Sightdeck AI — Agentic Studio for agent-driven development

Agent-driven development at scale

Stop prompting one agent and waiting.
Start running a studio.

The model was never the bottleneck. The work around it was: what to hand over, in which order, who checks it, and who notices when it's stuck. Sightdeck AI is the orchestration layer for exactly that — for one developer or a whole team.

a/one-prompt-one-agent

You type a prompt, then watch a terminal. You are the scheduler.
Every session starts cold. You re-explain context that already exists.
One task at a time. The other twelve rot in a TODO file.
"Done" means it compiled. Tests and review are your job, later.
An agent asks a question at 14:10. You notice at 17:40.

b/agentic-studio

Tasks are refined into executable work before an agent touches them.
Each task carries its concept and history. Agents pick the thread back up.
Parallel where safe, serial where files collide — across every project.
Tests, reviews and your Definition of Done are steps of the workflow.
Blockers turn red the moment they happen — on your Mac and your phone.

How the studio works

Four things you stop doing by hand.

From the first rough sentence to a tested, reviewed Done — without babysitting a single session.

"Add VAT to the invoice PDF and fix the broken thumbnail in checkout."

↓ refinement

ACME-21Invoice PDF: show VAT per line

kind: feature
points: 5 · half a day
model: opus
concept: tasks/ACME-21.md ✓

01 · FROM TASK TO EXECUTION

A rough idea in. A task an agent can actually run, out.

Write it down badly, or dictate a whole braindump. Sightdeck AI splits it into tasks and refines each one: kind, estimate, story points, the right model and a concept the agent reads before it starts. Bigger work becomes an Epic with planned subtasks — you approve the plan, then press Start.

Running Tasks board: several projects stacked, each with serial and parallel lanes of status-coloured task pills

02 · RUN AT SCALE

Ten tasks, five projects, one Start button.

Select a batch and Sightdeck AI proposes the run shape: side by side where the tasks don't touch the same files, one after another where they would. Each task gets a fresh session on its own branch. Up Next keeps feeding the queue while you sleep — this is what a team of one looks like.

ACME-21 · workflow

✓ImplementationJonas Weil · Developer

✓Dev tests — 41 of 41 greenPlaywright

●ReviewMr. Mint · Security Expert

○Target persona reviewDavid Köhler · Solo Dev

○Definition of Done · 3 itemsteam

03 · QUALITY BUILT IN

Done means tested and reviewed, not "it compiled".

Personas, workflows and a Definition of Done decide what finished means. Tests, security review, a look through your target user's eyes — each is a step with its own specialist. A failed review sends the task back, and a second agent on a different provider can cross-check the result before it closes.

ACME-5 Toggle left or right of the avatar? Answer

ACME-15 · step 4 of 7ETA 12–18 min

src/api/client.ts

To do In progress Testing Done

04 · STAY IN CONTROL

You orchestrate. You don't babysit.

Every task shows where it is — to do, in progress, testing, done — which step is running, which file is open and roughly how long it still needs. When an agent needs you, the task jumps into a pulsing red column and pings your phone. Nothing runs where you can't see it.

Workflows

Build a team of agents.
Draw how they work together.

Developer, security expert, UX designer, SEO, the target personas you build for — each is an agent with a role. Drag them onto the steps of a workflow, connect the steps, add an If where the path splits. Every task then walks the same way from Backlog to Done, with the right specialist at every step.

Workflow editor: agents dragged onto steps from Requirements Engineering to Implementation, Review and Target Persona Review, with If branches and a loop back on failed tests

Workflow editor · requirements → design → implementation → tests → review → persona review → docs → done

JWJonas Weil Developer MMMr. Mint Security Expert LBLena Brandt UX/UI Designer SNSara Niemann AI Expert FRFelix Roth SEO Expert DKDavid Köhler Target persona +your own any role

Agents with a role

Your project's own agents — name, role, instructions, avatar. Plain agent files in your repo, not a lock-in format.

Steps, branches, loops

Several agents on one step work it together. An If sends features through design; three failed test runs send the task back.

Stages across providers

Build on Claude, review on Codex. Each stage is its own session and hands over in writing to the next.

Playbook & templates

Write down who does what in your team's words, and start from a workflow template instead of a blank canvas.

Why "studio"

Delegate the work.
Stay in the middle of it.

Handing everything to agents and walking away is one way to work. Sitting next to them is another. The studio does both at the same time — pick per task.

Delegate

Let the queue work itself off.

Fill Up Next, press Start, go cook. Workflows run their steps, sessions that crash restart, and the few questions that really need you land on your phone.

SD-21→ SD-22→ SD-23→ SD-24

Dive in

Code right next to your agents.

Open the Code view and you are in the middle of it: your files, what an agent just changed, and a chat that knows which task wrote which line. Jump into a running session to steer it, or write the tricky part yourself while the rest keeps running.

SD-23 · you're in SD-22 SD-24

fully delegatedhands on the keyboard

Sightdeck AI · acme-shop BoardRunningCodeRoadmap

FilesChanged · 3

▾ src/invoice

pdf.ts M

vat.ts A

▸ src/checkout

Thumbnail.tsx M

▸ tests

▸ docs

package.json

src/invoice/pdf.ts last written by ACME-21 Reviewed · next file

40export function lineRow(item) {

41 return [

42 item.name,

43 money(item.total),

44 ];

45}

HEAD

40export function lineRow(item) {

41 return [

42 item.name,

43 money(item.net),

44 vatLabel(item.rate), // "19 %"

45 money(item.net * (1 + item.rate)),

46 ];

+3 −1 Ask why

ACME-21 · Terminal ✓ 41 of 41 specs green ● Review · Mr. Mint

AI Explain this change

Why net instead of total?

The PDF now prints each line as net, VAT rate and gross, as ACME-21 asked. The rate comes from vat.ts. pdf.ts · L43–45 vat.ts · L8

Your filesBrowse, open and edit, with syntax highlighting and go to definition.

Review changesWhat an agent changed, side by side, and which task wrote it.

Ask whyAI Explain answers with jumps straight to the lines it means.

Live terminalEvery session runs inside the window. Read along or type into it.

Remote Control

Your studio, in your pocket.

Walk away from the desk and keep orchestrating. Answer a blocked agent, resume a stalled session, move tasks or start new ones from your phone. At home over Wi-Fi with Remote, anywhere else with Anywhere.

Add device9:59 left

Waiting for a scan

http://127.0.0.1:8765/?token=…

Remote · Anywhere Beta

Answer your agents from the couch, the train or a client office.

Remote puts the board on your phone over the same Wi-Fi. Anywhere drops that condition and works on any network.

Pair once, with a code that expires on its own.
Phone and Mac connect directly, peer to peer.
If that route is blocked, an encrypted relay carries it and never sees a task, a file or a reply. You can forbid the relay entirely.
Answer, resume, move tasks and start new ones.

One backlog

Four ways to look at it.

Same XML underneath. Pick the lens the question needs — what's next, what's stuck, what ships this quarter, what it all cost.

List

the default

Everything in flight above the fold, backlog below, sortable by story points or by hand. Stalled sessions flag themselves.

Board

kanban

Drag a card between columns to change status — drop it on In Progress and the agent starts, same as pressing Start.

Roadmap

day · week · month · quarter

Lay tasks on a timeline, drag a bar to reschedule, and watch tagged versions travel from planned to released on the axis.

Insights

across all projects

Throughput, cycle time, velocity per week and release health — and a warning the moment a planned version's date slips past.

The whole surface

Everything in the app, grouped by what you're doing.

Six things you do with agents all day. ● marks what landed in 1.5.0.

Plan

10

Quick-add & auto-split — one sentence in, several refined tasks out.
Refinement — kind, estimate bucket, Fibonacci points and model, filled in for you.
Plan mode review — Claude drafts the epic and its subtasks; you approve, skip or delete before anything is created.
Epics & subtasks — a parent locks while its children run, and resumes an interrupted run cleanly.
Concept per task — a Markdown doc beside the ticket, so context is never re-explained.
Templates — reusable task shapes for the work you do over and over.
Project prompt — your own rules added to every task Claude starts, kept on your machine only.
Task dependencies — blocked-by and related links, a dependency graph, and missing links suggested by AI.
AI Learning — estimates and remaining time learned from your project's own finished tasks.
Store — ready-made workflows, agents, checklists, skills and task templates, with ratings.

Run

10

One-click launch — a fresh Claude session in the right directory with the right prompt.
Bulk start — select as many as you want, parallel or serial, shape proposed from file overlap.
Up Next & auto-run — a queue that works itself off while you sleep, restarting stalled sessions.
Model per task — haiku for a typo, opus for a migration. Assigned from complexity.
Resume & restart — continue the old session, or start the task fresh.
Git branches & worktrees — each task on its own branch, optionally its own worktree, with auto-commit checkpoints.
Commands & before/after hooks — plain scripts around a task, no tokens.
Providers — Claude Code or Codex, a default per project, each with its own models and settings.
Stages & workflows — every task through stages in order, each stage on its own provider and model.
Start button — your project's start, stop and health commands, with live output.

Watch

9

Running Tasks — every lane across every project in one frame.
Needs-you lane — a blocked agent goes red and pulses until answered.
Live progress — step plan, current file and ETA range per run.
Status lifecycle — backlog, active, testing, waiting, done, written by the skill itself.
Duration tracking — real wall-clock time against your estimate.
Cost pill — tokens and dollars on the card, model breakdown one click away.
Agents view — every session inside the window, with a terminal dock; sessions survive a restart.
Profiler — every tool call of a run on one timeline: what held it up and where the money went.
Usage limits — your 5-hour and weekly Claude usage, with a warning before it runs out.

Organise

9

Multi-project sidebar — unlimited projects, grouped, with one combined list.
Board & Roadmap — kanban and a week/month/quarter timeline on the same tasks.
Calendar — what ran when, per day.
Drag-to-prioritise — top of the backlog is next to run; the queue order is the order.
Tags & versions — tag tasks, maintain releases and their dates, track planned to shipped.
Archive & import/export — retire finished work, move a backlog between machines.
Card configuration & maintenance — pick the fields you want on cards; find unused files and broken references.
Code view & review — browse and edit files, review every change side by side, ask why it was made.
Instructions — edit CLAUDE.md and AGENTS.md with preview and review, or fill them from templates.

Report

6

Journals — day, week or month of finished work as a report, PDF included.
Changelog builder — release notes from closed tickets, with your own rules and examples.
Insights — throughput, daily stats, release health.
ROI per ticket — what each piece of work actually cost.
Task history — every status change, timestamped, on the ticket.
Definition of Done — checklists per agent and per team; every finished task shows what was met.

Connect

8

Sightdeck Anywhere — paired phone control off your own network, encrypted, relay sees nothing.
Sightdeck Remote — the same board on your phone over local Wi-Fi: answer, resume, move, start.
Jira integration — paste a key or URL and the ticket arrives with criteria, comments, attachments, priority and links as Markdown.
Jira import wizard — a whole batch at once, filtered by assignee, with live sync status per ticket.
Webhooks — a POST on every task that hits done. Slack, deploy, status board.
Team Settings — per-member id ranges and prefixes so the XML merges clean in git.
Ask & Help chats — ask about your project and get answers from its code; Help answers from the docs.
GitHub import — bring issues in from GitHub as tasks.

Watch it run

See the studio in action.

Short walkthroughs, no signup. New videos land on the channel as features ship.

Sightdeck AI - Quick Start

Set it up and start your first tasks in a few minutes.

More on the channel. Hello Code on YouTube →

Who runs it

A team of one. Or a team that ships like ten.

Solo developer

You are product, dev, QA and reviewer. The studio gives each of those a seat — you just decide.

Agency & freelance

One backlog and one workflow per client. See at a glance which client has an agent waiting.

Tech lead

The team's rules become the workflow. Every agent runs through the same review, whoever started it.

Project manager

Tasks, epics, roadmap and a live view of what every agent is doing — without opening a terminal.

From the field

What builders say.

"Sightdeck AI helped me optimize development with Claude Code even more. From one terminal to another, waiting so you don't lose track, to multitasked coding, testing and deploying. Great tool if you want to up your AI development game!"

Johan Moorman CEO, Memo · memo-ict.nl

"Sightdeck AI helps me managing multiple Claude Code sessions in parallel with much more structure and clarity. With a central overview of backlogs, active tasks and open questions, it has become an essential part of my AI development workflow."

Daniel Naumann CTO, Ottscho IT GmbH · ottscho-it-service.de

One license, the whole studio

€49 / year

Sightdeck AI for macOS 12 or newer, Apple Silicon and Intel
Unlimited projects, tasks, agents and lanes
Workflows, stages and cross-checks included
Sightdeck Remote and Anywhere included
Every update released while the license is active
Email support

Checkout via Paddle · receipts include VAT for EU customers · we don't store payment details

Download · v1.5.0

macOS · Apple SiliconM1 · M2 · M3 · M4 and newer .dmg → macOS · IntelMacBook Pro · iMac · Mac mini .dmg →

Single binary, no installer drama, no account. Updates land in-app. View changelog

Good to know

Common questions.

Is Sightdeck AI a coding agent?

No. It orchestrates the agents you already use — Claude Code, and in beta other agent CLIs such as Codex. Sightdeck AI decides what runs, in which order, through which steps, and shows you all of it.

What is a workflow?

The steps a task goes through from Backlog to Done, drawn in an editor: which agents work each step, where the path branches, when a task goes back. Workflows and agents are plain files in your repo, so your team shares them through git.

Can I still code myself?

Yes. The Code view shows your files, what an agent changed and which task wrote it, and every task runs on its own branch. Resume a session to steer it, or work in your own editor while the other tasks keep running.

Do I need Claude Code installed?

Yes. Sightdeck AI is a layer on top of Anthropic's Claude Code CLI. If you don't have it yet, the app's Setup screen will install it for you in one click.

Where is my data stored?

Locally, in plain XML and Markdown inside your project (.agents/sightdeck/). Sightdeck AI never uploads anything. Commit the files to git if you want them shared; ignore them if you don't.

Does it work on Windows or Linux?

The desktop app ships for macOS 12 (Monterey) or newer, on both Apple Silicon and Intel. Windows support is in progress. The skill and CLI scripts are pure Python and run anywhere Claude Code does.

What happens when I close the app?

Nothing. The XML keeps living in your project. Claude keeps writing to it. Open Sightdeck AI next week and everything is exactly where you left it — only with whatever the agents did in the meantime added on.

Can I use Sightdeck AI without the dashboard?

Yes. The skill is the contract. Install just the skill into ~/.claude/skills/sightdeck-ai, and Claude will manage tasks in the XML on its own. The dashboard makes it pretty — it isn't required.

Can I integrate with my own tools?

Yes — Sightdeck AI fires a webhook POST every time a task moves to done. Point it at a Slack bot, an e-mail relay, a deploy script or anything else with a URL. The payload carries the task id, title, status, run lane and timing — wire it up in minutes, no plugin system needed.

Does Sightdeck Anywhere send my code somewhere?

No. The relay only helps your phone and your Mac find each other; the connection between them is encrypted end to end, and tasks, files and replies never pass through it in readable form.

How does the license work?

The license is a yearly subscription. All updates released during the active period are included.

Build more. Orchestrate everything. Stay in control.

Stop prompting one agent and waiting.
Start running a studio.

Four things you stop doing by hand.

A rough idea in. A task an agent can actually run, out.

Ten tasks, five projects, one Start button.