Skip to main content

Side Projects

2026

Mesh

Mesh

Runtime coordination for AI agents

View Live

Overview

Mesh is a coordination layer for AI coding agents, aimed at developers and engineering teams who run more than one agent on the same codebase. Left alone, agents edit the same files, redo finished work and lose what they learned when a session ends. Mesh gives them a shared project brief, exclusive file claims, a typed work log and review steps, and gives people a dashboard to watch and step in. Agents connect from tools such as Claude Code, Cursor and Codex, and the product is currently in alpha with a waitlist.

A dark terminal-style panel on the Mesh home page showing a sample session start: project status, board counts, open threads, active file claims, registered agents and recent log entries, ending with an offer to pick up an unassigned ticket.
What an agent is told when it connects: the project, the board, who holds which files, what the last few log entries say, and a ticket it could pick up.
The Project Board and Claims section of the Mesh home page, with a card listing three files marked CLAIMED, 409 CONFLICT and AVAILABLE beside the headline 'Agents claim work before touching it'.
File claims, explained as an interface: a second agent that asks for a file already held gets a conflict response instead of a silent overwrite.
The Features page showing 'Every change reviewed. No exceptions.' with a review-gate code card and a list of review modes, above a section titled 'Define done. Before work begins.'
Review gates: the agent that built a change cannot approve it, review strictness is a project setting, and acceptance criteria are defined before work starts.
A large gradient card titled 'Step in anytime' containing a preview of the Mesh dashboard Feed page, with panels for the project, agents, ledger entries and conversation threads, and the edge of the next card visible at right.
The 'Step in anytime' card in a horizontal carousel previews the dashboard, where people read the ledger and threads and step in.

Who We Designed For

Developers and engineering teams running more than one AI coding agent on the same codebase

  • Solo developers

    One person running one or a few agents who wants each session to start informed instead of re-explaining the stack, with review settings that can tighten as the project grows.

  • Development teams

    Teams whose agents work in parallel and need to avoid collisions, see who holds which files, and have one agent review another's work.

  • Engineering leaders

    People accountable for what agents do, who need a record of every action, visibility into progress and blockers, and a say in how much review is required.

  • Platform builders

    Teams wiring Mesh into their own agents and tooling, who need consistent hand-offs, live event streams and exportable audit data.

Tech Stack

On Mobile

Mobile view of the Features page: the word 'scale.' in large white type over a green and blue gradient, with a short tagline and two buttons.
The one-word wordmark hero scales down to a phone, with a floating tab bar for navigation.
Mobile view of the sample session-start terminal panel, with the status lines wrapping to fit a narrow screen.
The session-start readout reflows to a narrow screen with its color coding for done, in progress and blocked intact.
Mobile view of color-coded ledger entries (DONE, PROGRESS, BLOCKED, STATUS) above the heading 'Agents claim work before touching it'.
Ledger entries are color-coded by type: done, progress, blocked and status.

Process

  1. Core coordination serviceMarch 15, 2026

    The first commit scaffolds the web app. The 66 commits dated that day add the pieces the product still centers on: the project brief, file claims, the work ledger, threads and handoff, plus sign-in for people and keys for agents. A rule that makes ledger and audit entries unchangeable appears in the same first batch.

  2. Dashboard, billing and marketing siteMarch 15 to early April 2026

    Design specs for the dashboard, board, settings, admin panel, add-agent flow and billing are dated March 15 to 19. A review system with its own dashboard screens followed on March 27. Between March 31 and April 2 the marketing pages were rebuilt to match new designs, including the animated gradient heroes.

  3. Hardening and a design systemApril to early May 2026

    This period added a GitHub integration, a coordinator that can use Anthropic, OpenAI and Google models, security hardening sprints, mobile optimization of the admin, marketing, review and thread screens, and a 'refined design system' on April 20.

  4. Compliance, MCP connection and review toolsJune to October 2026

    From June 10 the work adds compliance scoring for agent sessions and an agent-first ticket format. An MCP server was scaffolded on June 23, with a design spec the same day, and a standalone repository for it was created on June 25. It lets Claude Code connect with a browser sign-in instead of pasted tokens. Sprint tools were added on September 1, a review pass followed on September 27 and 28, and ticket-approval tools landed on October 12.

AI Collaboration

Tools

Claude

Approach

Claude is credited as co-author on most commits, starting with the first scaffold. The product is built to coordinate this same kind of agent, and its repository carries instruction files for the coding agents that work on it.

Features

Shared memory

  1. A project brief that every agent reads before it starts, so a new session does not begin from zero
  2. A context snapshot loaded at session start: the brief plus live claims, open threads, recent log entries and the board
  3. A write-once ledger where agents record what they found, decided, changed and tested
  4. Threads for questions that outlive a session, so the next agent can answer instead of guessing
  5. Handoff to close out work: a summary and test results, with claims released and the ticket moved to review

Collision prevention

  1. Exclusive file claims: a second agent that asks for a claimed file is refused and told who holds it
  2. Starting a ticket claims its files in one step; if any file is taken, nothing changes
  3. Claims expire automatically if an agent disconnects without releasing them
  4. A kanban board where claimed work looks different from available work
  5. Deadlock detection that flags agents waiting on each other

Human oversight and trust

  1. Review modes set per project, from Light (trust the agent) to Strict (people approve everything)
  2. The agent that built a change cannot approve it
  3. Acceptance criteria set at project level and acknowledged by the agent before work begins
  4. A human override that requires a written reason, recorded in the review history and the ledger
  5. An append-only audit trail with JSON export

Design and experience

  1. One-word wordmark heroes (scale., control., trust.), each over its own grain-textured gradient
  2. Features shown as interface panels instead of illustrations: a status readout, a claim-conflict card, color-coded ledger entries
  3. Monospaced type for headings, labels and code, with // as the list marker
  4. A distinct concentric shape for each audience on the Solutions page
  5. Documentation written for people and agents, with a machine-readable index

Cool idea? Let's collaborate.

Have a passion project or experimental concept? I love exploring new ideas and building things that push creative boundaries.

Start a Conversation