1 · Install dependencies

cd packages/orchestration && pnpm i && cd ../..
for service in data paper research factor evolver; do
  (cd "services/$service" && uv sync)
done

Docker self-hosting (recommended)

This is the supported Docker path for a personal server or local machine. It builds the complete stack locally: PostgreSQL/TimescaleDB, Redis, migrations, data, paper, research, factor, evolver, Mastra, and the Operator Console.

Prerequisites: Docker Engine with Docker Compose v2, Git, and OpenSSL. Clone the repository, then initialize the local-only environment file and start the stack:

git clone https://github.com/mirror29/inalpha.git
cd inalpha
bash scripts/selfhost.sh init
bash scripts/selfhost.sh up

The console is available only on the host at <http://127.0.0.1:3001>. Wait until bash scripts/selfhost.sh status shows the application services healthy, then create the first login account. The prompt reads the password without echoing it or placing it in shell history:

bash scripts/selfhost.sh create-user --email you@example.com

For a public deployment, create the review account with the admin role (use your existing owner subject if it must keep the current account data):

bash scripts/selfhost.sh create-user --email admin@inalpha.dev --roles admin --subject console:dev

Visitors can then choose Request a trial on the sign-in page. Registration creates a pending account without accepting a password. Approval in Trial Waitlist generates a one-time activation link: send it to the applicant's stated email address to verify mailbox ownership. The applicant sets a password from the link and only then becomes active. Links expire after 48 hours and can be regenerated; rejected applications remain blocked. Because the application endpoint is public, keep a persistent per-IP rate limit or challenge enabled at the reverse proxy / CDN layer; the in-process limits are only a second line of defence.

Sign in at <http://127.0.0.1:3001>, open LLM Settings, and add your provider, model, and personal API key. The Dashboard encrypts it with LLM_CONFIG_ENCRYPTION_KEY; the orchestrator and Evolver resolve that owner-scoped credential. The standalone Research service still uses the deployment-level LLM_PROVIDER / LLM_MODEL and matching provider key in infra/.env.selfhost, so configure that block if you need deep dives and treat it as a shared-credential boundary until per-owner propagation lands.

Useful operations:

bash scripts/selfhost.sh logs [service]
bash scripts/selfhost.sh status
bash scripts/selfhost.sh down

Public deployment

The self-host Compose file intentionally exposes only 127.0.0.1:3001. For remote access, place your own TLS-terminating Caddy, Nginx, or Cloudflare Tunnel in front of the Dashboard and proxy only that address. Never publish PostgreSQL, Redis, Mastra, or the Python service ports, and do not expose the login or API-key settings page over bare HTTP.

Quick Start

2 · Configure your LLM key (required)

A single .env at the repo root is read by Mastra (TS) and all Python services. Copy the template and fill in the LLM provider you want to use:

cp .env.example .env

Inside .env, set LLM_PROVIDER to one of deepseek | anthropic | openai | gemini | kimi | zhipu | ollama and fill in the matching key.

Defaults pick each vendor's current flagship as of 2026-05. Override with LLM_MODEL=... if you want a reasoning / cheaper variant.

Override the default model by setting LLM_MODEL=... in the same file. Mastra and services/research both read this one file — no per-service config to juggle.

If you want to run Evolver, also generate its Ed25519 credential-grant keypair and place the DER base64 values in EVOLUTION_CREDENTIAL_PRIVATE_KEY_B64 / EVOLUTION_CREDENTIAL_PUBLIC_KEY_B64. The exact commands are in services/evolver/README.md. Evolution is fail-closed without these keys and currently accepts only the priced default model for each of DeepSeek, OpenAI, Kimi, and Zhipu; ordinary chat can still use other models or custom proxies.

Already have keys in services/*/.env or packages/orchestration/.env from earlier? Those still work as cwd-level overrides while you migrate. Once you copy them up into the root .env, the per-service files can be deleted.

Optional · FRED key for macro factors. The factor library's macro factors (macro.* — rates, term & credit spreads, CPI, payrolls, real-economy, sentiment) read FRED data via venue=fred. Set FRED_API_KEY in .env to enable them — it's free and instant. Without a key the connector simply isn't registered and macro factors degrade gracefully (price/volume factors are unaffected). Note: macro factors are computed only at timeframe=1d/1wk — they're filtered out on intraday bars (monthly series would be a step function), so request 1d to see them.