catch AI mistakes before they ship
- One model's blind spots become your blind spots
- Tech debt accumulates silently with every AI-generated commit
Vibe coding doesn't scale. You need a system.
How Codev works
Four pillars that turn AI coding from chaotic into systematic.
Multi-Agent Orchestration
No single AI has all the answers. Codev consults three models — Gemini, Codex, and Claude — at every checkpoint. One flags a security issue, another catches a missing edge case, a third spots an API misuse. The cost of consultation is far less than the cost of bugs in production.
$ consult --type spec-review spec 0042
[gemini] APPROVE — comprehensive, well-structured
[codex] REQUEST_CHANGES — missing rate limit on /api/upload
[claude] APPROVE — minor: consider adding retry logic
Verdict: 2 approve, 1 request changesEnforced Protocols
SPIR, TICK, MAINTAIN, EXPERIMENT. These aren't suggestions — they're enforced by tooling. Gates require human approval before proceeding. Porch tracks state, runs checks, and blocks advancement until criteria are met. You can't skip tests, skip the review, or merge without consultation.
$ porch status 0042
PROJECT: 0042 - user-authentication
PROTOCOL: spir
PHASE: implement (Phase 2: API Endpoints)
GATE: plan-approval ✓ (approved 2h ago)
CHECKS: build ✓ tests ✓ lint ✓
NEXT: porch done 0042Agent Farm — An IDE for Agents
Agent Farm spawns builders in isolated git worktrees, each running in their own terminal session. An architect orchestrates; builders execute. You can watch them work, send them messages, review their output. Tower provides a unified dashboard across all your projects.
$ af status
BUILDERS:
#1 [0042-user-auth] implementing builder/0042
#2 [0043-search-api] reviewing builder/0043
#3 [0044-dashboard] testing builder/0044
ARCHITECT: watching 3 buildersDistinct Agent Roles
Not all agents work the same way. Codev defines three distinct roles: the Architect oversees the project, writes specs and plans, and orchestrates builders — but never writes code directly. Builders are autonomous implementation agents that execute specs in isolated worktrees. Consultants review work at every checkpoint but never build. Separation of concerns, applied to AI agents.
$ af status
ARCHITECT: watching project 0002-landing-page
└─ wrote spec, approved plan, monitoring builders
BUILDERS:
#1 [0002-landing-page] implementing builder/0002
└─ Phase 2: Landing Page Content & Layout
CONSULTANTS (last review):
[gemini] APPROVE — all six beats present
[codex] REQUEST_CHANGES — add viewport tests
[claude] APPROVE — clean component split
Architect → Builder #1: "Fix 4th pillar per correction"
Builder #1 → Consultants: ready for re-reviewThis isn't vaporware
We use Codev to build Codev
The spec for this landing page was written by the system. Every feature in codevos.ai goes through the same SPIR protocol — specify, plan, implement, review — with 3-way consultation at every checkpoint.
Open source, MIT license
No lock-in. Read the code, fork it, run it yourself. Every protocol, every template, every prompt is in the repo.
Real tools you can run today
$ codev init
✓ Created codev/specs/
✓ Created codev/plans/
✓ Created codev/reviews/
✓ Loaded protocols: spir, tick, maintain, experiment
✓ Created af-config.json
✓ Ready. Run "af start" to begin.A methodology you participate in
Codev is a methodology that evolves through use. The community shapes the protocols, contributes improvements, and shares what works.