Glow · Dec 2024– · Case study

Building the product and the automations people use it for.

Glow brings people and AI agents together on a CRDT-backed multiplayer automation canvas. Our vision at Glow is that anyone can build any automation or business process by voice, with multiple people and AI agents working on one canvas at the same time. With the team, I help shape product direction and designed our MCP and live-canvas agent workflow.

Read the documentation Product changelog See the automations ↓

In brief

Problem

Black-box chat vs. manual wiring

Teams need more than an opaque chat prompt or slow manual node wiring: they need a shared canvas where multiple people and AI agents can build, understand and iterate on automations at the same time.

My contribution

Multiplayer + Voice + MCP

With the team, I designed our MCP and live-canvas agent workflow, define and test product behavior, write code, build client automations and write the full documentation.

Result

Live in production

CRDT multiplayer and Glow MCP are live in production. Teammates and voice-driven agents build, validate, self-test across any MCP server and parameterize automations on one canvas.

Stage: Pre-seed
Role: Founding Product Engineer
Team: #4
Since: Dec 2024
Platform status: In production

My contribution

I propose and define features with the team, including multiplayer and the design for how external agents build, validate and test on the live canvas over MCP, and test the results.
I work with clients on requirements, onboarding and automation proposals, and handle support and sales demos.
I built the travel intake automation below. I also write code and build internal tooling.
I write every page of the documentation at docs.getglow.ai and record product walkthroughs.

Product decisions

Two examples of the direction I advocated and the product behavior I helped define with the team.

A shared CRDT canvas for multiple people, voice and AI agents

Our vision at Glow is that anyone, technical or not, can build any automation or company process by voice, working alongside multiple teammates and AI agents on the same live canvas at the same time. Because the canvas is backed by CRDTs, everyone sees and edits the same reality in real time instead of waiting for an opaque chat reply.

In that loop, the person is the conductor: setting intent on a sticky note or by voice, stepping in to switch a model or rule in place, and deciding what ships, while the connected AI agent handles the heavy lifting of wiring steps, testing and refactoring into templates.

Neither the canvas nor the agent blocks the other: teammates and AI agents keep editing, reviewing and iterating on the business process together.

The agent is the brain; Glow MCP is the orchestrator, validator and test harness

Pure chat agents are a black box, while manual node wiring is slow. That drove my design for Glow MCP with the team: let an external AI agent handle voice and reasoning, while Glow MCP enforces step schemas, deterministic layout, version safety and live-workflow mutation guards.

Because MCP is composable, the same voice agent can pair Glow MCP with any other MCP server (such as Gmail, Slack, CRM or database MCPs) to send realistic test events, run the automation end to end and verify the output on the canvas without manual mocking.

At TrustSoft, I spent four years working across enterprise cloud environments. That experience informs how I think about permissions, client separation and the work involved in operating a shared platform.

Client & product automations

Two automations built in Glow: a personalized travel intake automation in client testing, and the multiplayer voice-and-MCP workflow from a sticky note to cross-MCP testing and template parameterization.

Boutique travel agency

A tailored form for each inquiry

Client testing

A Glow automation that turns an inquiry email into an individually customized Tally form, with questions and a greeting shaped by that trip.

The client can test generated forms in their own Glow account and give feedback on the questions and tone.

From inquiry to review

Read the inquiry
The client supplies an inquiry, and an AI step separates confirmed trip details from missing information to guide the questions that follow.
Tailor the questions
A second AI step drafts the greeting, questions and answer options for that specific inquiry, using the trip details and the agency's guidance.
Create the form
Python builds and validates the Tally blocks, then a separate HTTP step calls the Tally API to create the individual form.
Review the form
The founder can inspect the generated form during client testing and must approve its questions and tone before sharing it with a traveler.

Client testing only, with no automatic sending; founder approval is required before a form reaches a traveler.

Brief & technical details

The brief

The agency plans custom journeys and needs follow-up questions that reflect what each traveler has already shared. I built the intake automation around a separate Tally form for each inquiry, while keeping approval with the founder. Partner briefing is a later phase, outside this delivery.

What shapes the form

Build on the inquiry: The prompts distinguish confirmed facts from preferences that need detail, so questions can explore the trip rather than repeat dates or group size.
Keep options relevant: Destination choices stay within the requested region, and budget options reflect the destination rather than assumptions about the traveler's wealth.

Voice + MCP · CRDT Multiplayer Canvas

From a sticky note and voice to a tested, reusable automation

Live in production

Glow MCP documentation ↗

With the team, I designed the workflow for building, validating and testing automations on Glow's shared canvas through MCP. Because the canvas is backed by CRDTs, multiple teammates and external AI agents operated by voice can build and iterate at the same time.

Instead of waiting on a black-box chat reply or wiring nodes by hand, people and AI agents co-edit one live canvas, run end-to-end tests through any external MCP server and refactor working flows into reusable templates.

From canvas note to tested template

Write or speak the brief
Anyone on the team drops a sticky note on the shared CRDT canvas and speaks to the agent. The agent reads the note and current graph over Glow MCP.
Build, validate & tweak live
Glow MCP validates step schemas and lays out the graph on the canvas while teammates can immediately adjust models or settings in place.
Test across any MCP
The agent uses a second MCP server (such as Gmail MCP sending a realistic Linear inquiry) to trigger a real run and verify the resulting draft and Slack summary.
Branch or parameterize
By voice or from a second sticky note, the agent branches logic or extracts hardcoded values into a Custom Variables step so the automation becomes a reusable template.

Runs under scoped workspace permissions, credit caps and live-workflow mutation guards, with a write-free preview mode available before live execution.

Brief & technical details

Why I designed it this way

Our vision at Glow is that anyone can build any automation or business process by voice, together with teammates and AI agents on one shared canvas at the same time. In the workflow I designed with the team, the human is the conductor: you connect any AI tool, voice client or IDE as the reasoning brain, while Glow MCP acts as the deterministic orchestrator and validator exposing canvas state, sticky notes, step schemas, live mutation guards and execution traces.

How the multiplayer MCP loop works

Simultaneous CRDT co-editing: Because the canvas uses CRDTs, several teammates and AI agents can work on the same automation at once: dropping sticky notes, tweaking models or prompts in place, and watching MCP edits land live.
End-to-end testing via any MCP: Pairing Glow MCP with any other MCP server in the same AI chat (such as Gmail, Slack, Linear or a database MCP) lets the agent fire realistic test events, inspect step-by-step execution traces and verify outputs automatically.

Watch the voice & MCP workflows in action

Two unscripted recordings starting from a sticky note on the shared canvas: one building and testing an inbox triage automation, and one building a web-researched lead agent and refactoring it into a reusable template.

Sticky Note & Self-Testing · 04:48

Claude Voice + MCP: Sticky Note to Self-Testing Automation in Glow

Moments

Loads YouTube only when you press play. Privacy

Two live canvas sessions side by side, starting from a sticky note with Gmail labeling requirements
Claude builds the automation through Glow MCP with schema validation and layout, then sends real test emails via Gmail MCP
A second sticky note triggers a live voice redesign that branches urgent emails to Slack

Lead Research & Template · 05:58

Claude Voice + MCP: AI Lead Research Agent to Reusable Template

Moments

Loads YouTube only when you press play. Privacy

Claude reads the sticky note over Glow MCP and wires a Gmail trigger, a native Glow AI Agent with live web search, a Gmail reply draft and a Slack DM
End-to-end test via Gmail MCP using a realistic Linear inquiry, verifying the web-researched Gmail draft and Slack high-fit summary on the first run
Voice refactoring into a reusable template: switching from Live to Draft after the live-mutation guard and extracting hardcoded values into a Custom Variables step

Unscripted product demos, built and tested live through MCP.

All 4 walkthroughs →Idea board (sebbuild.com) ↗YouTube channel ↗

Documentation I write

I write Glow's product and technical documentation. These pages explain how collaboration, versioning, human review and client access work, including the defaults and limitations people need to know.

Edit together on one canvas

Teammates can edit an automation together, with named cursors, presence indicators and synchronized changes. Edits to separate fields combine; when two people change the same field, the later value wins.

Real-time editing →

Save a version before you change it

Save named automation snapshots and restore an earlier state when needed. Draft disables triggers but allows manual runs. Live enables triggers, and edits apply directly to the current automation. Draft is not a separate staging copy, and most edits do not create a saved version automatically.

Versioning →

Choose where a person reviews the result

Combine deterministic logic and AI, then add Human Review where a person should decide what happens next. Review outcomes and expiry are configurable. Approval is not automatic, and a review without an expiry can wait indefinitely.

Human Review →

Give partners scoped access

Delegation grants define access tier, automation scope and expiry. Existing client workspaces require admin approval. Delegated operators cannot open the credential vault or manage connections, but authorized access can include automation results and business data.

Governance & impersonation →

Copy automations, reconnect client accounts

Review a non-writing preflight before creating an inactive copy in a client's workspace. Connections bind to matching client accounts or are cleared for setup. Source variable and secret values are not transferred, and captured samples are stripped. Preflight checks configuration, not whether a run will succeed.

Pushing automations →

Product context: teams & agencies

The problem

For a team: A shared automation needs more than a working sequence of steps. Teammates need to edit it together, review changes and choose where a person must approve an action.
For an agency: An agency needs to maintain automations across client workspaces while keeping permissions and app connections specific to each client.

The product requirements

Working together: Put integrations, deterministic steps and AI on one editable canvas, with saved versions and configurable human review.
Working across clients: Make delegated access explicit. When an automation is copied to a client workspace, reconnect it to the client's own accounts rather than transfer the agency's credentials.

About Glow: team, investors & recent releases

Seven members of the original doFlo team sitting together on wooden steps. — The original team during the doFlo days, before the rebrand to Glow.
Photo: Glow ↗

Investors in Glow

The investors listed on Glow's About page.

Glow's investor listing →

Let's connect

Let's talk about what you are building.

I'm not taking freelance work. Happy to talk about Glow, automation ideas, or a product engineering opportunity.

LinkedIn GitHub YouTube X Resume (PDF)