agentim
Your AI agents. Every chat app. Your Mac.
A secret-agent octopus on a switchboard of AI models. Message him on Telegram, WhatsApp, Slack, Claude or ChatGPT, and he plugs every chat into the right model, running privately on your own Mac. Free and open source.
Channels
No new app to open. Agent Tim lives in the chats you already use, with the same agents, skills and memory everywhere.
Private by default
agentim pulls open Hugging Face models into Docker on your own machine, but only after checking they fit in memory. Cloud models are optional, and anything that looks sensitive never leaves home.
per message on local models
default model, memory-checked before download
bytes of sensitive prompts sent to the cloud
The Mac app
A real macOS app to manage conversations, agents, skills and models, and to keep Agent Tim's switchboard running. It's right there in your menu bar too.
Every answer shows which model handled it, and why.
Onboarding
Open the app and Agent Tim does the rest. He installs Docker if you don't have it, asks which models you want for each job, checks they fit in your Mac's memory, then downloads and starts everything.
Qwen3, Llama 3.2 or Gemma 3, from 1.7B up to 30B, sized to your Mac.
Qwen2.5 Coder or Qwen3 Coder, used whenever a message looks like code.
Send a voice note; Whisper listens and Kokoro answers out loud, all on this Mac.
LTX Desktop renders clips and pictures straight from chat.
Startups & Ideas
Pitch as many startups as you like. Agent Tim interviews you with ten questions, five as a YC partner and five as a co-founder. Then a team of agents who act like real employees gets to work, meeting every 10–30 minutes, handing work to each other by name, and keeping a living plan. You get a short summary after every meeting.
Type in the agents room and the agents speaking next answer you first. Watch who's in the room and who's typing, live.
"@Sam come up with 2 designs" and Sam answers right away. Tab completes the name.
Every meeting ends with decisions and action items, each with an owner and a date, plus metrics and open questions.
The Launch tab takes it all the way: Claude Code designs and builds v1, your tests run, then it deploys to Vercel, Netlify, Cloudflare, Fly.io or Docker and checks the URL is live. You approve the build and the deploy.
Point the team at local folders or GitHub repos, describe a feature, and Claude Code plans it in every service before anyone touches a file.
Start an ad-hoc chat with any mix of agents and skills. Everyone answers in turn, or @tag one.
CPO: the problem, the MVP, what to cut.
Designer: onboarding, key screens, trust.
CTO: the smallest shippable version.
CMO: positioning, channels, launch.
Marketing Automation: funnels that run themselves.
CRO: pipeline, pricing and revenue.
BizDev: partners and channels.
CFO: unit economics, runway, raising.
Data Analyst: the numbers behind every decision.
CISO: security that helps you sell.
COO: owners, dates and a weekly rhythm.
CEO: keeps the plan and sends you the summary.
/startup new Leafy: a monthly subscription that swaps office plants # Tim asks 10 questions (YC partner + co-founder), then the team meets every 10–30 min /startup # your startups · /startup now 2 · /startup off 2
The router
Agent Tim reads each message and picks the model that fits: local first, cloud only when it's worth it. Want a specific one? Type @alias.
Built-in skills
Problem, customer, market math, riskiest assumption and a $100 test you can run this week.
/validate-your-startupNiche, names, a content calendar, titles, thumbnails and hooks that keep viewers watching.
/youtube-channel-creatorDebugging, design trade-offs, and reviews that list the real bugs first.
/software-engineerOne-page PRDs, RICE prioritization, and metrics with real baselines.
/product-managerPlans, critical paths, risk registers, and status updates under 150 words.
/project-managerTurns an idea into on-model shot prompts, then renders each clip with LTX.
/video-directorRisk-based test plans, Playwright tests, accessibility checks and bug reports you can act on.
/web-qaDevice matrix, Maestro/XCUITest/Espresso flows, offline and interruption testing, store readiness.
/mobile-qaContract and negative tests, OWASP API Top 10 checks, pytest suites and k6 load tests.
/api-testingNorth star, funnels, retention, tracking plans, dashboards and clean experiments.
/data-analyticsArchitecture, build vs. buy, delivery speed; product strategy, discovery and roadmaps.
/cto · /cpoPositioning, channels and launches; security programs, compliance and incident readiness.
/cmo · /cisoVideo
Ask from any chat. LTX Desktop renders the clip on your Mac, for free, and Agent Tim sends it right back. This one was made exactly that way.
Lid closed + charger = still answering. On battery it sleeps like normal.
Anything that stops gets started again, within a minute.
Boots again after a power cut and brings the whole stack back.
Models that won't fit are refused before a single byte downloads.
Works with what you already use
agentim is an MCP connector, so Claude and ChatGPT can ask your local models and agents, make images and videos with LTX, and run your startups. And Agent Tim can hand coding tasks to the Claude Code or Codex already on your Mac.
One click in This Mac → Add to Claude. Everything stays on your Mac.
This Mac → Connect ChatGPT opens a private HTTPS link to paste into ChatGPT's connectors.
@claude-code review the auth module. Continue your existing Claude Code sessions with /claude. Read-only unless you allow edits.
Get started
Onboarding installs Docker if needed, lets you pick chat, coding, voice and video models, and sets everything up. It's the whole quick start. Prefer the terminal? Read on.
git clone https://github.com/CrewOS-Agentim/agentim.git && cd agentim
brew install xcodegen && xcodegen generate && open agentim.xcodeproj # ⌘R
You need a Mac with Apple silicon and Docker Desktop.
git clone https://github.com/CrewOS-Agentim/agentim.git cd agentim
Creates .env with fresh secrets, then installs the restart watchdog, the LTX runner and the lid-close power guard. It asks for your password once.
./deploy/macos/install.sh
Add a Telegram bot token from @BotFather and/or Slack tokens to .env, plus who's allowed to talk to Tim. For WhatsApp, scan the QR at localhost:3000/dashboard.
open -e .env
docker compose up -d # Hugging Face models download on first start
Build it once with Xcode and drag it to Applications. Then in This Mac, click Connect this Mac's models.
brew install xcodegen && xcodegen generate
open agentim.xcodeproj # ⌘R
Message your bot. Try /help, /agents, /validate-your-startup my idea… or /video a cat surfing at sunset.
Open source
agentim is MIT licensed and made in the open. Every star, idea, bug report and pull request makes Agent Tim better for everyone.
Something broken? Tell us what happened; the app's Help → Report an Issue fills in your Mac's details.
Report a bugA skill, channel or model you'd love to see? Ask for it, or start with an idea in Discussions.
Suggest a featureQuestions, setups, show and tell: meet the people running Agent Tim on their Macs.
Open Discussions