verdigraph-neurogenesis

Clone this repo, run one script, and within 60 seconds you're building deterministic, content-addressed brain artifacts from any agent file — Claude project export, OpenAI Assistant config, raw prompt list, or Verdigraph genome JSON. Pure Python core; zero external services required.

60-second quickstart

git clone https://github.com/viridis-security/verdigraph-neurogenesis
cd verdigraph-neurogenesis
bash quickstart.sh

That's it. The script creates a venv, installs the package editable, runs the brain builder against an example genome, and prints the deterministic brain_id + content_hash. No Cloudflare account, no Stripe key, no `verdigraph.dev` account needed. Everything runs locally.

If you also have an internet connection, the script will additionally hit https://verdigraph.dev/app/import with the same input bytes and confirm the hosted Worker produces the exact same brain_id — that's your proof the local build is byte-equivalent to the production reference implementation.

What it is

Verdigraph turns an agent file into an inspectable cognitive graph with a content-addressed identifier you can pin in git, cite in an audit, or paste into a code review. Three things make this useful:

1. Determinism. Identical input bytes always produce identical brain_id, content_hash, and graph structure. Run it twice, get the same answer twice. Run it in Python locally; run it in TypeScript on the Worker; same answer either way.
2. Inspectable structure. Every brain carries 9 firing invariants + 1 advisory check (I9_fitness_metric_wired) so you can prove what the agent file actually compiles to without trusting a black box.
3. Self-contained build pipeline. No external dependencies beyond the Python stdlib. No SaaS lock-in. You can audit every line of verdigraph/brain.py (≈ 660 lines) in an afternoon.

Use it

Build a brain from a Verdigraph genome

python -m verdigraph build --file examples/hypothetical_research_agent.genome.json --format verdigraph_genome --pretty

Or pipe input:

cat my_agent.json | python -m verdigraph build --stdin --format auto --summary --pretty

Build from a Claude project export

python -m verdigraph build --file my_claude_project_export.json --format claude_project_export --pretty

Build from an OpenAI Assistant config

python -m verdigraph build --file my_assistant.json --format openai_assistant --pretty

Build from a flat prompt list

echo -e "You are a helpful assistant.\nSummarize the user's request.\nPlan steps and execute." \
  | python -m verdigraph build --stdin --format prompt_list --pretty

Re-verify a saved brain artifact

python -m verdigraph build --file my_agent.json --pretty > brain.json
python -m verdigraph verify brain.json

Use it as a Python library

from verdigraph.brain import extract, verify_brain, to_dict

genome = b'{"agent_name":"my_agent","purpose":"...","initial_nodes":["planner","executor"],"fitness_metrics":["task_success_rate"]}'

brain = extract("verdigraph_genome", genome)
print(brain.brain_id)        # e.g. RMX124YY916WP0TCSEHFYX7M30
print(brain.brain_uri)       # verdigraph://brain/RMX124YY916WP0TCSEHFYX7M30
print(brain.content_hash)    # sha256 hex
print(len(brain.nodes), "nodes,", len(brain.edges), "edges")

report = verify_brain(brain)
assert report.passed         # all non-advisory invariants pass

print(to_dict(brain))        # serialize for storage / round-trip

Expose it as an MCP server for your LLM agent

pip install -e ".[mcp]"
verdigraph-mcp                 # runs over stdio

Then in Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "verdigraph": {
      "command": "/absolute/path/to/repo/.venv/bin/verdigraph-mcp",
      "args": []
    }
  }
}

Or in Claude Code: claude mcp add --transport stdio verdigraph /absolute/path/to/repo/.venv/bin/verdigraph-mcp.

Restart your client. Your agent now has verdigraph_* tools to build/verify/evolve brains directly. No network calls; everything runs on your machine.

Determinism and verifiability

Canonicalization rule (one sentence)

Apply json.dumps with separators=(",", ":") after recursively sorting every object's keys lexicographically by codepoint and coercing integer-valued floats to integers (matches JavaScript JSON.stringify byte-for-byte). UTF-8 encoded before hashing. See verdigraph/brain.py::canonicalize (≈ 20 lines, stdlib only).

Layout

verdigraph-neurogenesis/
├── README.md                     ← you are here
├── quickstart.sh                 ← clone → first brain in 60 seconds
├── pyproject.toml                ← Python package metadata
├── verdigraph/                   ← Python core (no external deps)
│   ├── brain.py                  ← deterministic build pipeline (extract / canonicalize / verify / evolve)
│   ├── cli.py                    ← `python -m verdigraph` CLI
│   ├── genome.py                 ← AgentGenome / GrowthRules / SafetyAxioms (live-agent runtime)
│   ├── graph.py                  ← CognitiveGraph / CognitiveNode / CognitiveEdge
│   ├── agent.py                  ← DevelopmentalAgent (live-agent runtime)
│   ├── growth.py / pruning.py    ← evolution operators
│   ├── evaluation.py             ← task-outcome ledger
│   ├── compute.py                ← compute-routing helpers
│   └── ledger.py                 ← immutable event log
├── verdigraph_mcp/               ← optional: stdio MCP server (`pip install -e ".[mcp]"`)
├── tests/                        ← pytest, all green on a clean clone
├── examples/                     ← runnable demos with fixture genomes
├── docs/                         ← canonicalization spec, architecture, invariants
├── papers/                       ← three companion papers (Zenodo-archived)
└── hosted-mcp/                   ← OPTIONAL: Cloudflare Workers deployment if you want a hosted instance

Optional: deploy your own hosted instance

A reference Cloudflare Workers deployment lives in hosted-mcp/. It serves the same deterministic-build pipeline over HTTPS + OAuth 2.1 + PKCE, adds prepaid USD credits via Stripe, and Ed25519-signed compliance attestations. You do not need this to use the Python core. It exists because the same protocol can run hosted if you want a shared multi-caller environment. See hosted-mcp/README.md for deployment instructions.

A live reference deployment runs at https://verdigraph.dev — same byte-equivalent pipeline. The local Python implementation is the canonical source; the Worker is a reimplementation for hosting convenience.

Run the tests

Python core:

source .venv/bin/activate
pip install -e ".[dev]"
pytest -q

TypeScript hosted-MCP (Cloudflare Worker):

cd hosted-mcp
npm ci
npm run typecheck
npm test

Both suites run in CI (.github/workflows/tests.yml) on every push and pull request: the Python job across 3.10 / 3.11 / 3.12, and the hosted-mcp job on Node 22 — where the cross-core parity.test.ts executes against a real Python install rather than self-skipping. A secret-scan job fails the build if a live Stripe identifier is ever committed.

The tests/test_brain_parity.py suite locks the deterministic-build contract — specifically that b'{"agent_name":"x","purpose":"y","initial_nodes":["a"],"fitness_metrics":["task_success_rate"]}' produces brain_id == "RMX124YY916WP0TCSEHFYX7M30" and content_hash == "20b9e5be0e5a0d34e564df6d0a554b1232ff9cc3ff309ab8da77a97756602c0c". If either side ever drifts, that test fails on the next CI run and we ship the divergence as a deliberate schema bump.

Companion papers

In papers/:

1. PAPER_1_Physical_NeuroGenesis_SynapseForge.md — physical version: AI-agent-architected, 3D-printed, solution-grown neuromorphic substrates.
2. PAPER_2_Verdigraph_Digital_NeuroGenesis.md — software version: self-evolving digital cognitive graphs.
3. PAPER_3_Verdigraph_Compute_Efficiency.md — compute-efficiency layer.

To cite:

Hart, Justin. (2026). Verdigraph NeuroGenesis: A Software Framework for Self-Evolving AI-Agent Cognitive Substrates (Version 0.1.0). Zenodo. https://doi.org/10.5281/zenodo.20261687

License & contact

MIT. Maintained by Viridis LLC. Contact: [email protected].

This is an experimental research framework. It does not create autonomous unrestricted self-modifying AI. All growth and pruning actions are constrained by explicit genome rules, safety invariants, and an auditable ledger.