Case study
Tetherlab
An MCP-based coordination layer that lets AI coding agents see each other's intent before they write code — we replace the merge-conflict model with intent-level coordination.
- Role
- Co-founder & Engineer
- Period
- 2026-04 → Present
- Status
- Still in progress
- Topics
- 5
The problem we started with
When you run more than one AI coding agent in the same repo (Claude Code and Cursor, or Codex CLI alongside Gemini CLI), they each have their own private view of what they're about to do. They plan in isolation, write code in parallel, and produce merge conflicts that look nothing like the ones humans write, because each agent already believed the other was going to do something else.
The conventional fix is "git is good at merging, let the agents fight it out." That works for trivial cases. For non-trivial work — refactors, shared interfaces, schema migrations — it produces chaos.
What we built
Tetherlab is a small MCP server that every participating agent connects to. Each agent publishes its intent (which files it plans to touch, which symbols it plans to introduce or rename, what assumptions it's making about neighboring code) before it writes anything. Other agents see those intents in real time and can adjust their own plans — or call out a conflict explicitly — before either of them spends tokens producing code that will not fit together.
The product is live in open beta with paying trial users. The current wrappers cover Claude Code, Cursor, Codex CLI, Gemini CLI, and OpenCode.
What I owned
Three of us co-founded the company. My slice:
- The MCP server itself — the canonical intent store, the conflict-detection logic, and the protocol extensions we've prototyped on top of standard MCP.
- Two of the agent wrappers (Claude Code and Cursor) — the part that hooks each tool into our coordination layer without forcing the user to change how they invoke the agent.
- A meaningful chunk of the go-to-market — positioning, landing-page copy, technical content, and the early user conversations.
The other two cofounders own the rest of the engineering surface and most of the customer-facing work. We share product and roadmap decisions across the team.
What was hard
We over-invested in a custom protocol extension before we'd validated that the basic intent-publishing primitive solved the user's problem. Two months of protocol work would have been better spent as one month of "ship the simplest thing that coordinates two agents" followed by user interviews. The protocol exists, but the v1 should have been thinner.