One command in a repository you have lived in for years, and the tool you already run starts answering about the whole codebase, every claim opening the file and line it stands on. No new editor, no new habit: the project remembers, your tool just got a memory.
WebDesktopCLIthe web takes a zip or a GitHub connection
runlog link ~/code/some-repo
runlog syncSeeThe same quote-and-confirm as a folder of documents, and the project filling with declarations rather than pages: source files read at declaration grain, so a function asked for by name is cited at its own span, not the file’s head. What your .gitignore names stays home.
WebDesktopCLIthe snippet is the same everywhere
In the workspace, open Integrations → AI tools and copy the snippet for the tool you run. It carries the same auth as your sign-in, so no separate key is in play. Then ask: What does this repository do?
claude mcp add --transport http runlog-memory https://runlog-7613480744.us-central1.run.app/v1/mcp/ext{
"mcpServers": {
"runlog-memory": {
"url": "https://runlog-7613480744.us-central1.run.app/v1/mcp/ext"
}
}
}https://runlog-7613480744.us-central1.run.app/v1/mcp/extSeeThe tool calls memory_query and answers with the files and lines it rested on; what you tell it to remember lands through memory_write. Count the screens between sign-up and the first cited answer inside your tool: the project, the snippet, your tool.
WebDesktopCLIthe same thread is open in the desktop app and the web
runlog "what does the sweep do" # one ask, answered with the declarations it rested on
runlog # converses; resumes the project's coding thread where it left off
runlog --mode plan # this session proposes and shows; nothing is writtenThe conversation is the project’s reserved coding thread, so the next invocation resumes where the last one ended, from this terminal, the desktop app or the web page alike. A run lays itself out step by step and asks a clarifying question when the ask underdetermines the work.
--mode picks the agreement fresh each session: Plan proposes and displays, Manual asks per write and per command, Edits applies file edits and asks before commands, Auto asks for neither. Deny and allow rules authored as prose in .runlog/permissions.md apply in every mode, and a deny always wins. A file edit the person confirmed in the app is applied by the next runlog sync, before the folder is read.
Building against the API instead? Ask with citations is the same read, from your own process.