Connect your AI assistant
Native OAuth first, LiquidLM CLI login and stdio second, PAT last.
Client setup guides
- Connect LiquidLM in ChatGPT web — ChatGPT web setup: native OAuth first, CLI-managed OAuth second, PAT only as a last fallback.
- Connect LiquidLM in Codex — Codex setup: native OAuth first, CLI-managed OAuth second, PAT only as a last fallback.
- Connect LiquidLM in Claude Code — Claude Code setup: native OAuth first, CLI-managed OAuth second, PAT only as a last fallback.
- Connect LiquidLM in Cursor editor and CLI — Cursor editor and CLI setup: native OAuth first, CLI-managed OAuth second, PAT only as a last fallback.
- Connect LiquidLM in Claude Desktop — Claude Desktop setup: native OAuth first, CLI-managed OAuth second, PAT only as a last fallback.
- Connect LiquidLM in OpenCode — OpenCode setup: native OAuth first, CLI-managed OAuth second, PAT only as a last fallback.
Choose authentication
- 1. Native OAuth is always the default when supported by the client.
- 2. Use liquidlm login plus liquidlm mcp serve when the client can launch local commands and native OAuth is unsuitable.
- 3. Use a scoped PAT only when both OAuth paths are unsuitable. Every guide provides a client-specific example or explains the platform limitation.
- Headless operation alone does not require a PAT: the CLI can refresh an existing OAuth session. Initial authorization still requires user interaction.
- Configuration review is not a fresh end-to-end client certification. OpenCode native OAuth refresh remains unresolved; validate your installed client version.
Project conventions for assistants
- The canonical LiquidLM conventions (search-first retrieval, citations, scoped vaults, temporal updates, privacy rules) live in the public liquidlm/skills repository with per-client adapters and a reversible installer.
- CLI 0.2.0-alpha.1 and newer: liquidlm mcp install --client <client> --with-skill writes the MCP configuration and installs the conventions block (AGENTS.md, CLAUDE.md or .cursor/rules) in one reversible step.
- Without the CLI: node install.mjs --client <client> from the skills repository writes a checksum-guarded managed block; user edits are detected before replacement and uninstall is clean.
# one step: MCP config + project conventions
liquidlm mcp install --client opencode --with-skill
# conventions only
liquidlm skill --client opencode