Deploy and Host Helix Foundry with Railway

Helix Foundry is HelixDB's open-source (Apache-2.0) app for turning company data into one graph: connect databases, tools and documents, let an agent draft an ontology over them, load it into HelixDB, then explore it and ask questions with cited answers. This community template runs Foundry's 0.1 release images with a password-protected URL and the Acme Corp sample.

About Hosting Helix Foundry

Upstream ships Foundry as software for your own computer: no sign-in, one owner, seven containers sharing a network namespace, and a site that answers only localhost:3000. This template rebuilds that stack for Railway's private network. A gateway asks for a generated password before anything reaches the app and presents each request to it as localhost. Foundry and its ontology agent share one container and volume, as they share a pod upstream. HelixDB, MySQL and an S3-compatible store (SeaweedFS) run as their own services with volumes. Secrets are generated at deploy. The Acme Corp sample database is wired in, so "Use the sample" works once you add an AI key.

Common Use Cases

Build an ontology over PostgreSQL, MySQL, Stripe, GitHub, Linear, Notion, Slack or Google Workspace data, plus PDFs and notes, and explore it as one graph
Ask questions across tables and documents and get answers with charts, the queries run and a source for every claim
Give Claude Code, Cursor or your own agents an MCP server over your company graph
Trace every value back to the row or document quote it came from, and review values documents disagree on
Try ontology-driven analytics on a made-up company before connecting real data

Dependencies for Helix Foundry Hosting

Helix Foundry images by HelixDB: foundry-site 0.1.3, foundry 0.1.2, ontology-agent 0.1.1
HelixDB v0.0.14 (graph database), MySQL 8.0 (Foundry's store), SeaweedFS 4.47 (S3-compatible store)
PostgreSQL 18 with the Acme Corp sample data (optional demo source)
An AI provider: a Claude or OpenAI key, or Azure OpenAI, Amazon Bedrock, Ollama or any OpenAI-compatible server, entered in the app

Deployment Dependencies

Source and local guide: https://github.com/HelixDB/helix-foundry
Foundry docs: https://docs.helix-db.com/foundry
HelixDB: https://github.com/HelixDB/helix-db
SeaweedFS: https://github.com/seaweedfs/seaweedfs
Claude API keys: https://console.anthropic.com

Implementation Details

Service	Role	Public
Gateway	Caddy: password prompt, then the app; MCP and HTTP Ask with a bearer token	yes
Site	Foundry's web app (Next.js)	no
Foundry	API, MCP, sync/draft/load/ask workers and the ontology agent; volume `/work`	no
HelixDB	Graph database every ontology loads into; volume	no
Storage	SeaweedFS S3 API for synced tables, snapshots and documents; volume	yes (S3, signed requests only)
MySQL	Foundry's database; volume	no
Acme DB	PostgreSQL with Acme Corp's made-up sample tables	no

Storage has a public domain because your browser uploads documents straight to it, to URLs Foundry signs. Every S3 request needs Foundry's signature. HelixDB checks no key: never give it a domain or TCP proxy.

First login

Deploy. Nothing is required; optionally paste a Claude key into ANTHROPIC_API_KEY (or an OpenAI key into OPENAI_API_KEY) on Foundry.
Wait until all services are healthy. The first deploy builds five thin images on top of the published ones, which takes a few minutes.
Open the Gateway's URL. Sign in with BASIC_AUTH_USERNAME (default admin) and BASIC_AUTH_PASSWORD, both on Gateway → Variables.
Foundry opens in My workspace. Without a key, open Settings › AI models and pick a provider. Keys entered there are stored sealed, never in a variable.
Click Use the sample. It connects Acme Corp's database (accounts, contacts, subscriptions, invoices, tickets, usage), uploads its contracts and call notes, creates a ready-made ontology, loads it into HelixDB and asks a first question.

Your own data

Add a connector or upload documents, state a goal, and let the agent draft. A first draft can take 10 to 20 minutes on a hosted model. Review it, make it current and load it. Connectors may reach other services in this Railway project by their private names ({service}.railway.internal). Foundry seals every credential you give it with FOUNDRY_CONNECTOR_KEY on Foundry. Never change that key, and keep a copy: without it, the stored credentials cannot be read.

Connect your agent (MCP)

With FOUNDRY_AGENT_ENDPOINTS=on (the default), https://{your-domain}/mcp is Foundry's MCP server and https://{your-domain}/v1/ontologies/{id}/ask its HTTP Ask. Both skip the password prompt and require Authorization: Bearer {token}, where the token is FOUNDRY_LOCAL_TOKEN on Foundry. It acts as the owner of everything, so keep it in your client's environment. For example: claude mcp add --transport http helix https://{your-domain}/mcp --header "Authorization: Bearer $FOUNDRY_LOCAL_TOKEN". The in-app Set up with your agent prompt describes the Docker install and does not apply here. Set the variable to off to keep these routes behind the password.

Custom domain

Add it to Gateway, then set HELIX_LOCAL__SITE_ORIGIN on Foundry to https://{your-domain}. Also add the domain to STORAGE_ALLOWED_ORIGINS on Storage, or uploads fail.

Scaling and limits

Keep every service at one replica: Foundry's workers, HelixDB and the store keep state in-process or on their volumes. Scale vertically. Foundry needs about 1 to 2 GB while drafting or loading, so Hobby or Pro is recommended. On Free (0.5 GB) or Trial (1 GB), large drafts and loads can be stopped for lack of memory. FOUNDRY_WORK_MAX_GB (default 2) caps the snapshots one draft or load unpacks into the 5 GB Hobby volume. Foundry's own spend caps on your keys apply: $20 per load job, $100 a month for loads, $30 a month for questions.

Upgrades

Each image is pinned by tag and digest in a service variable: FOUNDRY_SITE_IMAGE on Site, FOUNDRY_IMAGE and FOUNDRY_AGENT_IMAGE on Foundry, HelixDB's image source, FOUNDRY_STORAGE_IMAGE on Storage. Upgrade the three Foundry images together to the set upstream's local/compose.yaml names, after reading its changes. Foundry migrates MySQL at start, and migrations only go forward, so back up the volumes first. The Site checks at build that the sample host it rewrites is still where it expects.

Costs

Estimated at Railway usage prices: about $12/month idle and $30/month with regular syncs and questions. Foundry reaches the store through its public URL, so synced data counts as egress. Model calls are billed by your AI provider.

Why Deploy Helix Foundry on Railway?

Railway runs the whole Foundry stack, from the web app and agent to the graph database, store and sample, in one project with HTTPS, volumes, private networking and logs. You get a shareable, password-protected Foundry instead of one bound to localhost, plus an MCP endpoint for your agents. This is a community template, not affiliated with or endorsed by HelixDB. Helix Foundry and HelixDB are HelixDB's open-source projects under the Apache-2.0 licence.