Flyspace

A canvas for thinking with AI over live data. Drop blocks that pull real numbers - prediction markets, economic series, crypto, news, research - wire them into personas, and ask a question that is answered from what the wires are actually carrying.

Run

npm install
npm run dev

Open http://localhost:3000.

Without any keys the canvas still runs: every built-in data source is public (Polymarket, CoinGecko, Hacker News, World Bank, OpenAlex, and others). A persona needs an LLM - Ollama runs locally with no key.

The idea

A block is a live view of one source. A wire says this feeds that. A persona is a mind whose entire context is what its incoming wires carry - so the question "what does this thing actually know?" has a literal answer you can point at on screen.

Shells are saved canvases. The Shell Store spawns pre-wired ones; the Investor shell arrives with its blocks and wires already connected.

Capability compiler

An unknown API can be compiled into a native block instead of hand-written. The compilers in src/core/capabilities/ turn an OpenAPI document, MCP tool schemas, or a structured "bring" description into a CapabilityManifest. installProposal is the only gate that registers one. Write and destructive capabilities stay pending until approveCapability, and a manifest never grants itself authority. Execution has idempotency keys and reports EFFECT_UNCERTAIN when a write may have left the process. See src/core/capabilities/DESIGN.md. Claim ceiling: this is tested against fixtures and stubbed transports; no third-party API or MCP server has been measured in production use.

Spoken canvas control

Hold Talk or Space and say a command from a fixed grammar ("wire hacker news to the analyst", "delete this", "undo"). Speech and pointer input only propose commands; the interaction engine (src/core/interaction/) decides. It refuses unknown or ambiguous names, previews a delete until you confirm, and supports undo. Claim ceiling: tested with typed-in transcripts; there is no live microphone measurement of recognition accuracy or latency.

Configuration

Keys are read server-side from process.env and are never sent to the browser. Copy the example and fill in only what you need:

cp .env.example .env

Variable	Purpose
OLLAMA_BASE_URL	Local LLM (Ollama). Default http://localhost:11434. No key.
ANTHROPIC_API_KEY	Claude for Mind / personas.
GOOGLE_API_KEY	Gemini for Mind / personas.
DATABASE_URL	Optional. Postgres for the inference ledger. Blank = off.

Only Ollama streams an answer token by token today. Anthropic and Google answers arrive whole and are shown when complete; the ledger records those runs as streamed: false.

The inference ledger (optional)

Set DATABASE_URL and run npm run db:migrate to keep a durable server-side record of every LLM execution - provider, model, status, latency, token counts, and which wired sources actually fed the turn:

npm run db:migrate
curl 'http://localhost:3000/api/inference-runs?limit=5'

# everything that made one answer, however many personas deep
curl 'http://localhost:3000/api/inference-runs/42/lineage'

When one persona feeds another, the ledger records which run was consumed, not just which block - so the full chain behind a cascade's answer stays walkable long after the upstream blocks have refetched.

Postgres owns only server-side run records (these, plus the capability broker's execution rows). The canvas - blocks, wires, shells, personas, memory - stays in IndexedDB, because it belongs to the person at the keyboard and must work with no server at all. A run belongs to the server: it is the only party that held the key, called the provider and timed it. Leave DATABASE_URL blank and the ledger is a no-op. See INFERENCE_LEDGER.md.

Hosting: this is local-first and single-user. Keys are shared server-side, and there is no application authentication, so a public URL would hand every visitor your API credits and the prompt and answer excerpts in the ledger. The decision on record is that a deployed OmniOS goes on a private network (Tailscale / WireGuard / IP allowlist), not a public one. The dev and start scripts bind 127.0.0.1. The one exception is the limited public preview at omni.syberlabs.io, keyless and run with OMNI_PUBLIC_DEMO=1; DEPLOYMENT.md records it, and the checklist that has to clear before anything more is reachable.

Development

npm run typecheck   # tsc, 0 errors
npm run lint        # eslint, 0 errors (warnings are tracked debt)
npm test            # vitest
npm run build       # the Workers bundle that deploys (vinext); dist/
npm run scan:bundle # no secret reached dist/client (build first)
npm run build:next  # the Next server the e2e harness drives
npm run test:e2e    # playwright golden path (needs npm run build:next first)

Node version comes from .nvmrc, which CI reads too, so local and CI cannot drift apart.

CI runs four jobs on every push and pull request:

Job	What it proves
Typecheck, Test & Build	the five checks above, including the Playwright golden path
Client bundle carries no secrets	builds with a canary value per secret env var, fails if one reaches .next/static
Inference Ledger (Postgres)	migrations apply and re-apply against a real postgres:16; constraints, foreign keys and the recursive lineage walk execute
Dependency audit	production deps block on high; dev-only advisories are advisory

npm run scan:bundle is the interesting one - it is the check that used to be a grep someone had to remember. See DEPLOYMENT.md.

The ledger's Postgres-backed tests are skipped unless you give them a scratch database of their own. They TRUNCATE, so never point this at the database in your .env:

OMNI_TEST_DATABASE_URL=postgres://localhost:5432/omni_test npm test

Not in this repo

• The agent surface - keyless local browser automation, /surface and /api/agent - moved to SyberLabs/omni-agent on 2026-09-01. It shared no module with the canvas.
• The Garden and the life OS - /garden, life-system domains, stability and equilibrium modelling - deleted on 2026-09-01. It was a second product in the same repo, ~15k lines whose value was gated on history that was never built. See APEX_PLAN.md §5.

The canvas is the product.

Docs

APEX_PLAN.md is the live roadmap. vision.md is the north star. TYPED_PORT_SYSTEM.md covers the port layer (declared ports are enforced when a wire is created; blocks without ports stay untyped). AGENTS.md holds the rules for agents working in this repo. INFERENCE_LEDGER.md covers the one thing Postgres owns, and why the rest stays local. DEPLOYMENT.md records the hosting decision and what CI guarantees.