Lavra (/ˈla.vɾɐ/ — Portuguese for "harvest")

Lavra turns your AI coding agent into a development team that gets smarter with every task.

A plugin for coding agents that orchestrates the full development lifecycle -- from brainstorming to shipping -- while automatically capturing and recalling knowledge so each unit of work makes the next one easier.

Quick Start | Full Catalog | Architecture | Security | Command Map | v0.7.1 Release Notes

Without Lavra The agent forgets everything between sessions -- you re-explain context every time Planning is shallow: it jumps to code before thinking through the problem Review is inconsistent: sometimes thorough, sometimes a rubber stamp Knowledge stays in your head. When a teammate hits the same bug, they start from zero Shipping is manual: you run tests, create the PR, close tickets, push -- every time

With Lavra Automatic memory. Knowledge is captured inline during work and recalled automatically at the start of every session. Hit the same OAuth bug next month? The agent already knows the fix. Structured planning. Brainstorm with scope sharpening, research with domain-matched agents, adversarial plan review -- all before a single line of code is written. Disciplined execution. Agents follow deviation rules (what to auto-fix vs. escalate), commit per task with traceable bead IDs, and verify every success criterion before marking work done. Built-in quality gates. Every implementation runs through a review-fix-learn loop. Knowledge capture is mandatory, not optional. Team-shareable knowledge. Memory lives in .lavra/memory/knowledge.jsonl, tracked by git. Your team compounds knowledge together.

The workflow

Most of the time, you type three commands:

/lavra-design "I want users to upload photos for listings"

This runs the full planning pipeline as a single command: interactive brainstorm with scope sharpening, structured plan with phased beads, domain-matched research agents, plan revision, and adversarial review. The output is detailed enough that implementation is mechanical.

/lavra-work

Picks up the approved plan and implements it. Auto-routes between single and multi-bead parallel execution. Includes mandatory review, fix loop, and knowledge curation -- all automatic.

/lavra-ship

Rebases on main, runs tests, scans for secrets and debug leftovers, creates the PR, closes beads, and pushes the backup. One command to land the plane.

Add /lavra-qa between work and ship when building web apps -- it maps changed files to routes and runs browser-based verification with screenshots.

Who this is for

Anyone using coding agents who wants consistent, high-quality output instead of hoping the agent gets it right this time.

• Non-technical users: /lavra-design "build me X" handles the brainstorming, planning, and research. /lavra-work handles the implementation with built-in quality gates. You get working software without needing to know how to code.
• Solo developers: The memory system acts as a second brain. Past decisions, patterns, and gotchas surface automatically when they're relevant.
• Teams: Knowledge compounds across contributors. One person's hard-won insight becomes everyone's starting context.

Install

Requires: beads CLI, jq, sqlite3

npx @lavralabs/lavra@latest

Or manual:

git clone https://github.com/roberto-mello/lavra.git
cd lavra
./install.sh               # Claude Code (default)
./install.sh --opencode    # OpenCode
./install.sh --gemini      # Gemini CLI
./install.sh --cortex      # Cortex Code

Pipeline (4): /lavra-design, /lavra-work, /lavra-qa, /lavra-ship

Supporting (9): /lavra-quick (fast path with escalation), /lavra-learn (knowledge curation), /lavra-recall (mid-session search), /lavra-checkpoint (save progress), /lavra-retro (weekly analytics), /lavra-import, /lavra-triage, /changelog, /test-browser

Power-user (6): /lavra-plan, /lavra-research, /lavra-eng-review, /lavra-review (15 specialized review agents), /lavra-work-ralph (autonomous retry), /lavra-work-teams (persistent workers)

30 specialized agents across review, research, design, workflow, and docs. Each runs at the right model tier to keep costs 60-70% lower than running everything on Opus.

See docs/CATALOG.md for the full listing.

How knowledge compounds

brainstorm  --DECISION-->  design
design      <--LEARNED/PATTERN--  auto-recall from prior work
research    --FACT/INVESTIGATION-->  plan revision
work        --LEARNED (inline)-->  mandatory knowledge gate
review      --LEARNED-->  issues become future recall
retro       synthesizes patterns, surfaces gaps

Six knowledge types (LEARNED, DECISION, FACT, PATTERN, INVESTIGATION, DEVIATION) are captured inline during work and stored in .lavra/memory/knowledge.jsonl. At session start, relevant entries are recalled automatically based on your current beads and git branch. The system gets smarter over time -- not just for you, but for your whole team.

Configuration

.lavra/config/lavra.json can be created manually or by the /lavra-setup command.
It allows users to toggle workflow phases, planning and execution behavior:

{
  "workflow": {
    "research": true,             // run research agents in /lavra-design
    "plan_review": true,          // run plan review phase in /lavra-design
    "goal_verification": true,    // verify completion criteria in /lavra-work and /lavra-ship
    "testing_scope": "targeted"   // "targeted" (hooks, API routes, complex logic only) or "full" (all tests)
  },
  "execution": {
    "max_parallel_agents": 3,     // max subagents running at once
    "commit_granularity": "task"  // "task" (atomic, default) or "wave" (legacy)
  },
  "model_profile": "balanced"     // "balanced" (default) or "quality" (opus for review/verification agents)
}

/lavra-setup: run this to generate a codebase profile (stack, architecture, conventions) that informs planning.

Acknowledgments

Built by Roberto Mello, extending compound-engineering by Every. Task tracking by Beads.

Inspired by Every's writing on compound engineering, with ideas from Mario Zechner, Simon Willison, Get Shit Done, gstach by Garry Tan and many others. Thanks to my friend Dan for rubber-ducking Lavra.

Lavra (/ˈla.vɾɐ/) is the Portuguese word for "harvest" — the idea that every session plants knowledge that the next one reaps. Named by Roberto Mello, who is Brazilian.