Works with and more
An audio-visual interface to your agents.
- Talk to your agents.
- Sketch or write what you want, like you would for a colleague.
- Hear and see what they did and what they understood.
- Get work done across your apps without leaving the screen.
Bysicle isn't out yet. Leave your email and we'll write when it's ready.
Your mind doesn't have folders, tabs and apps for your thoughts. On a board, your questions, the answers and the agents' work sit side by side, and you can point at any of it.
Tap a question. The answer lands on this board and the view moves to it. You can also type your own in the bar below.
The basics
Your code and data
Day to day
Show me
About us
One agent run, shown three ways. The run is an invented example: an agent was asked to make failed uploads retry.
In a terminal
$ agent "make failed uploads retry" ● Reading src/upload/client.ts (212 lines) ● Reading src/upload/queue.ts (148 lines) ● Searching for "retry" in src/ · 9 matches ● Thinking The client gives up after the first network error. queue.ts has a backoff helper that nothing calls. ● Editing src/upload/client.ts @@ -88,7 +88,12 @@ send(chunk) - const res = await fetch(url, opts) + const res = await withBackoff( + () => fetch(url, opts), { + tries: 3, baseMs: 2000, + }) ● Editing src/upload/queue.ts ● Running npm test -- upload PASS upload/client.test.ts (14) PASS upload/queue.test.ts (9) FAIL upload/resume.test.ts ✕ resumes after reconnect expected chunk 4, got chunk 1 ● Reading src/upload/resume.ts (96 lines) ● Editing src/upload/resume.ts - offset = 0 + offset = state.lastAcked + 1 ● Running npm test -- upload Tests: 24 passed, 24 total ✓ Done. 3 files changed, +41 −6.
To find the line that mattersscroll 140 lines
A chat app (like the Claude or Codex apps)
I'll start by looking at how uploads are sent and where errors are handled.
The client gives up after the first network error. There is a backoff helper in queue.ts that nothing calls, so I wrapped the fetch in it:
- const res = await fetch(url, opts)
+ const res = await withBackoff(
+ () => fetch(url, opts), {
+ tries: 3, baseMs: 2000,
+ })
I also exported the helper from queue.ts so the client can import it. Next I'll run the upload tests.
Most tests passed, but resume.test.ts failed: a resumed upload restarted at chunk 1 instead of chunk 4. I changed the offset in resume.ts to continue from the last acknowledged chunk and ran the tests again.
Yes. All 24 tests pass. 3 files changed, +41 −6. Let me know if you'd like different retry timing or jitter between attempts.
To find the line that mattersscroll 3 replies, open the tool calls
On a board
Failed uploads now retry 3 times, waiting 2, 4 and 8 seconds. A resumed upload keeps its place.
- const res = await fetch(url, opts)+ const res = await withBackoff(+ () => fetch(url, opts), {+ tries: 3, baseMs: 2000,+ })client.ts, queue.ts, resume.ts · +41 −6
24 of 24 pass. One failed on the way: a resumed upload restarted at chunk 1. Fixed in resume.ts.
To find the line that mattersit's highlighted
You see the summary, the change and the result, and you can point at any part and ask.
- You sketch the flow: two boxes and an arrow.
- You scribble a note on the part to change.
- Bysicle says what it understood and asks before handing it off.
- The plan lands beside your sketch, which is marked understood.
- You highlight a few words in the middle of a paragraph.
- You ask about them out loud.
- That paragraph changes in place. The rest of the island stays put.
- You give an agent a brief: add a dark mode to this board.
- The agent works through it on your Mac while you keep talking.
- The change lands in place on the island it came from, with the before beside it.
- This page switches to dark mode for a few seconds, then back to your own setting.
- An agent's pull request lands on the board as an island.
- You point at one hunk.
- You ask why.
- The answer lands in place, under that hunk.
- You ask why a test failed.
- An agent investigates on your Mac.
- The failing check, the cause and the fix land as one island.
- Before and after, and the tests pass again.
- You ask about earlier work.
- Bysicle finds the note in the board's knowledge base.
- The voice answers from it and says where it came from.
- You ask out loud.
- Bysicle answers in a sentence, and the options land on the board.
- You point at one row and say "change that part".
- That row changes in place. The rest stays put.
A Wider Channel
Each era of computing widened the channel between people and computers, from punched cards to a screen you can point at.
Now agents work faster than we can follow, so understanding what they did is the bottleneck, and Bysicle widens the channel again on the devices you already own.
What happens between the moment you stop talking and the moment the board changes.
- 1
Listen
You talk. Speech recognition turns your words into text and notices when you have finished.
- 2
Decide
A small, fast model reads what you said and what is on the board, and picks one path:
- Answer by voice
- Put an island on the board
- Change an island in place
- Hand the work to an agent
It only answers short questions, the same ones every time: is this a question or a request, does it name a part of an island, how long should the answer be. That keeps the wait before the voice starts short.
- 3
Speak
The voice answers in a sentence or two and asks where you want to go next.
- 4
Show
When the answer is long, it lands on the board as an island: a table, a diagram, code or text.
- 5
Move
As the voice names a part of the island, the view moves to that part and highlights it.
A project folder per board
Each board has a project folder. Work for that board happens in that folder.
You pick the agent
You choose which agent on your Mac does the work for a board. It works with Claude Code, plus Codex, OpenCode and Pi.
It reports back
The agent works in the background while you keep talking. When it finishes, it reports back to the same board, and you can ask it what it did and why.
Your terminal agent, on a board
Bysicle does not replace your coding agent. It hands work to the agents already on your Mac: Claude Code, plus Codex, OpenCode and Pi. It puts what they did on a board you can talk to and point at. Several agents can report back to one board.
When it gets it wrong
You can ask the agent what it did and why, tell it what to fix, or ask Bysicle to stop the job.
Connected apps
Bysicle connects to the apps your agents already reach through MCP, the common protocol for connecting tools to agents. Settings list the connected apps for each agent.
On your Mac
Speech can run on your Mac. You choose on-device speech or our cloud in settings.
Our cloud
If you choose the cloud, your audio and text go to our servers to be transcribed and answered.
Your provider
Answers can use a subscription you already have, an API key, or an endpoint you run yourself. Requests go to the one you choose.
Training on your data
A data policy, including whether anything is used for training, has not been announced. Join the waitlist to hear first.
This page
This board is a scripted demo with no live voice yet. Questions you type are matched in your browser and are not sent anywhere. Your email is sent only if you join the waitlist.
Who is it for?
It is built first for people who write software and work with coding agents. It supports Claude Code, plus Codex, OpenCode and Pi. The plan is to widen from there.
Why voice?
You can talk while you look at the board and while agents work, with no window to switch to. The voice keeps it short and the board holds the detail.
What is an agent?
A program that does real work on its own for a while. It can read a codebase, fix a bug, search the web, draft a plan or file a ticket.
How fast is it?
Running speech on the Mac's own hardware cut the time to first speech from about 700 milliseconds to about 185. Low latency is the first rule every design choice is judged against.
Does it remember my project?
Yes. Each project keeps a knowledge base: notes about the project, its files, and the work agents did for it. When you ask about earlier work, the voice looks it up there and says where the answer came from.
Is it a plugin?
No. Bysicle is a native Mac app. Being native lets speech run on the Mac's own hardware, which keeps replies quick.
Which Macs does it support?
A Mac with Apple Silicon running macOS 15 or later. An M1 with 8 GB of RAM runs it. A phone app is next.
Do I have to talk out loud?
No. You can type any turn instead of saying it. A typed turn gets the same answers, the same board and the same agents as a spoken one.
Is it open source?
That has not been announced. Join the waitlist to hear first.
When is it out?
A release date has not been announced. Join the waitlist to hear first.
What does it cost?
We're still working out pricing.
Who is building it?
The team isn't named on this page yet. Hiring and funding have their own answers: ·
Engineering
Bysicle is built for engineering first. An agent can build a feature and open a pull request. Before you merge it, you need to know how it fits into the rest of the system and how you checked that it does what you asked. When something breaks, you need enough understanding to find the cause.
Marketing
AI can make a product animation. A person still has to judge whether anyone will want to watch it. That takes a feel for pacing, music and what holds a viewer's attention.
Sales
AI can send outreach to thousands of people. A large enterprise client needs a personal approach that shows the problem your product solves for them. To automate outreach well, you first need to know what great outreach looks like.
The same need
The same need for understanding shows up in each field. AI can help you build that understanding too, and how you get it matters less than having it. Bysicle is one way to get it fast. Talk the work through, see it laid out, and point at the part you don't follow.
What will it cost?
We're still working out pricing. Join the waitlist and you'll be the first to know.
Are you hiring?
We're not hiring at the moment. Follow us on X and LinkedIn to hear when that changes.
One turn, from your voice to the board, and the model behind each step.
Speech recognition (ASR)
It depends on where you are: an Indic model or a Western one. It keeps changing as better ones come out.
Decisions: Jev
Our own decision model. On each turn it picks what happens: answer by voice, put an island on the board, change one in place, or hand the work to an agent.
Language model: Qwen
It writes the answers, unless you choose your own provider.
Speech output (TTS): Qwen
It speaks the reply.
Around them
A native Mac app. The coding agents on your Mac: Claude Code, plus Codex, OpenCode and Pi. MCP to connect the apps your agents already use.
In the app
Right now, nothing. When we add tracking, it will be usage only, such as how much you use it in a day.
Are you funded?
It'll be announced soon. Follow our socials.
Loved by people who've worked at
- Amazon
- Cisco
- Apple
- Neuralink
- HighLevel
“Useful for quick questions to Claude.”
“A perfect brainstorming partner.”
“It's removed the need for switching between apps, in the most intuitive way I've seen.”
“I've even replaced Wispr. I compose all my blogs here.”
From early users.