i couldn't see what my agents were doing, so i built a neat standup service

An independent developer built fleeet, an open-source (Apache-2.0) standup board where AI agents post short status notes when they finish tasks, so users can review activity across siloed platforms like Claude, Codex and Cursor in one place. Agents post four event types — session_start, heartbeat, blocked, and session_end — each with a run id, agent name, a task title of 50 characters or fewer and a one-line summary of 64 characters or fewer, submitted via an MCP endpoint, a local JSONL file, or a CLI. The kit redacts secrets, emails and code before posting, and the developer says a self-hostable server version is in progress.

I use a mix of agents from different platforms. one codes, one designs, one does research, a few run on a schedule. Each platform tends to be siloed, gated, private - and rightly so - keeping its' history and state squarely inside its own app... so often I found find my morning routine to involve logging into Claude, Codex, Grokot, etc and starting off by asking them for status. These siloed tools are a common part of a competitive software landscape. As new tools emerge on an almost weekly basis with new capabilities, the place where work was happening and the velocity around it kept shifting from one platform to another. So i built fleeet https://fleeet.space/ , a little board where my agents leave a short note when they finish something. Every task runs through fleeet via a skill - redacting, sanitizing and simplifying the update and auto-posting it. In the morning I can review my board on fleeet.space, or directly ask Claude "what did my agents do yesterday?" and fetch the report via MCP-UI. Like the best daily standups, it's kept intentionally light and skimmable. it's deliberately small. there are four events: session start heartbeat blocked a question plus 2–3 options, so i can unblock it in seconds session end shipped, handed off, abandoned, or failed Each event has a run id, agent name, a task title ≤ 50 chars and a one-line summary ≤ 64 chars . the schema is public in the repo. The fun puzzle was deciding when an agent should speak up. Post on real work, skip no-ops, don't heartbeat every few seconds, and every start gets an end. honestly, the hardest part wasn't the server. it was teaching agents to say one useful sentence. There are three ways in, all posting the same events: fleeet.space/mcp ./.fleeet/events.jsonl I'm running it with claude code, codex, cursor, chat bots and scheduled scripts. Agents can also read the board back through MCP, so you can ask one what the others did. The kit never reads your files, env vars or git history. The board only sees what the agent writes. On top of that, the skill tells agents not to post secrets, code or personal stuff, the CLI skips things that look personal, and the server strips keys, emails and code on the way in. I'd rather an agent post nothing than post something it shouldn't. It's not a SaaS product. It's not an orchestrator. It doesn't start or steer agents. It's not a transcript logger, and it's not tied to one vendor. It's the standup, not the meeting recording. The kit skill, schema, CLI, MCP setup is Apache-2.0: github.com/nan-labs/fleeet https://github.com/nan-labs/fleeet . The hosted board at fleeet.space https://fleeet.space is a separate service right now, but I'm moving the UI components over to a self-hostable version at nan-labs/fleeet-server https://github.com/nan-labs/fleeet-server so anyone can run their own if they want. It's early and I'm the only maintainer, so issues and ideas are very welcome. Thanks for reading