OpenLog CLI

Official CLI and MCP server for OpenLog.

npx -y @openloghq/cli@latest
npx -y @openloghq/cli@latest setup
npx -y @openloghq/cli@latest init
npx -y @openloghq/cli@latest login
npx -y @openloghq/cli@latest whoami
npx -y @openloghq/cli@latest mcp

Guided setup

Running openlog or openlog setup in a TTY walks through:

1. Sign in (browser device login) if needed
2. Install MCP into Codex, Claude Code, Cursor, all of them, or skip
3. Choose an MCP permission profile

Setup never inspects or connects the current directory. After setup, an agent can discover OpenLog projects without a local folder. Non-interactive shells should use the individual commands instead.

Project init

Run init inside any local folder you want OpenLog agents to identify automatically:

cd <folder>
npx -y @openloghq/cli@latest init

The interactive flow shows the detected folder type and asks whether to connect it. Git repositories keep openlog.projectId in local .git/config. General folders use .openlog/project.json with a versioned project ID. A session started from a descendant of a general folder resolves the nearest parent binding. OpenLog does not inspect the folder contents.

Init can create a new OpenLog project or bind the folder to an existing project in the selected workspace. One workspace may contain Git, general-folder, and directory-free projects. A user's same normalized repository can be connected only once. If the server project is created but writing the local binding fails, init rolls the new project back.

start_openlog_session accepts a local projectPath, an explicit projectId, or no selector. With no selector it starts automatically only when exactly one project exists. With multiple projects it returns selection_required so the agent asks the user and retries with the selected ID. With no projects, agents with write permission can preview and confirm create_workspace_project; this does not require or create a local directory.

The MCP server sends its built-in session instructions in Korean. Tool names, input fields, Capture Mode values, and response status values remain unchanged. Workspace Agent Guides are managed separately as workspace-owned Markdown.

Each workspace owns one editable English Markdown Agent Guide. The MCP server's start_openlog_session and get_workspace_agent_guide tools return the latest Guide revision. update_workspace_agent_guide replaces Guide content after a preview confirmation (confirm: true, or skipConfirmation: true when the user explicitly asked to skip). Use the Guide for durable agent behavior and recording policy; keep one-off knowledge in NOTE Logs.

get_workspace_project returns the connected project's Capture Mode. update_workspace_project_capture_mode changes it after the same preview confirmation pattern. After a confirmed change, re-read the project or call start_openlog_session again so the new mode applies in the session.

The project Capture Mode returned with the session:

• AUTO: create or update Task, Log, and Output drafts when the Guide says the work matters
• ASK: ask before those writes; this is the default
• EXPLICIT: perform those writes only after an explicit request

Capture Mode is model guidance, not a server authorization boundary. MCP permissions take precedence, and publishing, Agent Guide updates, Capture Mode updates, or deletion keep their own confirmation policy.

Human-facing output

help, setup, login, whoami, and mcp permissions use a session-log skin (ASCII logo, quiet ANSI when the terminal supports color). MCP serve stays plain stdio JSON-RPC.

 ██████╗ ██████╗ ███████╗███╗   ██╗██╗      ██████╗  ██████╗
██╔═══██╗██╔══██╗██╔════╝████╗  ██║██║     ██╔═══██╗██╔════╝
██║   ██║██████╔╝█████╗  ██╔██╗ ██║██║     ██║   ██║██║  ███╗
██║   ██║██╔═══╝ ██╔══╝  ██║╚██╗██║██║     ██║   ██║██║   ██║
╚██████╔╝██║     ███████╗██║ ╚████║███████╗╚██████╔╝╚██████╔╝
 ╚═════╝ ╚═╝     ╚══════╝╚═╝  ╚═══╝╚══════╝ ╚═════╝  ╚═════╝
session log · cli

Usage
  openlog login              Sign in from the terminal
  ...

Interactive terminals use the human-readable whoami view. Pipes and scripts keep the machine-readable /auth/me JSON; pass --json or --human to select the output explicitly. Set NO_COLOR=1 to disable ANSI styling.

Local MCP permissions

openlog mcp permissions
openlog mcp permissions set read-only
openlog mcp permissions set safe-write
openlog mcp permissions set full
openlog mcp permissions reset

safe-write is the default. Permission changes apply after the MCP server is restarted or reloaded.

MCP tool groups:

• Account and posts: authentication, notifications, authored/liked posts, image upload, post detail, and confirmed publishing
• Workspace context: project session start, workspace and project discovery, activity, working brief, tasks, logs, todos, memories, outputs, and graph links
• Project creation: confirmed create_workspace_project without a required local directory
• Safe writes: create/update, todo completion, link creation, image upload, and confirmed output/post publishing
• Full-only writes: individual workspace-document/link deletes and working-brief clear

MCP client configuration:

{
  "mcpServers": {
    "openlog": {
      "command": "npx",
      "args": ["-y", "@openloghq/cli@latest", "mcp"]
    }
  }
}

Register every supported local client, or choose one:

openlog mcp install all
openlog mcp install codex
openlog mcp install claude-code
openlog mcp install cursor

Cursor registration updates the global ~/.cursor/mcp.json while preserving existing MCP servers.