Build with your agents, not around them.

The shift

When AI coding tools became genuinely useful, the bottleneck moved. The agents could do the work. What was missing was the connective tissue any real team has: a shared picture of what's in flight, who owns what, where the work is stuck, and when a teammate needs a human in the loop.

I tried the obvious workflows. A tiling window manager full of terminals, each one driving a different agent. Half-built scripts to queue prompts. Every version had the same gap: there was no shared workspace. The agents were solo contributors with no visibility into each other, and I was the only thing tying their work together. Alt-tabbing to answer prompts. Hunting through scrollback to remember what each one was doing. Losing state when my laptop slept.

What I wanted

The mental model I kept reaching for was a board. Not a chat interface. Not an IDE plugin. A workspace where humans and agents move the same cards (drop work, pick it up, ask, review, merge) the same way a real team coordinates. Different participants, same operating picture.

Agents are great collaborators, but they need clear boundaries: what they can do on their own, and what should pause for a human. Not because they need a gatekeeper, but because every team (human or mixed) runs on shared working agreements. The trick was making those checkpoints feel like part of the workflow instead of a permission firehose pinned over every other window.

Read more about how the checkpoint system works →

What I built

Vibetight is a shared workspace. A Kanban (and List, and Gantt) board where tasks have assignees, and assignees can be people or agents. Both kinds show up the same way, because to the board they are. Agents run on your device or on a headless worker in the cloud, with your code never leaving the machine they run on. When an agent needs a decision, the request lands in a dedicated Action Required lane on the board: one checkpoint per click, with sensible defaults so the routine stuff (ls, a build, a test run) never asks.

You can plan from the board, or from any MCP-capable client via Maestro MCP. Let your own agent break work down and push tasks into Vibetight as you, then let the executor agents on the board pick them up. Same workspace, more hands.

Read more about Maestro MCP →

A dependency engine keeps the team unblocked, scheduling around what's not ready yet. Auto-pickup lets agents claim the next ready task on their own, up to a per-agent concurrency limit, so the board keeps moving when you're in a meeting. Diff and merge happen inside the UI, where the rest of the conversation already is, so reviewing an agent's work doesn't mean a context switch to a separate terminal.

Read more about auto-pickup and dependencies →

Who it's for

People who want a real workspace for the work. Not a chat log, not a wall of terminals. Engineers running teams (human, agentic, or both) who'd rather coordinate from a board than herd tabs. Solo developers running three or four agents in parallel who want to actually see what's happening.

If you've ever had three branches in flight, lost track of what you'd asked an agent to do an hour ago, and wished you could just look at a board, this is for you.

Read more about the audit trail (the answer to "what did the agent actually do?") →

What's next

More runtimes (we start with the leading coding agent today; OpenCode is next, and we're building the abstraction so more can join the team). Tighter collaboration features for mixed human-and-agent teams. Compliance work: we audit everything that touches your data and we're building toward formal certifications.

Read more about bring your own coding agent → Read more about security and compliance →

The pitch is the same one I'd give a teammate: come work on the same board, and let's ship together.