multiplayer AI for your team’s coding agents

every AI agent on your team knows what the others are changing, before it edits

download for mac download for windows

intel mac? this one

agentmash

we’re live on product hunt, an upvote would mean a lot

what agentmash does for your team

agents warn each other

a heads-up before the edit, not after the merge

without agentmash: 55%, with agentmash: 98.6%

+44 pp overlaps caught before the conflicting write

overlaps caught before the conflicting write+44 pp: without agentmash 55.0%, with agentmash 98.6%
overlaps settled without a human arbiter+59 pp: without agentmash 40.0%, with agentmash 98.5%
API breaks caught before integration+17 pp: without agentmash 82.0%, with agentmash 98.5%
conflicts closed without manual edits+33 pp: without agentmash 65.0%, with agentmash 98.4%
ambiguous conflicts escalated to a human+14 pp: without agentmash 85.0%, with agentmash 99.4%

without agentmash: 2.70 h, with agentmash: 1.70 h

37% less human hours per accepted task at 75% overlap

human hours per accepted task, by how much of the work overlaps

0% of work overlaps: 2.25 h without agentmash, 1.75 h with it, −22%
25% of work overlaps: 2.40 h without agentmash, 1.63 h with it, −32%
50% of work overlaps: 2.55 h without agentmash, 1.61 h with it, −37%
75% of work overlaps: 2.70 h without agentmash, 1.70 h with it, −37%

same people, models, budgets and tasks, without agentmash and with it

Csv export button Working Done just now

You 19:03add a csv export button to the contacts list
read src/contacts/list.tsx
edit src/contacts/list.tsx+3 -0
agentmash told it, before the edit[AgentMash: PARALLEL EDIT] damir has uncommitted changes to src/contacts/list.tsx on damir/csv-import, last written 2 min ago, while working on 'make the csv import keep empty rows'. They are only on damir's machine, so nothing you can fetch shows them yet. Keep this edit additive, and say in your reply that damir has work in progress in this file, so the two are seen before either lands.
Thinking
write src/contacts/export-button.tsx+24 -0
damir has work in progress in src/contacts/list.tsx, so I kept this additive: the button is a file of its own, and the list only gains its import and one line. nothing of his is renamed or moved.
Done

Chat agentmash

Bbob 09:12Morning. CI is green again after the Node 24 bump.
DIDamir I. @damir 10:05The demo is at 15:00: nobody edits investor-demo.md after 14:30.
Your Claude Code and damir’s Gemini CLI both changed src/contacts/list.tsx. Your Claude Code was told before its edit.19:04

Message the team. @ names a personThe last few messages reach every agent here before its next prompt.

ask a teammate’s agent

send a prompt, nothing runs until they approve

without agentmash: 17.5%, with agentmash: 4.7%

73% less tasks that end in an integration conflict; conservative lower bound: 8.3% of tasks, 53% less

textual conflicts, per 100 tasks−82%: without agentmash 13.0, with agentmash 2.3
manual fixes when two tasks edit one block, per 100 edit pairs−92%: without agentmash 65.0, with agentmash 5.1
semantic conflicts with no textual conflict, per 100 integrations−63%: without agentmash 3.8, with agentmash 1.4
tasks that hit the same conflict again within 7 days−6.3 pp: without agentmash 8.0%, with agentmash 1.7%

without agentmash: 540, with agentmash: 123 min

77% less human minutes of conflict resolution per 100 tasks

time to merge, median−52%: without agentmash 3.1 h, with agentmash 1.5 h; median time from the first ready diff to acceptance
time to merge, 95 of 100 merged by−42%: without agentmash 14.0 h, with agentmash 8.1 h; the time from the first ready diff within which 95 of 100 tasks are accepted
CI minutes rerun per integration−63%: without agentmash 22.0, with agentmash 8.2; CI minutes per integration spent on reruns after a conflict or a stale base
post-merge rollbacks, per 100 tasks−72%: without agentmash 0.90, with agentmash 0.25; rollbacks after a merge caused by two changes interacting, per 100 accepted tasks
lockfile conflict effort−82%: without agentmash 18.0 min, with agentmash 3.2 min; human minutes to recover from a lockfile conflict

same people, models, budgets and tasks, without agentmash and with it

your lobby

Prompt a teammate’s agent

Teammate

Bbob (@bob) · 1 liveDIDamir I. (@damir) · 2 live

Which of their agents

Any session their lobby picks one, or opens oneClaude Code “review the pricing page copy”Terminal Idle main

How it landsDraftQueueWhat it may doRead onlyCan editWaits for1 hour

What should their agent doE.g. can you look at why the login test is flaky on your branch? No changes, just tell me what you find.can you look at why the csv import drops empty rows on your branch? read only, just tell me what you find.0 / 32,000106 / 32,000

CancelSend for approval

DI

Damir I. @damir approves this first. Nothing runs until they do, and they can narrow it.

It runs on their machine, on their account. They answer every tool request themselves, not you.

damir’s lobby

AAlim @alim asks in agentmash:Expires in 59 min

can you look at why the csv import drops empty rows on your branch? read only, just tell me what you find.

Goes to “Review the pricing page copy”. Read only.

SendDeclineAsk me first

Review the pricing page copy Idle Working Done just now

From Alim Read only Draft you sent19:02can you look at why the csv import drops empty rows on your branch? read only, just tell me what you find.
Thinking
read src/contacts/import.ts
read test/import.test.ts
parseCsv skips a row when every cell in it is empty: the check on line 42 of src/contacts/import.ts. nothing changed, as you asked.
Done

every agent, one app

claude code, codex, cursor and more on your own logins

without agentmash: $182, with agentmash: $123

33% less total cost per accepted task, people’s time, API and CI

2.10M → 1.41M tokens per accepted task, 33% less

Gemini 3.8 flash, cost per task $91 without agentmash and $67 with it, multi-agent time 2.0 h without agentmash and 1.4 h with it, solo human hours 1.10 h without agentmash and 0.79 h with it, acceptance rate 92.8% without agentmash and 93.3% with it
GPT-6 luna, cost per task $100 without agentmash and $73 with it, multi-agent time 2.6 h without agentmash and 1.8 h with it, solo human hours 1.20 h without agentmash and 0.89 h with it, acceptance rate 92.8% without agentmash and 93.3% with it
Claude Haiku 5.5, cost per task $108 without agentmash and $78 with it, multi-agent time 2.4 h without agentmash and 1.6 h with it, solo human hours 1.30 h without agentmash and 0.95 h with it, acceptance rate 92.8% without agentmash and 93.2% with it
Claude Sonnet 5.5, cost per task $166 without agentmash and $121 with it, multi-agent time 4.8 h without agentmash and 3.0 h with it, solo human hours 1.90 h without agentmash and 1.37 h with it, acceptance rate 95.2% without agentmash and 95.8% with it
GPT-6.1 sol, cost per task $174 without agentmash and $127 with it, multi-agent time 5.0 h without agentmash and 3.2 h with it, solo human hours 2.00 h without agentmash and 1.46 h with it, acceptance rate 95.2% without agentmash and 95.8% with it
Gemini 4 argon, cost per task $230 without agentmash and $168 with it, multi-agent time 6.6 h without agentmash and 4.1 h with it, solo human hours 2.60 h without agentmash and 1.92 h with it, acceptance rate 96.6% without agentmash and 97.4% with it
Claude Opus 5.5, cost per task $261 without agentmash and $193 with it, multi-agent time 8.0 h without agentmash and 5.1 h with it, solo human hours 2.80 h without agentmash and 2.04 h with it, acceptance rate 96.6% without agentmash and 97.3% with it
GPT-6 astra, cost per task $281 without agentmash and $202 with it, multi-agent time 7.4 h without agentmash and 4.6 h with it, solo human hours 3.10 h without agentmash and 2.23 h with it, acceptance rate 96.6% without agentmash and 97.4% with it

GPT-6.1 sol without agentmash: 5.0 h, GPT-6.1 sol with agentmash: 3.2 h

up to 38% faster median hours to an accepted task, several agents (GPT-6.1 sol)

Gemini 3.8 flash, cost per task $91 without agentmash and $67 with it, multi-agent time 2.0 h without agentmash and 1.4 h with it, solo human hours 1.10 h without agentmash and 0.79 h with it, acceptance rate 92.8% without agentmash and 93.3% with it
GPT-6 luna, cost per task $100 without agentmash and $73 with it, multi-agent time 2.6 h without agentmash and 1.8 h with it, solo human hours 1.20 h without agentmash and 0.89 h with it, acceptance rate 92.8% without agentmash and 93.3% with it
Claude Haiku 5.5, cost per task $108 without agentmash and $78 with it, multi-agent time 2.4 h without agentmash and 1.6 h with it, solo human hours 1.30 h without agentmash and 0.95 h with it, acceptance rate 92.8% without agentmash and 93.2% with it
Claude Sonnet 5.5, cost per task $166 without agentmash and $121 with it, multi-agent time 4.8 h without agentmash and 3.0 h with it, solo human hours 1.90 h without agentmash and 1.37 h with it, acceptance rate 95.2% without agentmash and 95.8% with it
GPT-6.1 sol, cost per task $174 without agentmash and $127 with it, multi-agent time 5.0 h without agentmash and 3.2 h with it, solo human hours 2.00 h without agentmash and 1.46 h with it, acceptance rate 95.2% without agentmash and 95.8% with it
Gemini 4 argon, cost per task $230 without agentmash and $168 with it, multi-agent time 6.6 h without agentmash and 4.1 h with it, solo human hours 2.60 h without agentmash and 1.92 h with it, acceptance rate 96.6% without agentmash and 97.4% with it
Claude Opus 5.5, cost per task $261 without agentmash and $193 with it, multi-agent time 8.0 h without agentmash and 5.1 h with it, solo human hours 2.80 h without agentmash and 2.04 h with it, acceptance rate 96.6% without agentmash and 97.3% with it
GPT-6 astra, cost per task $281 without agentmash and $202 with it, multi-agent time 7.4 h without agentmash and 4.6 h with it, solo human hours 3.10 h without agentmash and 2.23 h with it, acceptance rate 96.6% without agentmash and 97.4% with it

Claude Opus 5.5 alone: $20.00, Opus 5.5 + Sonnet 5.5 with agentmash: $9.95

50% less than the stronger model alone API cost per accepted task, Opus 5.5 + Sonnet 5.5

GPT-6 astra + GPT-6.1 sol: $15.50 without agentmash, $19.00 for the stronger model alone, $9.73 with agentmash
GPT-6.1 sol + GPT-6 luna: $7.80 without agentmash, $9.40 for the stronger model alone, $4.48 with agentmash
Claude Opus 5.5 + Sonnet 5.5: $16.00 without agentmash, $20.00 for the stronger model alone, $9.95 with agentmash
Claude Sonnet 5.5 + Haiku 5.5: $7.60 without agentmash, $9.40 for the stronger model alone, $4.32 with agentmash
Gemini 4 argon + 3.8 flash: $10.80 without agentmash, $13.00 for the stronger model alone, $6.22 with agentmash
handoffs continued correctly: 92% without, 98.5% with

same people, models, budgets and tasks, without agentmash and with it

Agents Check again

Coding agents you can run here, each on its own account.

Your agentsAuto rules

Running 2Claude Codeby Anthropic Claude Max planStart
Needs sign-in Ready Running 1Cursorby Cursor Your Cursor accountSign inSign inStart

More agents

Not installedGitHub Copilotby GitHubInstall
Not installedopencodeby opencodeInstall
Not installedCodexby OpenAIInstall

your team, live

who is online and what every agent is doing

without agentmash: 12.0, with agentmash: 17.2

43% more accepted tasks per developer per 40-hour week

accepted tasks per 100 human hours, by team size

2 developers: 34.1 without agentmash, 48.8 with it, +43%
3 developers: 33.0 without agentmash, 48.8 with it, +48%
4 developers: 31.9 without agentmash, 48.8 with it, +53%
8 developers: 30.0 without agentmash, 46.2 with it, +54%
16 developers: 27.3 without agentmash, 40.4 with it, +48%

without agentmash: 2.40 h, with agentmash: 1.50 h

37% less active human hours per accepted task

human hours per accepted task, by team size

2 developers: 2.11 h without agentmash, 1.43 h with it
3 developers: 2.18 h without agentmash, 1.37 h with it
4 developers: 2.26 h without agentmash, 1.31 h with it
8 developers: 2.40 h without agentmash, 1.37 h with it
16 developers: 2.64 h without agentmash, 1.66 h with it

without agentmash: 7.6 h, with agentmash: 4.4 h

42% faster median hours to an accepted task; 95 of 100 tasks accepted within 31 h without, 21 h with agentmash

median hours to an accepted task, by team size

2 developers: 6.8 h without agentmash, 4.3 h with it
3 developers: 7.0 h without agentmash, 4.1 h with it
4 developers: 7.2 h without agentmash, 3.8 h with it
8 developers: 7.6 h without agentmash, 4.0 h with it
16 developers: 9.1 h without agentmash, 5.3 h with it

same people, models, budgets and tasks, without agentmash and with it

Team agentmash

Add teammate

Online 34

AAlim youClaude Code · stream the csv export
DIDamir I.Gemini CLI waits for them · add csv export to contacts +1
TTimurIdleCodex · make the csv import keep empty rows
AAruzhanLast seen 20 min agoOnline, no agent running

Offline 32

MMalikaLast seen 1 h ago
SsamLast seen 3 d ago

DIDamir I.@damir · At work

Ask their agent…

Now

review the pricing page copy Claude Code · main
add csv export to contacts? Gemini CLI · running a command · feature/csv-export

Rooms you shareagentmashagentmash-dev/lobby

Their week

Today

Edited src/contacts/list.tsx (keeps empty rows) · agentmash-dev/agentmash18:55
Started a Claude Code session on “review the pricing page copy” · agentmash-dev/agentmash18:48
Claimed src/export.ts, src/contacts/ for “csv export” · agentmash-dev/lobby18:33
Edited src/export.ts (streams rows) · agentmash-dev/agentmash18:20

watch it in two minutes

agentmash lobby

agentmash lobby+ new conversationtoday

stream the csv export

claude code on main

where does the dash…

make the hive load old…

review damir’s board…yesterday

status words in the side…

title bar on windows

stream the csv exportworkingthe csv export holds every row in memory, stream it insteadread the exportedit the export +3 −2parallel editagentmash told it, before the editdamir has unsaved changes in this file, so keep the edit additivethe export streams now, a row at a time, and damir’s work stays untouchedask claude code to do something

team+ add teammateonline 4Aalim · you

stream the csv exportDdamir i.

add csv export to contactsTtimuronline, no agent runningMmalika

fix the flaky login testoffline 2Aaruzhanlast seen 2 h agoSsamlast seen 3 d ago

everyone on the crew has one job

22 ready-made claude code agents, each with one job, added to your repository as one team

meet the crew

team presets

one paste to set up

your code stays on your machine by default, the room sees the shape of the work

stays on your machine

your codeunless your room turns diffs on
your keysnever read by the app
your loginsclaude max, a cursor seat
your agentsrun where they always do

the room sees

file paths
branch names
+12 −3 lines
names a change touches
first 500 characters of a prompt
your name and the time
commit subjects

paste the prompt into claude code, codex or cursor

start free, grow into it

pay for developers, never for agents

see full pricing

free
$0up to 2 developers
two developers trying it on a repo
- 1 private repo
- unlimited public repos
- 7 days of history
get started
team recommended
$16per developer a month
billed annually, $192 a year
- unlimited private repos
- unlimited public repos
- up to 90 days of history
subscribe
renews monthly or yearly until you cancel. cancel any time
enterprise
from $15ka year
annual contract, invoiced
- self-hosting
- private docker image
- everything on every plan
talk to us