Linger is an open-source personal agent with a Claude-class brain running locally on your PC.
Open-Source · Local AI · Consumer Macs & PCs
- A Claude-class brain, running locally. A 125B model on a 12 GB GPU, on everyday Macs and PCs.
- The Grok Bot experience on your PC. Zero token fees for local inference.
- A loyal AI sidekick. It remembers what matters, proactively gets things done, and keeps your data on your device.
Efficient AI model–harness co-design brings a full local agent to consumer Macs and PCs. Our MoE inference engine, PolyStrata, keeps a model's busiest experts on your GPU and the rest in RAM or on SSD, so larger models run with less GPU memory: the 125B Qwen3.8-Flash-Next runs on a 12 GB card with 64 GB of RAM. The model runs on your own machine instead of a remote server, so there is no cloud bill and no token fee.
Speed: how fast seven MoE models write on 8, 12 and 32 GB GPUs
| Writing, tokens per second | Laptop RTX 4070 8 GB |
RTX 3060 12 GB |
RTX 5090 32 GB |
|---|---|---|---|
| Qwen3.8-Flash-Next | 38.1 (Q2_0)¹ | 45.5 | 179 |
| LFM2.5-24B-A2B | 70–73 | 97.8 | 480–485 |
| Flint 40B-A2.5B | 61–63 | 65.5 | 251–252 |
| Qwen3.6-35B-A3B | 94–100 | 91.8 | 394–395 |
| Ornith 1.5 | 76–82 | 80–82 | 316–319 |
| Gemma 4 26B-A4B | 92–93 | 78.0 (Q4 file) | 353–356 |
| Tiel Coder 35B-A3B | 83 | 78.8 | 337–345 |
¹ Not from PolyStrata's results: our lab measured the Q2_0 file with a short prompt on a 96 GB card held to an 8 GB budget, with 32 GB of RAM and a 64K context. It is not a real 8 GB card, and it does not settle the earlier 95 tok/s IQ1_M figure: Flash-Next Coder IQ1_M gave 32.1 under the same conditions, and the full-model UD-IQ1_M was not tested.
Full results (5 more models!) can be found in PolyStrata.
One command detects your hardware, recommends the best model, installs it and checks it. In our lab we run an architecture search: 49 language models, two inference engines, quantizations from 1-bit to 8-bit and context windows from 64K to 256K, tested tier by tier on NVIDIA cards and Macs. On your computer, the installer reads your GPU, memory and bandwidth, keeps only what fits with at least a 64K context, picks the smartest model and then the fastest, and counts it as installed only after it answers a real question.
By machine: what each Mac and RTX tier gets, how fast it writes and which cloud model it matches
| Hardware | Recommended model | Writing, tok/s | SimplexAI | Compares with |
|---|---|---|---|---|
| Mac 16 GB | Spark-X2.6-4B Q4_K_M† | 37.6 short (37.4–37.8); 13.3 after a 126K prompt (13.30–13.31) | Not ranked | — |
| Mac 24 GB | Qwen3.8-27B UD-IQ3_S | 46.06 single session (median of 3); 48.43 for 2 sessions together | 51.9 | ≈ Claude Opus 4.5 |
| Mac 32 GB | Qwen3.8-27B UD-Q4_K_M | 45.44 single session (median of 3) | 54.0 | ≈ Claude Sonnet 4.6 |
| Linux NVIDIA 8 GB + 32 GB | Qwen3.6-35B-A3B UD-Q3_K_M | 121–146 short; 88–103 after a 62K prompt | 35.4 | ≈ Gemini 2.5 Pro |
| Linux NVIDIA 12 GB + 32 GB | Qwen3.8-Flash-Next GSQ-RCO Q2_0† | 21.3 short (19.8–25.4); 34.1 after a 62K prompt (12.0–37.3), first token after about 3 minutes | 57.3 | ≈ Claude Opus 4.6 |
| Linux NVIDIA 24 GB + 64 GB | Qwen3.8-27B UD-Q4_K_M | 32.5–33.2 single session at 130K; 28–41 per session with two sessions at about 8K | 54.0 | ≈ Claude Sonnet 4.6 |
Decode speeds were measured in our lab, mostly on a simulated tier (a larger card held to the tier's memory budget, or a larger Mac planned as the tier), so a real machine of the tier can be slower. SimplexAI scores sit on the third-party AIHOT scale, where 50 is the reference average, and scores within 3 points count as the same level. † Not this machine's default recommendation. Not ranked: AIHOT has no third-party results for the model. Test conditions, defaults and self-tests are in Local inference.
Tell one agent, and every agent knows. Another Me reads back through your inbox, then each night files your mail, chats and meetings into wiki pages that every agent can search. It keeps the information, preferences and decisions that matter across conversations, on your device instead of in a provider's account.
It reaches out first. Another Me watches your inbox, calendar and Slack, follows up and moves work forward, and messages you only when something needs you. And it improves itself recursively: every night it reads how its own messages landed and rewrites the rules it decides by, so every "not now" makes tomorrow's judgment sharper.
Every message Another Me sends on its own is also a test. Each night, Dreaming reads how those messages landed and rewrites the rules Another Me decides by; the next day it decides with the new rules, and how those land feeds the next night. No retraining: what it learned stays a short list of rules.
Implicit reinforcement: push back on the same kind twice, and it becomes a rule
You never state the rule; Another Me reads it from how you react. Want it sooner twice, and it tells you right away; turn it down twice, and it stays quiet.
Direct response: say how you want it, and one Dreaming makes it a rule
No second time needed. Name the kind and it changes whether you are told; name the hour and it changes when.
- Desktop and browser: Your apps, your Chrome, already signed in. Agents open the apps on your Mac, Windows PC or Linux desktop and click and type the way you do, and on a Mac or Windows PC they work in your everyday Chrome with the sign-ins you already have.
- Agent teams: One plan for every agent. When a job needs more than one agent, Another Me brings in the agents you have, creates the ones it is missing, and writes one
team.mdthat they all follow. - Task delegation: Long work runs in the background while the chat stays with you. An agent hands long work to background copies of itself and keeps talking with you while the work runs.
- Model scheduling: One model, many agents, your work first. Every agent, approval and background job shares the local model, so work you are waiting for starts at once and background work waits until the model is quiet.
- Bring your own runtime: Claude Code, Codex or Pi, on your GPU or your subscription. Each runs as it always does, on local models or on the plan you already pay for.
- Voice: Record a meeting with a live transcript, ask what was said, and hand the follow-ups to Another Me.
For one person on one computer. Choose your platform below. The app opens Linger, connects this computer and runs its agents. Quick Install on its setup screen sets up a Linger server on this computer too.
| Platform | Package |
|---|---|
| macOS 14.0 or later, Apple silicon | Linger-<version>-macos-arm64.dmg |
| Windows x64 | Linger-<version>-windows-x64-setup.exe, or -windows-x64-portable.zip to run without installing |
| Linux x64 | Linger-<version>-linux-x64.tar.gz (glibc 2.35 or newer) |
There is no package for Intel Macs; use Docker Compose there. Check each
download against SHA256SUMS.txt from the same release.
See Install the desktop app for the Windows and Linux steps and for help when setup fails.
For a server that several computers share, or to build from source. You need Docker with Compose and at least 5 GiB of memory for its virtual machine.
git clone https://github.com/simplexai-labs/linger.git
cd linger
cp .env.example .env
docker compose up --build --detach --waitCreate the owner account. As written, it signs in as owner@linger.local with
the test password test1234; replace either one to use your own:
printf '%s\n' 'test1234' | docker compose exec -T teamhub \
teamhub-account create-owner --email owner@linger.local --password-stdinChange the test password in Settings › Security before other people can reach the server.
Open http://localhost:3000 and sign in. The app then shows the commands that connect your computer. See Install with Docker Compose to serve other computers on your network and to turn on voice.
The full documentation lives at docs.heylinger.com.
| What it covers | |
|---|---|
| Desktop app | Install the app, or build it from source. |
| Choose a runtime | Claude Code, Codex or Pi, on local models or a subscription. |
| Self-hosting | Run it with Docker Compose to serve several computers. |
| Troubleshooting | Fixes for when something does not start. |
New here? Read the contribution guide and pick a good first issue.
Choose the channel that best fits your question:
- GitHub Issues: report reproducible bugs and track engineering work. Read the contribution guide before opening one.
- Email: write to tonyzhang@simplexai.com for anything that does not fit an issue.
- Security policy: report a vulnerability privately, never in a public issue.
Star Linger on GitHub to follow along, and choose Watch → Custom → Releases to hear about every new release.
Linger is built on these open-source projects:
- PolyStrata: Runs local models, keeping the experts a model uses most on your GPU.
- Strata: The inference engine PolyStrata is built on. It runs Qwen3.8-Flash-Next.
- BlockNote, Yjs and Hocuspocus: Power the documents that people and agents edit together.
- Hindsight: Where agent memory's fact extraction and search come from.
- Cua: Lets agents use the apps on your desktop.
- LiveKit: Carries the audio for voice and meetings.
- River: Runs TeamHub's background jobs.
- NOTICE: Lists every third-party component, what we adapted from it, and the license that applies.
Thank you to the open-source community for libraries this powerful and this simple: they let us focus on the product itself, and we hope our projects will in turn give everyone an easier-to-use knowledge base.
Have an idea? Open a feature request and describe the problem it would solve.
This repository is licensed under the Linger Open Source License, based on Apache 2.0 with additional conditions. NOTICE lists third-party code and the logos and images that this license does not cover.









