Latch

The Problem

You can't see what you're not looking at.

Five AI sessions running across two projects. You have no idea what session 4 is doing. Is it still running? Did it error? Did it finish ten minutes ago while you were focused on something else? You'd have to cmd-tab through three terminal windows to find out.

~ the agentmaxxer's desk — right now

// Five agents. One glance.

PAD 1  [###]  refactor-auth       running  — 2m 14s
PAD 2  [###]  write-tests         running  — 0m 48s
PAD 3  [.#.]  deploy-staging      WAITING  — approve deploy?
PAD 4  [!!!]  migrate-db          ERROR    — type mismatch line 42
PAD 5  [___]  lint-cleanup        idle     — completed 8m ago

// You glance at your pad. Tap pad 4. Terminal switches focus. You're there.
// No alt-tab. No hunting. No wondering.

Every orchestration tool — Conductor, Ruflo, Composio, Claude Agent Teams — is screen-only. The control surface for the most significant shift in developer workflow in a decade is alt-tab. The problem isn't speed. It's awareness.

Status Mapping

Color is state. Pressure is action.

Green

Session running

Amber

Waiting for approval

Red

Error / failed

Dim

Idle / complete

Architecture

MIDI in. Status out. MCP in the middle.

Hardware

MIDI Controller

Any controller you own

→

The Product

Latch

MCP Server + MIDI Bridge

→

Sessions

AI Coding Tools

Claude Code, Cursor, etc.

Latch is not a hardware product. It's the intelligence layer between your physical controls and your AI agents.

Hardware → Software

▶ Pad press → switch to session N

▶ Pad hold → approve / reject pending action

▶ Knob turn → select AI model

▶ Fader → adjust tokens, temperature

▶ Velocity → action modifier (light tap = preview, hard press = approve)

Software → Hardware

● Session running → pad glows green

● Waiting for you → pad pulses amber

● Something broke → pad turns red

● Idle / complete → pad dims

● Encoder ring → parameter value indicator

Why Now

Three things converged to open this gap.

⚡

The Workflow Exists

Agentmaxxing isn't theoretical. Developers are running 2-5+ concurrent AI sessions daily, right now. The role of "Agent Orchestrator" is showing up in job descriptions. Collins Dictionary named "vibe coding" Word of the Year 2025.

🔌

The Protocol Exists

Model Context Protocol gives Latch a native communication channel to every AI tool that adopts it. No custom integrations per tool — the emerging standard. As MCP adoption grows, Latch's reach grows automatically.

🎛

The Hardware Exists

MIDI controllers are mature, affordable, diverse, and sitting in drawers. Decades of tactile control surfaces — velocity-sensitive pads, endless encoders, motorized faders, programmable LEDs. Available for a new job.

Product Ladder

Open source core, commercial layers earned.

1 — Ship

Latch

Open source MCP-MIDI bridge. Free. Works with any controller. The protocol layer.

Open Source

2 — Learn

Latch Pro

Learning engine: auto-assigns pads by project type, detects approval patterns, preset profiles, multi-tool orchestration.

Paid Tier

3 — Build

Latch Board

Purpose-built hardware designed from real usage data. The right pads, knobs, and layout — nothing generic.

Crowdfunded

4 — Share

Latch Hub

Community-shared configurations, controller mappings, and workflow presets.

Community

Controller Breakpoints

Responsive design for hardware.

Latch adapts to what's on your desk — the way CSS adapts to screen size. Every mapping decision starts with: "How does this work on an APC Mini?" Then scales up and down from there.

Tier 1 — Grid Only

Pure session switching

Pads only, no knobs/faders

Novation Launchpad Mini · Midi Fighter 64

$110 – $350

Tier 2 — Grid + Faders ☆

Reference device

Pads + linear analog controls

Akai APC Mini MK2 (design target)

$99 – $200

Tier 3 — Encoder-Focused

Parameter control

Rotary encoders, push-click, LED rings

Midi Fighter Twister · Arturia BeatStep

$90 – $277

Tier 4 — Dev Peripherals

Non-MIDI native

LCD keys, dials — SDK bridge required

Elgato Stream Deck · Keychron Q0

$90 – $350

Tier 5 — Full Console

Everything

Pads + knobs + faders + displays

Ableton Push 3 · NI Maschine MK3

$200 – $999

Tier 6 — DIY / Hackable

Build your own

Custom builds, community-driven

Adafruit MacroPad · Arduino/RP2040

$20 – $100

// Breakpoint cascade — design for Tier 2, degrade gracefully

Tier 5 (full console)   richest experience, all features
Tier 2 (grid + faders)  reference experience, design target
Tier 3 (encoders)       parameter-focused, pairs with grid
Tier 1 (grid-only)      core experience, status + switching
Tier 4 (macro pads)     secondary support, bridge required
Tier 6 (DIY)            community-defined, spec-driven

Launch Strategy

Open source protocol play. RFC with working code.

MCP has no physical interface layer. Every MCP server today connects AI tools to software — APIs, databases, file systems. None connect AI tools to physical hardware. Latch ships as the missing layer, not a product announcement.

Beta targets three overlapping communities

MCP Builders

Already building MCP servers, see the gap, want to extend the protocol into physical space

Claude Code Power Users

The agentmaxxers who feel the pain daily and will dogfood it immediately

MIDI-Curious Devs

Developers who are also musicians/producers/DJs — already have controllers, immediately get the vision

Competitive Landscape

Nobody has built this.

Software Orchestrators

Conductor / Ruflo / Composio

Screen-only. Same tabs and windows problem they claim to solve.

Dev Hardware

Stream Deck / QMK Macro Pads

Generic macro launchers. Not AI-native, no session awareness.

Platform Multi-Session

Claude Agent Teams / Cursor Background

Vendor-locked. Status buried in their UI, no cross-tool view.

MIDI-Desktop Bridges

BetterTouchTool / MIDI2LR / Duct tape

Hacky middleware. No protocol, no status feedback, no AI awareness.

MCP-native + any MIDI hardware + live session status + open source protocol. First physical interface layer in the MCP stack.

Design Philosophy

Software-first. Hardware is Phase 3.

Zero Barrier

If you have a controller, you can use Latch today. If you don't, an APC Mini costs less than two months of your Cursor subscription.

Product is the Protocol

What makes Latch valuable is the mapping — how pad velocity maps to approval confidence, how LEDs encode multi-dimensional status, how software learns your workflow. That's all software.

Data-Driven Hardware

If thousands of developers use Latch daily, we'll know exactly which controls they reach for, which layouts stick, which affordances are missing. Purpose-built hardware from usage data beats guessing.

Thesis Fit

"You Are an API" — AI session management is locked in opaque tool UIs. Latch extracts the orchestration control plane into physical hardware the user owns. Same inversion: locked-in interface → user-controlled intelligence layer.

Wherewithal Ventures — Business Model Inversion thesis

7/7

Stratum Founder Rubric
Real job. Recurring pain. Clear spec. Clean distribution. No hardware logistics.

Pressure Test

What could kill this — and why it probably won't.

Strengths

Category creation: no hardware solution exists for AI session orchestration

Zero barrier: works with controllers developers already own

MCP-native: reach grows automatically as the protocol is adopted

Open source wedge: protocol contribution, not just a product launch

Cognitive separation: physical surface for commanding, keyboard for typing — two different jobs

Latency advantage: MIDI is sub-millisecond, speculative LED updates keep feel instant

Risks to Watch

Native multi-session UIs: if Anthropic or Cursor ships a visual session manager, urgency for hardware drops

MCP instability: breaking protocol changes could force rewrites (mitigated: MIDI engine is protocol-agnostic underneath)

Multi-tool orchestration is hard: each tool exposes session state differently (Phase 2 problem, not Phase 1)

Pad-to-LED round trip must feel under 100ms or the device feels broken — hard design constraint

Market may be early: "agentmaxxing" is emerging, not mainstream yet

The Ask

Ship open source. Build community. Raise seed.

v0

Open source launch
MCP server + MIDI bridge
npm install, GitHub, HN

$500K

Seed round
12 months runway
Latch Pro + multi-tool

$99

Reference hardware cost
Akai APC Mini MK2
Less than 2 months of Cursor

~ next steps

// 1. Build v0
MCP server + MIDI bridge. Test with APC Mini + Claude Code.

// 2. Dogfood it
Daily use. Iterate on mapping. Discover what's missing.

// 3. Ship open source
GitHub + npm. RFC framing. Dev Twitter, Hacker News, MCP community.

// 4. Learn and layer
Usage data → Latch Pro → Latch Board → Latch Hub