Deploy and Host DSH with LongMemory on Railway

DSH is DeepSeek Harness, DeepSeek's open-source coding agent: a model that reads your files, runs shell commands and keeps working through a long task, with a browser UI. This template runs DSH 0.1.5-rc.1 on Railway and pairs it with a LongMemory server as an MCP tool server, so the agent can recall, store decisions and track task state across sessions. Built on DeepSeek Harness; not affiliated with or endorsed by DeepSeek.

About Hosting DSH

DSH's web surface is built to run on your own machine: it binds loopback only and refuses to listen on all interfaces. This template therefore puts Caddy in the same container, answering the platform health check itself and forwarding to the harness. The forwarding has one non-negotiable property — the browser's Host header passes through untouched — because DSH checks every API call and WebSocket against a trust fence that requires the request Host to be a trusted authority and any Origin to match it. The template registers your Railway domain as trusted at boot.

Sign-in is DSH's own. Each start prints a one-time URL carrying a token; opening it sets a signed 30-day cookie. That cookie's signing secret lives on the volume, so a redeploy prints a new token but does not sign you out. Your workspace, sessions, settings, credentials and any toolchain you install also live on the volume.

Common Use Cases

Run a coding agent on a machine that stays awake, and check on it from any browser.
Give the agent memory that outlives a session: decisions, task state and project context stored in LongMemory and recalled through MCP tools.
Keep a project's toolchain and repositories in one persistent place, separate from your laptop.

Dependencies for DSH Hosting

A DeepSeek API key, set at deploy time or later under Settings → Models.
Two volumes, created by the template: one for the agent's home and workspace, one for LongMemory's database.
Optionally, an OpenAI or Gemini key on the longmemory service for semantic recall. Without one, LongMemory matches on shared words rather than meaning, and says so in its log.

Deployment Dependencies

DeepSeek Harness — upstream project (MIT); documentation.
LongMemory — memory server (Apache-2.0).
Template repository (MIT) and the LongMemory template repository it builds the memory service from.
Caddy (Apache-2.0) and mise (MIT).

Implementation Details

Two services:

dsh — the only public service. node:22-bookworm-slim with @deepseek-ai/dsh 0.1.5-rc.1 installed from npm (the package hard-depends on the built web frontend), the Caddy 2.11 binary, and mise for toolchains. Health check /up, answered by Caddy. Volume at /data holding the home directory, the harness home, the workspace, the npm prefix and mise data. Runs as an unprivileged user whose home is on the volume.
longmemory — private, no public domain. Built from a pinned upstream commit, SQLite on its own /data volume, health check /health. Its API key is generated per deployment and handed to dsh as a service reference, along with its private-network address.

The LongMemory connection is a Cordis configuration overlay applied at boot when the service reference is present. The MCP client negotiates at startup and exposes thirteen tools to the model as mcp__longmemory__* — recall, ingest, remember a decision, update task state, explain, code graph and more. If LongMemory is still starting, the harness comes up without those tools and reconnects with backoff.

First run: open the dsh service's deploy logs, find the dsh web: line, and open https:///?token=… with its token. If you left the key blank, paste it under Settings → Models. An optional edge gate adds HTTP basic auth in front of DSH's sign-in when you set DSH_GATE_PASSWORD on the service.

Three things to know. Upstream has shipped only alphas and release candidates so far, and releases land every few days; the pinned version moves accordingly. Running sessions end on redeploy — files survive, in-flight agent turns do not. And the token exchange, the fence and the MCP client were all measured against the running service: an unauthenticated index request is 401, a foreign Host or Origin on the API is 403, and a cookie issued before a redeploy authenticates after it.

Why Deploy DSH on Railway?

Railway is a singular platform to deploy your infrastructure stack. Railway will host your infrastructure so you don't have to deal with configuration, while allowing you to vertically and horizontally scale it.

By deploying DSH on Railway, you are one step closer to supporting a complete full-stack application with minimal burden. Host your servers, databases, AI agents, and more on Railway.