onto — a living compiler of ontologies

A program is stored not as code, but as a genome: entities, rules with executable bodies, contracts, invariants. Code is the phenotype: printed into dialects (Go, Python), proven by the court (SMT), living under a warden, and regrown from scratch. A change in behavior without confirmation is a question with an executable counterexample, not a silent merge.

uv tool install ontogen                # -> global `onto`
onto validate genomes/hotel.yaml       # typecheck the genome
onto court    genomes/hotel.yaml       # SMT court: contracts proven, mutants distinguished
onto serve    genomes/hotel.yaml --port 8090 --data ./data [--store sqlite]
onto warden   genomes/hotel.yaml ...   # daemon: watch -> gates -> molt -> monitors
onto materialize genomes/hotel.yaml --dialect go-stdlib   # print the phenotype
onto mcp      genomes/hotel.yaml       # a mouth for the LLM (propose = the only write)

Models: bring your own (any OpenAI-compatible provider)

Configure providers and per-task model ladders in .onto/config.toml — OpenRouter, OpenAI, Groq, Together, DeepSeek, a local Ollama/vLLM, anything OpenAI-compatible. Name a known provider and give only a key (base_url is a preset); reference models as provider:model or bare. Keys come from ${ENV} or @keyfile, never hardcoded. onto models shows the registry.

[default]
provider = "openrouter"
[provider.openrouter]
api_key = "${OPENROUTER_API_KEY}"
[provider.local]
base_url = "http://localhost:11434/v1"   # Ollama
api_key = "ollama"
[ladders]
skills = ["qwen/qwen3-coder", "qwen/qwen3-coder-plus"]
nl     = ["anthropic/claude-sonnet-4.5", "local:qwen2.5-coder"]

Quickstart (zero machine)

curl -LsSf https://astral.sh/uv/install.sh | sh   # 1. uv (brings its own Python)
uv tool install ontogen                           # 2. the engine, isolated
onto init                                         # 3. couple this project + harness
#    -> .onto/ + genome/ + MCP registration + edit-guard hook + skill
onto court genome/genome.yaml                     # prove (ALL PROVED)
onto serve genome/genome.yaml --data .onto/data   # live organism from second zero

The harness (Claude Code) sees the MCP server immediately; it changes behavior only through propose (checkers + SMT court + semantic diff), and a hook blocks any direct edit of the genome or the printed phenotype. Dialect toolchains (e.g. Go/JVM) are pulled lazily, only at materialize.

Status: 1.3.0 — the U1-U8 constitution is closed, IR v1.0 is frozen (D72)

End-to-end scenario (IDEAL, works)

A natural-language product description -> nlfront builds a genome+acceptance -> the COURT proves -> the organism lives (warden timers, /admin, /list, /q?params, webhooks) -> any language is grown by the model (growdialect). Exams: exams/fnl.py, exams/fideal.py, exams/fgrow.py.

Where to look

What the exams have proven (exams/, all green)

the organism lives via an interpreter without code generation; one judge is green on interpreter+Go+Python; contracts are proven (COURT), 21/21 mutants distinguished, an equivalent mutant recognized; gene composition (payments in three domains byte-for-byte); metabolism by measured heat with failure-with-arithmetic; warden: molt in 0.3s, migration of 30k live events by functor, REVOKE of rights; live CEGIS with an SLM (~$0.001/skill); a membrane for others' organisms (drift -> trust revocation); 25k ev/s ingest, p99 1.8ms, kill -9 -> start from snapshot 0.01s; jsonl|sqlite store with a byte-identical fold. Running the whole battery: tools/check.sh.

Status: the engine is built and delivered exam by exam (see docs/log/PLAN.md); open — calendar operation and the external operator (F7).