mercury-agent

Mercury Agent is an open‑source AI assistant that lives on your own computer and can be reached from a terminal, a web browser, or messaging services such as Telegram, Discord, Slack, and Signal. The tool positions itself as “soul‑driven,” meaning it builds a personal memory of how you work, what you value, and the decisions you have made, then uses that memory to act on your behalf when you are not present.

The core of Mercury's operation is a SQLite database stored in ~/.mercury that holds ten different memory types, full‑text search indexes, and markdown files you control (for example soul.md or persona.md). You can query the memory directly, for instance by running `search_memory "weekly update"` which returns hits formatted in your preferred bullet‑point style, ready to be posted to a channel like #eng‑leads. Reads happen without prompting, while any command that writes, runs a shell command, or accesses external accounts triggers a permission dialog. You can choose “approve once,” “always approve,” or “deny,” and the choice can be remembered per command type. A built‑in shell blocklist prevents dangerous operations such as `rm -rf /` from ever executing, regardless of the mode you select.

Mercury ships with roughly fifty built‑in tools that are permission‑gated. Typical inputs include natural‑language requests such as “draft the weekly update for the team” or “run `npm test` in the settings folder.” The agent translates the request into a series of low‑level actions, reading files, editing them, committing to git, or invoking a shell command, and then presents the result. Output is shown in the console or sent back through the chat channel you used, often with a completion banner that indicates whether the task succeeded, was verified by a build or test, or was paused because a blocker was encountered.

For continuous operation, Mercury can be installed as a system daemon with a single `mercury up` command. The daemon starts on boot, restarts after crashes, and can run scheduled cron jobs. Persistent “Mercury Bots” can be defined, each with its own model and memory, allowing you to create specialized assistants, for example a daily‑brief bot that posts a summary to Telegram at 07:00, or a fact‑checker that validates sources before posting.