Self-hosting

Run your own LoreKit backend — full self-host on your own Supabase project, or self-host plus BYOD for memory rows in a second database. Compare the three setups and connect your agent to your own endpoint.

Tip:

The hosted instance at lorekit.io needs none of this; it's ready to use today, at this MCP endpoint:

https://pqokxlhvnosogizsjztg.supabase.co/functions/v1/mcp

This page is for running LoreKit on infrastructure you control.

Compare the three setups

	Hosted	Full self-host	Self-host + BYOD
What you run	Nothing	Your own Supabase project, all 8 Edge Functions	Your function project, plus a second Supabase project for storage
Where memories live	LoreKit's hosted project	Your project's own database	The second (storage) project's database
Dashboard / REST / hooks see your memories?	Yes	Yes	No — they all read the function project's own database, never the BYOD storage project
BYOD available?	No	N/A	Yes — layered on top of a full self-host deployment

Full self-host is the right default unless your memory rows must live in a separate database — data residency, an existing Supabase subscription billed separately, or org policy. BYOD is never a substitute for deploying your own function project; it only changes where rows are written, and the hosted endpoint has no BYOD at all.

Deploy your own backend

Follow the Installation Guide on GitHub — it covers linking a Supabase project, applying migrations, setting Edge Function secrets, deploying all 8 Edge Functions (not just mcp + health — the memories function is required for hooks, the remote CLI commands, and the dashboard's REST API), and generating an API token on a dashboard connected to your project.

supabase link --project-ref <your-project-ref>
pnpm nx db:push supabase
supabase functions deploy --no-verify-jwt --project-ref <your-project-ref>

Verify with a health check:

curl https://<your-project-ref>.supabase.co/functions/v1/health
# → {"status":"ok","version":"1.0.0","db":"ok","ts":"…"}

(Optional) Bring your own database

Only if your memory rows must live in a database you already run. Apply the LoreKit schema to that second project and set LOREKIT_STORAGE_URL + LOREKIT_STORAGE_ANON_KEY + LOREKIT_STORAGE_SERVICE_KEY as Edge Function secrets on your function project (not in .mcp.json — that env never reaches a deployed function). Follow the BYOD guide for the exact schema commands.

Remember: the dashboard, REST API, remote CLI commands, and hooks all read the function project's own database. None of them see rows written to the BYOD storage project.

Point your agent at your endpoint

One endpoint setting covers MCP, REST, hooks, and the remote CLI commands, because the CLI derives the REST base from the MCP endpoint:

npx @lorekit/cli install \
  --endpoint https://<your-project-ref>.supabase.co/functions/v1/mcp \
  --token    lk_rw_<your-token>

(or export LOREKIT_MCP_URL instead of passing --endpoint every time — install reads only --endpoint, then LOREKIT_MCP_URL, then LOREKIT_ENDPOINT. A committed "mcp.endpoint" in .lorekit.json is never read by install. The remote CLI commands and hooks check the project's .mcp.json first, then the global ~/.claude.json, and use mcp.endpoint only when neither has a LoreKit entry. A leftover hosted install --global entry in ~/.claude.json therefore still wins over a committed mcp.endpoint, so re-run the global install with --endpoint too.)

Warning:

Almost every npx @lorekit/cli install rewrites the MCP entry — a new project, a global install, --force, --hooks, --mcp-json, --completions <shell>, or a re-run after an upgrade adds a skill — and without --endpoint it writes the hosted URL. Only a plain re-run on a fully installed project leaves .mcp.json alone. Always pass --endpoint, or export LOREKIT_MCP_URL in your shell profile, on every install.

Use the guided runbook

The Claude plugin ships a slash-only /lorekit-self-host skill that walks every step above as a gate — do the thing, verify it, stop and diagnose if the verification fails — and links to the docs instead of duplicating their commands. Claude plugin users get it automatically with /plugin install lorekit-memory@lorekit. Everyone else can fetch it directly:

curl --create-dirs -o ~/.claude/skills/lorekit-self-host/SKILL.md \
  https://raw.githubusercontent.com/mthines/lorekit/main/packages/cli/skill/lorekit-self-host/SKILL.md

Note:

Next steps: once you're connected, see Remote storage for token scoping and CI, which work the same way against a self-hosted endpoint. The GitHub App does not: it is off until you register your own App, point its webhook at your project, and set GITHUB_APP_ENABLED=true — see the GitHub App guide.