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.
roberto-mello