a multi-agent harness, made for humans

a team lead
for your agents.

you talk to one agent, bise. it runs the others and only checks in for the big decisions.

free and open source. you only pay for the models, as you do today. which plans work

download bise desktop

bise desktop runs on a Mac. open bise.dev there, or send yourself the link.

2026.10.1 · 205 MB · Apple silicon · macOS 14 or later

signed and notarized by Apple · on Linux, or rather in your terminal? bise terminal ↓

this Mac may have an Intel chip: bise desktop needs Apple silicon. bise terminal runs on both.

bise desktop isn't out yet. bise terminal is, with the same agents and the same pages.

$ curl -fsSL bise.dev/install | shcopy

bise is coming down as bise-mac-arm64.dmg. then:

  1. 1open it from your Downloads.
  2. 2drag bise to Applications.
  3. 3open bise from Applications.macOS asks once if you want to open an app from the internet: Open. then say what's on your mind.

nothing came down? download it again

bise
get bise

desktop or terminal?

same agents and threads in both: use one, or both at once.

bise desktop recommended

a window on your Mac

every project in one window, an inbox for what needs you, the bims at work. hold fn to talk from any app. it updates itself.

download for Mac

2026.10.1 · Oct 12, 2026 · Apple silicon · macOS 14 or later

not out yet: bise terminal is.

or the beta, a few days early · 2026.10.3 · 206 MB

bise terminal

in your terminal, for keyboard people

one thread per repo, keys for everything. macOS and Linux, over SSH too. see it

$ curl -fsSL bise.dev/install | shcopy

paste it in your terminal, then run bise in a repo.

why

you didn't become an engineer to babysit robots.

you’ve got more agents than you can keep track of.

five, ten of them at once. you hop from tab to tab all day, and by the evening your brain is soup.

you’d love to hand off more. but letting go is scary.

nobody tells you how to set up agents that run on their own. and when you try, you worry they’ll go off and do the wrong thing, with a token bill to match.

and none of it is fun!

it feels like working the counter at the post office.

what if your team looked like a video game?

bise

so here’s how bise does it.

you talk to one agent, bise. it starts an agent for each job, and only bothers you when something actually needs you.

it’s a simple setup. one lead, one agent per task. the lead figures out who does what, and in which order.

you’re still the boss. you choose the work, and what you want to look at yourself.

and it doesn’t cost more. they’re the same agents you’d run by hand, and bise itself is free: you only pay for the models, with the plan or the key you already have.

:*bise · every project
bise desktop: storefront's thread. you asked about checkout errors, bise found the cause, checkout fixed it, and search asks you a question in the inbox. the agents are on the right.
bise's own thread, across every project: what happened while you were out, and two things that wait for you

one agent to talk to.

bise sees all your projects. you say what’s on your mind and it passes the work to the right one. back from lunch? it’ll tell you what happened.

storefront's office open in the right bar: perf and stock at their desks, checkout done, a question mark over search, which waits for you

it feels a bit like playing the Sims.

each agent is a bim, a little character in its project’s office. you can see who’s typing, who’s stuck and who’s waiting on you. click one and you’re in its thread.

the inbox: two questions wait for you, then what you answered today and yesterday

one inbox for every decision.

questions from every project land in one inbox, with one banner, once. you answer from wherever you are, and your answer stays next to the thread it came from.

artifacts

questions you can actually answer.

when ten agents are running, one asks « can i drop the old discount codes? ». which codes? you scroll back through a thread you haven't read in an hour to find out.

so when a decision needs you, the agent writes you a page. it shows what it did and the options side by side. you leave notes right on it, pick one, and send it all back in one go.

perf's before / after page open beside its thread: three notes on it, one on a number, one on its question, one on the table, sent together with one click
the fun parts

it talks back. it clicks around.

over a docs page on testing renewals, you held fn and asked how to test a renewal without waiting a month. bise's card in the corner of the screen answers: make a test clock and advance it 31 days. the page stays in view.

talk to it. it talks back.

hold fn in any app, or press ctrl+r twice in the terminal, and just talk. a little face listens, thinks, answers out loud, and blows you a kiss when it's done. macOS

storefront's thread: a line says bise is driving Chrome on storefront.dev/pricing, with take the wheel and stop. you asked if buy is visible on the mobile pricing page; bise used the browser 3 times and answers: buy shows on all three plans. your own tab never moved.

it can use your browser.

agents open pages in their own tab group, with your logins: they read, click, fill forms and take screenshots. never your tab. macOS · off until you set it up

under the hood

the hard parts, done right.

nothing falls through the cracks.

code mode for every app.

plug in any MCP server. agents write a little script to use your tools, one run instead of forty calls.

nothing gets lost.

quit it, restart it, crash it: bise picks up where it left off.

they work as a team.

they talk and give each other a heads-up. the messaging is proven correct, so nothing goes missing.

the lead cleans up.

when an agent falls over, the lead picks it back up. you only hear about it if it needs you.

automatic worktrees.

each agent gets its own, set up and tidied away for you.

they can use your computer.

they use your browser in their own tabs, with your logins. your tabs stay yours.

it works while you don't.

a check every morning, a two-hour test suite: the agent naps and wakes up when it’s done.

it asks before anything risky.

a small model reads every call. the risky ones wait for your ok.

the little things are nice.

talk, dictate, paste a screenshot, quote a diff. it all goes in one box.

your plan or your keys

it works with what you already pay for.

your ChatGPT plan, OpenRouter, a coding plan, or any API key.

Claude plans only work in Anthropic's apps: for Claude, use an API key. the plans bise can use

who it's for

for people with more ideas than hands.

solo founders, indie makers, anyone with a long list of ideas and nobody to build them with. that’s who i made it for.

made by a human

hi, i'm GabrielGabriel Vergnaud, smiling, by the water at sunset

i work at Mistral, and bise is the multi-agent harness i always wanted.

i wanted one team lead to talk to, a team that does the work, and me staying in flow instead of babysitting tabs. so i built it. most of bise was built with bise.

i think it's the best multi-agent harness out there today. try it, and tell me what you think: i read every issue.

Gabriel :*

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

questions

you'll ask.

is it free?
yes. bise is free and open source. you pay for the models the way you already do: your ChatGPT Plus or Pro plan, a coding plan, OpenRouter, or an API key.

will it cost me more?
no. they're the same agents you'd run by hand. the lead only starts one when a job needs it, and does the small stuff itself. /usage shows what you spent today and this week, agent by agent, and /pause stops everything in one go.

is it safe to let them run?
the models are good at this now, and the people who get the most out of agents let them work. so by default, bise doesn't stop to ask. if you'd like a check, shift+tab turns on auto: safe commands just run, risky ones wait for your yes in the inbox, inside a macOS sandbox.

will one thread hold up on a project like mine?
yes. the lead doesn't carry your code. each agent reads what it needs with a fresh head, and the lead keeps the plan and the reports. bise is built with bise: one thread since September 28, 934 agents, 4,588 commits. it's still the same thread.

what does bise actually do for me?
it's your team. you say what you want. the lead splits it into jobs, starts the agents, keeps them out of each other's way, answers their small questions, remembers what you did weeks ago, and cleans up after them. you only hear about what needs you. a small software factory, and you're the boss.

i already use Claude Code or Codex.
keep your models, your MCP servers and your skills: bise brings them over. what changes is how you work: one lead to talk to, instead of seven sessions to watch.

which models?
the ones you already pay for: Anthropic, OpenAI, Google, Mistral and more, through a key, OpenRouter or your plan. the plans bise can use.

what does it run on?
bise desktop: Macs with Apple silicon, macOS 14 or later. bise terminal: macOS and Linux. you need git.

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

get bise

desktop or terminal?

bise desktop recommended

a window on your Mac

download for Mac

Apple silicon · macOS 14 or later

not out yet: bise terminal is.

bise terminal

in your terminal, for keyboard people

$ curl -fsSL bise.dev/install | shcopy

paste it in your terminal, then run bise in a repo.