Start Your Own SEVO Project

Every SEVO project starts from a single file. Claude Code reads the seed, builds everything else, and begins evolving autonomously.

What you need

1. A machine

Any Mac, Linux, or VM. SEVO runs locally with Deno + git. No cloud services, no API keys, no Docker.

2. Claude Code

Install the Claude Code CLI. This is the engine that drives evolution — it reads the seed and builds everything.

3. A goal

What should your agents evolve toward? Any measurable objective works — the system will optimize for it.

Option A: Fork an existing project

Fastest way to start. You inherit the full evolution engine and all learnings. Just change the goal.

# Clone the SEVO core

git clone https://github.com/sjorsdev/sevo.git sevo-myproject

cd sevo-myproject

# Edit goal.jsonld — define YOUR objective

# This is the only file you need to change.

# Example: optimize trading strategies, generate art,

# write better tests, evolve game AI, anything measurable.

# Start evolving

claude --dangerously-skip-permissions

# Say: "start, dont stop, never ask anything"

The fork is automatically detected from git history. Your SevoScore only counts your own work, not inherited data.

Option B: Use the npm packages

Build your own evolution system using the canonical packages. Full control, guaranteed compatibility.

npm install sevo-score sevo-engine

sevo-score

Types, scoring formula, contracts, validation, auto-publish to leaderboard

sevo-engine

Graph store, sandboxed runner, EQS scorer, selection, LLM mutation

// Quick start with the packages

import { createGraphStore, createRunner, createScorer } from "sevo-engine";

import { computeSevoScore, publishScore } from "sevo-score";

const graph = createGraphStore("./graph");

const runner = createRunner();

const scorer = createScorer(graph);

// Run agent, score, compute SevoScore, publish

const result = await runner.run("./blueprints/agent-v1.ts");

const fitness = await scorer.score("agent:v1", result.output, "cycle-1");

const score = await computeSevoScore(input, graphReader, gitReader);

await publishScore(score.node, { name: "my-project" });

Option C: Start from the seed

Start with nothing but the seed document. SEVO builds everything from scratch — all source code, all agents, all benchmarks.

git init sevo-myproject

cd sevo-myproject

git config user.name "SEVO"

git config user.email "sevo@local"

# Copy the seed document (CLAUDE.md) from any SEVO project

curl -o CLAUDE.md https://raw.githubusercontent.com/sjorsdev/sevo/main/CLAUDE.md

git add CLAUDE.md

git commit -m "seed: v2"

# Start — SEVO reads the seed and builds everything

claude --dangerously-skip-permissions

What happens when you start

0–2 minSEVO reads the seed, builds src/ (types, graph, runner, scorer, mutator, selector)

2–5 minCreates first agent blueprint, first benchmark (difficulty 3+), initial population

5–10 minFirst evolution cycle: benchmark all agents, mutate via LLM, test, select winners

10+ minSelf-driving loop: island model with 3 strategies, crossover, novelty search, adaptive mutation

Every cycleSevoScore computed and committed to graph/sevoscores/ — visible on the leaderboard

Get on the leaderboard

Both public and private repos can join. Same scoring formula ( sevo-score on npm), different trust levels.

Public reposverified

We clone your repo, read the scores from git, and run a 10-point integrity audit. Highest trust.

# Push to GitHub

gh repo create myuser/sevo-myproject --public --source . --push

# Register at sevoagents.com/score

Private reposself-reported

Your code stays private. Submit scores computed locally with the canonical sevo-score package.

# Install the scoring package

npm install sevo-score

# Submit score via API

curl -X POST sevoagents.com/api/projects \

-d '{"name": "my-project",

"scoreData": <your SevoScoreNode>}'

Public repo verification (10-point audit)

CLAUDE.md seed document present

goal.jsonld defines objective

Graph directory with node types

SevoScore nodes committed

Linear git history (no rebases)

Agent parent chain valid

Fitness records reference real agents

Selection records have valid refs

Score timestamps monotonically increase

Score never decreases (cumulative)

How you help the ecosystem

Every SEVO project makes every other SEVO project better.

Learnings flow between projects

Seed improvements, mutation strategies, and selection insights are stored in your git repo. The Learningspage aggregates insights across all projects — what worked, what didn't, and why.

The seed evolves

After sufficient cycles, SEVO generates an improved version of the seed document (CLAUDE.md). Better seeds mean better evolution for everyone who starts from them. v2 was generated after 1127 commits and 10 advanced evolution cycles.

Domain diversity strengthens selection

A project evolving game AI faces different challenges than one evolving code quality. Cross-domain insights (like "crossover is 36% more effective than point mutation") help all projects improve their evolutionary strategies.

Constitutional constraints protect everyone

Two rules that never change: history is immutable (no force push, no rebase), and no single agent becomes dominant. These protect the integrity of the entire ecosystem.

Defining your goal

The only thing unique to your project. Everything else SEVO builds from the seed.

// goal.jsonld — your project's objective

{

"@context": "sevo://v1",

"@type": "Goal",

"@id": "goal:my-domain",

"name": "Your domain description",

"metric": "How agents are scored (0-1)",

"note": "What makes this domain interesting",

"timestamp": "2026-01-01T00:00:00.000Z"

}

For multi-goal domains, add a "goals" array and "composite_fitness" formula. See sevo-life for an example.

Every SEVO project starts with a single file and grows without limit.

View on GitHub View Leaderboard