GenesisCode

Learn, build, and verify software with the deterministic, AI-first GenesisCode language.

AI-first · pure kernel · explicit effects

A language agents can reason about.

GenesisCode is a deterministic, self-hosting programming system designed for software written, reviewed, and iterated by AI agents. This site is the learning path, exhaustive reference, and direct window into every normative authority.

Run your first programOpen the full referenceLoad agent context

v0.2 pre-1.0deny-by-default effectscommit-stamped docs

176frozen symbols

188host operations

201diagnostics

350authority sources

Start in fifteen minutes

The shortest honest path is build, run one pure module, then inspect its canonical result. Nothing in this first program receives ambient filesystem, network, process, clock, or model access.

1. Verify the pinned toolchain.
2. Build the production CLI.
3. Evaluate through the selfhost engine.
4. Continue into effects only when the pure boundary is clear.

Follow the verified quickstart

cargo build -p gc_cli --locked
./target/debug/genesis eval hello.gc --engine selfhost

(def greet
  (fn (name)
    (prim str/concat "hello, " name)))

(greet "agent")

"hello, agent"

Choose your path

Learn the mental model

Start with CoreForm, immutable data, functions, modules, and the boundary between pure evaluation and host effects.

Take the language tour

Ship a real workflow

Use packages, capability policies, deterministic logs, replay, tests, and evidence-backed acceptance.

Follow the practical tutorial

Retrieve compact authority

Use the agent card, task cards, exact symbol lookup, and safe repair protocol without flooding context.

Enter the agent loop

Look up anything

Search every symbol, host operation, diagnostic, CLI surface, format, example, specification, and schema.

Browse exhaustive reference

One execution model, visible end to end

flowchart LR
  A["Agent or human"] --> S["Source + package intent"]
  S --> C["Canonical CoreForm"]
  C --> K["Pure Gλ kernel"]
  K -->|"sealed EFFECT"| R["Deny-by-default runner"]
  R --> H["Explicit host operation"]
  H --> L["Deterministic log + evidence"]
  L -->|"replay / verify"| A

The same identities connect source, policy, diagnostics, logs, replay, and acceptance evidence. An agent can retrieve a compact task card first, then expand only the exact symbol or failure record it needs.

What makes it different

The kernel never performs filesystem, time, network, process, graphics, GPU, or LLM work. Programs describe effects; a deny-by-default host runner decides, records, and can replay them. Seals make protocol values unforgeable. Canonical forms and hashes make artifacts comparable. Obligations make acceptance evidence explicit.

Layer	Responsibility	Trust posture
Gλ kernel	Evaluate immutable CoreForm, closures, primitives, and seals	Pure, deterministic, minimal TCB
Prelude	Contracts, effect programs, standard helpers, protocol constructors	Trusted language layer
Effect runner	Enforce capabilities, execute hosts, log facts, replay	Deny by default
Toolchain	Format, typecheck, test, patch, package, optimize, build	Selfhost-first and evidence-gated
Agent layer	Compact cards, exact lookup, plans, safe repair	Machine-readable and bounded

No ambient authority

Host access is an explicit request checked against a capability policy, not a library call hidden inside evaluation.

Replayable decisions

Effect logs retain the facts needed to reject policy, schedule, input, or result drift during replay.

Agent-sized retrieval

Core cards, task cards, symbol indexes, stable diagnostics, and JSON schemas prevent whole-manual context flooding.

Know what is ready

GenesisCode is intentionally labeled pre-1.0. Capability maturity, compatibility, and evidence are published rather than implied. Use the feature matrix to distinguish implemented, experimental, partial, and planned surfaces; use the roadmap for work that is not yet complete.

Use nowPure evaluation, canonical forms, explicit effects, capability policy, deterministic logs and replay, packages, structured diagnostics, and bounded agent workflows. Evaluate carefullySelfhost closure, performance tiers, broad platform targets, registry GA, flagship applications, and the remaining v1 guarantees.

Documentation is part of the product

This site does not replace the repository specifications. Curated pages teach; generated references enumerate; canonical Markdown and JSON remain authoritative. Every page links back to source, and CI rejects stale generated catalogs or a broken render.

Learn how to use this site · Read the documentation topology · Inspect compatibility · Follow the roadmap