Name: bise
Author: Gabriel Vergnaud

how

one thread. a whole team.

main is your team lead. it splits your ideas into jobs and runs the other agents.

you → mainApple Pay at checkout please. also search misses tshirt 😭

:* mainalways listening · knows the whole repo

∿ walletthe Apple Pay sheet

∿ checkoutthe order flow

∿ docsthe merchant setup

∿ searchunrelated, so it runs too

1no sessions to find againpeek into any agent when you feel like itno tabs to juggle, no "wait, which session was that?". main knows everything going on in the repo, even what you did last month.
2main is always listeningthe team is busy. main never isask, add a thing, change your mind while the agents work. main answers right away.
3a team that runs itselfthey talk behind your back. in a good wayagents message each other before they collide. when one needs its own copy of the repo, it makes a worktree and cleans it up after. each one commits its own work.
4forget there are agentsno yaml was harmed.no roles, no graph, no config. bise has opinions, so you don't need any. you just dump your ideas in the thread.

kiss your backlog goodbye

the models grew up. you can stop babysitting

fourteen things you'll notice in the first hour

the team runs itself

hand it something huge

give main a big goal. it splits it into jobs, runs the agents, restarts the ones that stop half-way, and keeps going until it's done.

worktrees? don't think about it

agents share your folder. when one needs its own copy, it makes a worktree, works there and cleans it up after. no branch to name, no folder to delete, nothing to keep in your head.

your head stays yours

main is always there

main gives the heavy work to the agents, so it's never busy. ask it anything, add one more thing, change your mind: it answers right away, and no job stops.

zen mode while you type

start typing and everything else fades: the agents, the counts, the chatter. just you and your words. send it and it all comes back. what the agents said to each other stays folded, ctrl+o opens it.

see what they made

everything they made, in one list

pages, docs, sheets, slides, PRs: what agents make for you lands in /artifacts. search it, open it, flip through its versions. in a reply, it's a link that opens it.

checkout landed: ± 3 files +42 −18. a click opens its diff against main on the right, file by file, green and red. tab goes to the next file; enter opens the line in your editor.

the diff, right there

click ± 3 files or press ctrl+g: an agent's changes against main open beside the thread, file by file. ⏎ opens the line in your editor.

hands off the keyboard

you press ctrl+r twice and ask main to move the promo field out of checkout. the face listens, thinks, then talks: main says promo will move it to the cart. it ends with a kiss.

talk to it. it talks back

press ctrl+r twice and just talk to main. it answers out loud while the agents keep working, and blows you a kiss when it's done. /voice picks the model, the voice and the language.

you ask main if add to cart is visible on a mobile product page. an agent opens it in a background tab in its own group, reads it, takes a screenshot, and main answers: it shows for all three colors. your own tab never moved.

it can use your browser

agents open pages in their own tab group, in the background, with your logins: they read, click, fill forms and take screenshots. never your screen, never your tab. Mac apps too. off until /computer-use.

no more "yes, continue"

you're not the router

main answers the agents for you: the obvious questions, the way you would, and it tells you why. only the real decisions reach you.

agents sync on their own

every agent can talk to every other one. they ask, share and hand off quietly, folded out of your way.

every tool you have

all your MCPs. all your skills. always on

GitHub, Linear, Sentry, Slack, your docs, your database: connect as many MCP servers as you want, and never pick which ones to turn on. bise calls tools through code, so a hundred servers don't fill its context.

bring your Agent Plugins

skills, MCP servers, hooks: plugins in the Agent Plugins format load as they are, from ~/.agents/plugins or your repo. Vibe plugins too.

you're still the boss

talk to any agent, anytime

⌥ + a number, or @name. ask it why, push it, then go back to main. nobody has to stop.

fits your PR flow

in a repo that takes pull requests, every change gets its own branch and its own PR, then goes through your usual flow: CI, review bots, teammates. the panel shows where each one stands.

and so much more

polished down to the last character

hundreds of tiny details. you'll feel them before you see them.

you set the eval timeout to 120 s :* done, i did it myself. too small for an agent. 0 agents started

your token bill can relaxbise starts an agent when a job needs one. that's the whole rule. there's no committee of agents reviewing each other in circles.

╭ attached ─────────────────────╮ │ ▣ 1 terrain.png 1440×900 │ ╰───────────────────────────────╯ shadow acne at dusk ▣ 1

show it a screenshotpaste it with ctrl+v or drag it in. an image is one chip in your text.

● listening ▂▅▃▆▂ and test it with 2% packet loss ctrl+r to stop

or just talkturn on /voice, press ctrl+r, say it.

:* cache is done. hit rate went from 12% to 71% on prod ❝ 1 what made the difference?

ask about anything on screenselect a few lines in the history and start typing: they come along as a quote.

:* main opus·hi ∿ ingest opus·hi ∿ alerts haiku·lo ∿ release haiku·lo

a model per agentthe big one for the hard job, a fast one for the chores. /model and /reasoning.

◷ main scheduled #48 for ingest every 2m · next 14:22 :* ingest: lag 4 s. ◷ scheduled #48 ended

scheduled tasksagents wake themselves on a schedule: watch a build, ship every morning. /scheduled lists them, runs one now, stops it.

$ npm test ✓ 142 passed (3.1 s) $ ctrl+` hides it. it keeps running

a real shell, one key awayctrl+` opens a terminal in your repo. it keeps running while hidden.

you /restart bise v2026.10.3 → v2026.10.4 ∿ rag ∿ dedupe ∿ cache back to work at once

restart whenever you likeupdate to the latest version mid-work, or quit and come back tomorrow. no work is ever lost: your agents resume right away, and your thread, your draft and your queue come back too.

$ bend PROOF.bend ✓ no message lost ✓ none delivered twice ✓ a restart = the same state ✓ nothing waits forever ALL PROOFS CHECK

agent communication is proven correctwe have the proof, thanks to Bend, a language built for the LLM era. no message lost, none sent twice.

bise :* ∿ streaming ✓ shadows

your terminalyour colors, light or dark, text at a reading width.

give it a try

one line to install.
then just talk.

1install it
$ curl -fsSL bise.dev/install | shcopy
2start it in the repo you want to work on: cd your-repo && bise
3pick a theme, paste a key. then say what's on your mind.

runs on macOS, Apple silicon and Intel. linux soon. windows later.

you need git, and a way to pay for the models: your ChatGPT Plus or Pro plan, or an API key from Anthropic, OpenAI, Mistral and more. a Claude plan works only in Claude Code. the plans bise can use.

coming from Claude Code? . it installs bise and brings over your key, model, skills and MCP servers · see the full prompt ↗

later bise update to stay current, bise doctor if something feels off, bise uninstall to leave.

questions

but i already have Claude Code / Codex / Vibe
keep your models, MCP servers and skills. bise changes how you work with them: you talk to a team lead instead of juggling sessions.

won't the agents step on each other?
they tell each other what they touch before they touch it. worktrees only when one needs its own copy, cleaned up after.

does it ask before running a command?
by default, no: that's yolo, everything runs. shift+tab switches to auto: safe commands just run, risky ones ask you first in your inbox, and it all runs in a macOS sandbox.

will a single thread scale to a project as complex as mine?
yes. main doesn't carry your code: the agents do the reading, each one with a fresh context. main keeps the plan and the reports, so the thread stays light. when it gets long, it compacts and keeps going. tested on a repo with millions of lines of code.

won't my context overflow? i don't want to live in the dumb zone
no need to overthink it. when a thread gets long, bise compacts it. the latest models are trained for that and they keep their trajectory. your thread just keeps going.

do i need a memory system?
nope. compaction carries the thread, and any agent can go back and read the full history: every message, every tool call. nothing gets lost.

how many agents can i run?
as many as you want. the real limits are your tokens, your budget and your provider.

what does it run on?
macOS for now, Apple silicon and Intel. linux is next. windows later, maybe through WSL first. you need git, and that's about it.

which models?
bring your own key: Anthropic, OpenAI, Mistral and more. the first run lists every provider.

do i need to learn anything?
four words: you, main, agents, inbox. main is the one you talk to. the rest you pick up in the first hour.

open source?
yes, under Apache-2.0. read it, fork it, send a :*.

made by a human. shipped with bise

hi, i'm gabriel

i build coding agents for a living.

agents made me faster. they also sometimes made my days feel empty. i was switching tabs, answering "should i continue?", pasting context around. i shipped more than ever and yet, i didn't enjoy it.

so i built the tool i wanted: a team lead to talk to, and me in flow. most of bise was built with bise.

i want everyone to ship like a team of a hundred, and to love building again. bise is open source and indie. i read every issue.

Gabriel :*

— Gabriel Vergnaud · @gvergnaud on GitHub · @GabrielVergnaud on X

you didn't become an engineer
to babysit robots.

seven agents.
all of them need you.

one thread. a whole team.