Skip to Content
V1Using Lore SaaSMCP Server Setup

MCP Server Setup

Connect your AI harness to Lore via the Model Context Protocol (MCP). The Lore MCP Server is hosted at mcp.lore.infraestrutura.app using the MCP spec’s Streamable HTTP transport (spec revision 2025-03-26 and later) — not the older, deprecated “HTTP+SSE” transport.

Today this is a remote-only connection: Claude Desktop, claude.ai, Claude Code, GitHub Copilot (VS Code Desktop and Copilot CLI), OpenAI Codex CLI, ChatGPT, and Cursor all talk to the hosted server directly and log you in through Dex. No token to copy, nothing to install. A local (stdio) path — running the lore-mcp binary on your own machine — exists in the codebase but isn’t distributed yet, so it isn’t documented here.

Claude Desktop / claude.ai

  1. Open Settings → Connectors → Add custom connector.
  2. Set the URL to:
    https://mcp.lore.infraestrutura.app/mcp?auth=required
    Belong to more than one org and want every tool call on this connector scoped to one of them without passing org every time? Append ?org=<slug>&auth=required instead, e.g. https://mcp.lore.infraestrutura.app/mcp?org=laca&auth=required. This is optional — see Scoping a connection to one org below.
  3. Under Authentication, choose Sign in now (public tools work without logging in, but this connects the private-namespace path in the same step instead of waiting on a prompt that may not appear — see the note below).
  4. Under OAuth Client, choose Use your own OAuth client and set:
    Client ID: lore-mcp
    Leave the client secret blank — lore-mcp is a public client.
  5. Continue — you’ll land on Lore’s sign-in screen (GitHub, or your org’s SSO if configured). Complete it, and the connector is ready.

Why “Sign in now” instead of “Log in when needed”?

lore-mcp forwards whatever credentials it has straight through to the Registry, and only the Registry decides whether a call succeeds — lore-mcp itself never issues its own 401 challenge for a specific tool call. In practice, that means a private-namespace call that fails doesn’t reliably prompt Claude to log you in on the spot; it just comes back with an “unauthorized” message in the tool result. Choosing Sign in now sidesteps that by authenticating up front. Public packages work either way.

What does auth=required do?

It makes your AI tool sign you in when it connects. With auth=required, lore-mcp answers the first connection request with a standard OAuth challenge (HTTP 401 pointing at /.well-known/oauth-protected-resource), so Claude Code, Cursor, Codex and Copilot start the sign-in flow right away, and an expired sign-in is renewed instead of silently continuing without it. Without the parameter, the connector still works, but anonymously: public packages can be read, and anything that needs your identity (private namespaces, org memory, publishing packages) fails with a message telling you to reconnect. An anonymous call that names an org (as org, or through ?org= on the URL) still returns that org’s public packages, and ends with a note saying the connector isn’t signed in and that the org’s mandatory Rules and memory weren’t delivered. Combine it with an org scope as ?org=<slug>&auth=required.

Claude Code

Same connector Claude Desktop uses, over claude mcp add:

claude mcp add -s user --transport http lore "https://mcp.lore.infraestrutura.app/mcp?auth=required"

Claude Code identifies itself automatically, so no client ID or callback port is needed; your browser opens Lore’s sign-in as soon as it connects. Older instructions added --callback-port 8090; connectors set up that way keep working.

To scope this connection to one org (see Scoping a connection to one org below), append ?org=<slug>&auth=required to the URL instead: https://mcp.lore.infraestrutura.app/mcp?org=laca&auth=required.

Where to install it in Claude Code

claude mcp add stores the connector in one of three scopes. Pick one:

ScopeFlagStored inAvailable in
User (recommended)-s user~/.claude.jsonEvery directory, for your account on this computer
Local (Claude Code’s default)-s local or no flag~/.claude.json, keyed by directoryOnly the directory you ran the command from
Project-s project.mcp.json at the repository rootEveryone who opens that repository (committed to git)

Without -s user, the connector exists only in the directory you ran the command from. If you added it there by mistake, remove that copy with claude mcp remove -s local lore from the same directory.

For a team, commit a .mcp.json to the repository instead:

{ "mcpServers": { "lore": { "type": "http", "url": "https://mcp.lore.infraestrutura.app/mcp?org=<your-org>&auth=required" } } }

Each teammate still signs in with their own Lore account; the file only holds the URL, never a token.

For IT administrators: machine-wide configuration (managed-mcp.json)

Claude Code also reads a system-wide file, deployed by an administrator (for example through MDM):

  • macOS: /Library/Application Support/ClaudeCode/managed-mcp.json
  • Linux/WSL: /etc/claude-code/managed-mcp.json
  • Windows: C:\Program Files\ClaudeCode\managed-mcp.json

This file is exclusive. When it exists, it replaces every MCP server configured by users on that machine (they are not merged), and claude mcp add stops working. Every connector people need must be listed in it, not just Lore. Claude Code’s documentation doesn’t confirm that the oauth block (and so the fixed callback port Lore requires) is honored in this file, so test the sign-in on a pilot machine before rolling it out. For most organizations, distributing the -s user command above is the safer choice.

GitHub Copilot

VS Code Desktop

Run MCP: Add Server from the Command Palette, choose HTTP, and fill in:

Server name: lore Server URL: https://mcp.lore.infraestrutura.app/mcp?auth=required

Leave Headers and OAuth Client ID blank — VS Code discovers lore-mcp automatically via its Client ID Metadata Document, and the default request timeout is fine.

Copilot CLI

From the terminal:

copilot mcp add lore "https://mcp.lore.infraestrutura.app/mcp?auth=required"

Leave the OAuth Client ID blank — Copilot CLI discovers lore-mcp the same way VS Code does.

Copilot CLI has known upstream port-binding bugs (github/copilot-cli#4793 , #4795 , #4800 ) that can cause the OAuth callback to fail intermittently — this isn’t a Lore-side issue. VS Code Desktop isn’t affected.

OpenAI Codex CLI

Add the connector to ~/.codex/config.toml:

[mcp_servers.lore] url = "https://mcp.lore.infraestrutura.app/mcp?auth=required"

Then run codex mcp login lore. Codex identifies itself automatically, so no client ID or callback port is needed. Older setups with client_id = "lore-mcp" and mcp_oauth_callback_port = 8092 under [mcp_servers.lore.oauth] keep working.

Cursor

Cursor has no automatic OAuth client discovery — add the connector to mcp.json with an explicit client ID:

{ "mcpServers": { "lore": { "url": "https://mcp.lore.infraestrutura.app/mcp?auth=required", "auth": { "CLIENT_ID": "lore-mcp" } } } }

No client secret — lore-mcp is a public client.

ChatGPT

In ChatGPT’s connector settings (developer mode), create a connector with the URL https://mcp.lore.infraestrutura.app/mcp and OAuth authentication. ChatGPT identifies itself automatically, so leave the client ID and secret empty. It works from any ChatGPT account or workspace; your browser opens Lore’s sign-in when you connect.

To have your org’s mandatory Rules loaded at the start of every conversation, use https://mcp.lore.infraestrutura.app/mcp?org=<slug>&auth=required. The connector then tells ChatGPT to call lore_context first. ChatGPT keeps that instruction until the connector is refreshed, so refresh it in the connector settings after changing the URL.

Gemini CLI

Not yet supported. Gemini CLI only speaks to an MCP server over stdio today, which Lore doesn’t distribute a ready-to-run binary for — see the note at the top of this page. Skills published to Lore still reach it via lore pull (the CLI’s local-install path); see the CLI docs.

Signing in: GitHub or your company account

When your AI tool opens Lore’s sign-in, you choose how to sign in:

  • Continue with GitHub, for personal and community use.
  • Company account: type your work email. If your company’s Lore organization has set up its Microsoft Entra ID (365) sign-in, Lore finds it from your email’s domain and takes you straight to your company login. Your email is only used for that lookup and is never stored.

Lore remembers your last choice in this browser and offers it first next time. To skip the choice entirely, scope the connector to your org with ?org=<slug> (see below). Your tool then goes straight to your organization’s sign-in: its company login when one is configured, GitHub otherwise.

Scoping a connection to one org

If you belong to more than one org, every tool call defaults to the flat/global namespace unless it’s told otherwise. Two ways to scope a session to one org, on the hosted (http) transport only:

  • Per call — pass org: "<slug>" on any individual lore_get/lore_search/lore_list/lore_resolve call.
  • Per connection — append ?org=<slug>&auth=required to the connector URL (see the Claude Desktop and Claude Code sections above). This is captured once, the first time the connector is used, and stays fixed for that connection’s entire lifetime — there’s no tool or mid-session switch to change it, so a session working across two orgs needs a second connector instead.

Not sure which slug to use? Call the lore_orgs tool — it lists every org you belong to and your role in each. If you’re a member of more than one org and haven’t scoped your connection, tool responses also include a one-line hint reminding you lore_orgs and ?org=<slug>&auth=required exist.

Org-scoped access requires that org to be on an active Pro plan. If it’s on the free tier, an org-scoped call (by any of the methods above) fails with a message pointing at upgrading  — public content is unaffected either way, and lore_orgs itself always works regardless of plan, so you can always check which orgs you belong to.

Available Tools

ToolDescription
lore_getRetrieves the full content of a package, rendered for the active target runtime
lore_searchSearches the registry for packages matching a query
lore_resolveGiven a natural-language goal, finds and pre-fetches the most relevant packages
lore_listLists packages in a namespace, or recent public packages
lore_versionsLists all published versions of a package
lore_orgsLists the orgs you belong to, and your role in each
lore_contextLoads the org’s context: its mandatory Rules and Skills, then its most relevant memory (hosted connector only)
© 2025 Lore · Built by M1 Cloud