BetaEditorial Listing

Linger

SimplexAI Labs · Linger: Local-First Personal Agent That Runs on Your Mac or PC

Open Linger

Linger is a personal agent that runs on your own Mac or PC: it installs a local model sized to your hardware, watches your Gmail or Outlook inbox, Google Calendar and Slack mentions, files your mail, chats and meetings into a nightly-built wiki, and messages you first when something needs you. It suits technical individuals who want a proactive assistant whose memory, and optionally its model, stay on their machine, and who are comfortable installing early 0.1.x software.

PricingFree
Setupmedium
Runs onDesktop · Self-hosted
Open sourceNo
DocsYes
CategoryProductivity
Personal AgentProactive AgentLocal AILocal LLMSelf-HostedMemoryPersonal WikiEmailCalendarSlackComputer UseBrowser AutomationAgent TeamsDesktop App

Best for

Technical individuals with an Apple silicon Mac or an NVIDIA GPU PC who want a proactive personal agent for inbox, calendar and Slack follow-ups, with its memory and, optionally, its model running on their own hardware

Not ideal for

Non-technical users who want a polished, signed app or a hosted service today, teams that need a mature shared deployment with documented support, and anyone whose machine lacks an Apple silicon chip or NVIDIA GPU but still wants local inference

Who it's for

Developers, founders and other technical knowledge workers who want a private, proactive personal agent on their own Mac or PC

Capabilities

  • Proactive Another Me agent that polls Gmail and Outlook inboxes, Google Calendar changes and Slack mentions every 10 minutes and decides whether to message you now, follow up later or stay silent
  • Nightly Dreaming pass that rewrites Another Me's notification rules from your reactions, so the same pushback twice, or one direct instruction, becomes a rule
  • Personal wiki built each night from mail, chats, meetings, files and agent conversations, searchable by every agent and stored on your server
  • Hardware-aware installer that detects GPU, VRAM, RAM and bandwidth, recommends a local model with at least a 64K context, and verifies it with a real question before counting it as installed
  • PolyStrata engine (MIT) that keeps a mixture-of-experts model's busiest experts on the GPU and the rest in system RAM or on the SSD
  • Runs agents on local models via Pi, on an existing Claude Code or Codex subscription, or on your own OpenAI-compatible or Anthropic endpoint (BYOK)
  • Desktop control on macOS, Windows and Linux, plus work inside your everyday signed-in Chrome on Mac and Windows
  • Agent teams that share one team.md plan, with Another Me creating missing agents for a job
  • Background task delegation so long work runs while the chat stays responsive
  • Dedicated commands for Gmail, Outlook, Google Calendar, Slack, Notion, Google Drive, Figma and GitHub, with other apps reachable through Composio
  • Confirmation cards before sending mail, posting to Slack, editing Notion or commenting on GitHub
  • Meeting recording with live transcript and voice mode (with your own Soniox and xAI keys)

Limitations

  • Early software: first released on 2026-11-10 and at desktop v0.1.6 as of 2026-11-20; the homepage still lists downloads as coming soon and points to a waitlist for hosted access
  • The macOS and Windows packages are not signed, so macOS and Windows SmartScreen block the first launch until you confirm it (on a Mac, by clearing the quarantine flag or allowing it in System Settings)
  • Local models need an Apple silicon Mac or an NVIDIA GPU on Windows or Linux; 8 GB Macs get no automatic local-model plan, there is no Intel Mac package, and no AMD or Intel GPU path is documented
  • Speed and quality figures come from SimplexAI's lab, mostly on simulated hardware tiers; the docs say acceptance on physical 8 and 12 GB NVIDIA cards and 16 GB Macs is still pending
  • The 125B-on-a-12 GB-GPU headline is not the default plan: the documented default for a 12 GB Linux card is Qwen3.6-35B-A3B, and 12 GB Windows cards get Spark-X2.6-4B
  • Connected apps require a Composio API key, and app requests and sign-ins go through Composio; voice requires Soniox and xAI keys
  • Agents act with your account's full access on the connected computer; the Guardrail's model review is off by default and agents on Pi skip the computer checks entirely
  • Signed-in Chrome control is documented for Mac and Windows only, and Linux desktop control works only on X11, not Wayland
  • Docs pages disagree on background memory and wiki work: some say it runs on a signed-in Claude Code or Codex (counting against that account), while the local-inference guide says Pi can run it on its own model from Computer Daemon 0.55.0
  • Source-available under a modified Apache 2.0 license that requires a commercial license to offer Linger as a hosted service to people outside your organization or to remove the frontend logo

Use cases

  • Having an agent watch your inbox, calendar and Slack and nudge you only about client replies, schedule changes and deadlines
  • Building a searchable wiki of clients, contacts and projects from months of mail and meetings, kept on your own machine
  • Running a personal agent on a gaming PC or Mac without paying per-token fees for routine background work
  • Recording a meeting, asking what was decided, and handing the follow-up email to Another Me for approval
  • Delegating a multi-step job, such as research plus a draft in Google Sheets, to a team of agents using your signed-in Chrome
  • Reusing a Claude Code or Codex subscription as the brain for an always-on personal agent

Our take

Linger stands out among personal agents because it ships the model stack too: a planner that sizes a local model to your machine and an expert-offloading engine that lets 8 and 12 GB Linux cards run a 35B mixture-of-experts model. That can make a proactive, always-watching agent cheap to leave running, and the nightly rule rewriting is a sensible answer to the main problem with proactive agents, which is noise. Read the fine print before relying on it. It is less than two weeks old, the impressive hardware numbers come from simulated tiers rather than shipped acceptance tests, inbox and Slack access route through Composio, and the default security posture trusts agents with your full account. Treat it as a promising tool for technical early adopters, not yet a set-and-forget assistant.

Who should use it

Developers and technical founders with an Apple silicon Mac or an NVIDIA GPU PC who want an agent that follows their inbox, calendar and Slack, keeps its memory locally, and can run without per-token costs, and who are comfortable with early releases and some terminal setup.

Who should skip it

People who want a hosted assistant that works out of the box, organizations that need audited security defaults or vendor support, users on Intel Macs or machines without an NVIDIA GPU who want local inference, and anyone unwilling to route connected-app access through Composio.

Strengths

  • Model, wiki and memory can all stay on your own machine, with no token fees on a local model
  • Installer picks and verifies a local model for your hardware instead of leaving model selection to you
  • Proactive agent learns when to speak up from your replies, and you can change a rule by replying in plain words
  • Writes that reach other people wait for a confirmation card
  • Can reuse a Claude Code or Codex subscription when local hardware is not enough

Weaknesses

  • 0.1.x software released in November 2026, with unsigned macOS and Windows packages
  • Headline local-model claims rest on simulated lab tiers; physical-device acceptance is still pending
  • Connected apps depend on Composio, an outside service with its own key
  • Action review is off by default and does not cover agents on Pi
  • No page lists the notification rules Another Me has learned yet; you adjust them only by replying
  • Modified Apache 2.0 license restricts offering it as a hosted service

Linger pricing

Self-hosted

$0

  • Desktop app for macOS (Apple silicon), Windows x64 and Linux x64, or Docker Compose server
  • No token fees when running a local model
  • Use an existing Claude Code or Codex subscription, or your own API endpoint, instead

Note: Linger itself is free to download and self-host. Optional outside services are billed by their providers: a Claude Code or Codex subscription or your own API endpoint if you do not use a local model, Composio for connected apps, and Soniox and xAI for voice. Hosted access is waitlist-only, and no pricing for it is published.

Technical specs

Modalities

Text, Voice, Desktop and browser control

Available models

Qwen3.8-Flash-NextQwen3.8-27BQwen3.6-35B-A3BSpark-X2.6-4BClaude Code (subscription)OpenAI Codex (subscription)OpenAI-compatible or Anthropic endpoint (BYOK)

Where Linger excels

Catching client follow-ups without living in the inbox

Another Me checks Gmail or Outlook, Google Calendar and Slack mentions every 10 minutes and learns from your reactions which kinds of messages deserve an immediate ping.

Keeping a private knowledge base of your work relationships

The nightly wiki turns mail, chats and meetings into pages about clients, contacts and projects that every agent can search, stored on your own server rather than a provider's account.

Running an always-on agent on a gaming PC

The installer fits a local model to the GPU and RAM you have, so background work does not run up per-token API bills.

Linger vs. competitors

Linger vs. OpenHuman

OpenHuman is a GPL-3.0 desktop personal agent that connects email, chat and work apps and keeps a persistent memory, aimed at people who do not want to wire API keys or run a server. Linger also runs on the desktop but brings its own local-model planner and PolyStrata engine and a proactive agent that tunes its own notification rules nightly, at the cost of more setup and outside keys for connected apps.

Linger vs. OpenClaw

OpenClaw is an open-source assistant you reach through messaging apps such as Telegram, WhatsApp and Slack, running the model of your choice. Linger is built around its own web and desktop app, a nightly-built personal wiki and a proactive agent that watches your inbox, calendar and Slack, with local-model installation handled by its planner.

Linger vs. Hermes Agent

Hermes Agent from Nous Research is a model-agnostic agent that runs on your own server, accumulates memory and creates reusable skills from experience. Linger targets a personal machine instead, with hardware-sized local models, desktop and Chrome control, and an inbox-and-calendar-focused proactive agent.

Linger vs. IrisGo

IrisGo is a desktop agent for Windows 11 and Apple silicon Macs that learns a task by watching you once and reruns it on a schedule, alongside a daily brief and email triage. Linger focuses on proactive follow-ups and a searchable personal wiki, with self-hosted, source-available code and a local-model stack.

Frequently asked questions

What does Linger do?

Linger is a personal agent that runs on your own computer. Its main agent, Another Me, watches your connected Gmail or Outlook inbox, Google Calendar and Slack mentions, follows up on work and messages you only when something needs you. Each night it files your mail, chats and meetings into a searchable wiki and updates its rules for when to notify you. It can also control desktop apps and your signed-in Chrome, run agent teams and hand long tasks to background agents.

Is Linger open source?

Its code is public on GitHub under the Linger Open Source License, which is Apache 2.0 with added conditions: you need a commercial license from SimplexAI to offer Linger as a hosted or managed service to people outside your organization, and you may not remove the logo or copyright information from its frontend. Because of those conditions it is source-available rather than an OSI-approved open-source license. The PolyStrata inference engine it uses is MIT-licensed.

How much does Linger cost?

The desktop app and Docker Compose server are free, and a local model has no token fees. You pay separately for any outside services you choose: a Claude Code or Codex subscription or your own API endpoint, Composio for connected apps, and Soniox and xAI for voice. Some docs pages say background memory work runs on Claude Code or Codex, which would count against that account. Hosted access is on a waitlist with no published pricing.

What hardware do I need to run Linger's local models?

Local models need an Apple silicon Mac or an NVIDIA GPU on Windows or Linux. Setup checks your hardware and available memory before proposing a model. The smallest documented plans are Spark-X2.6-4B on a 16 GB Mac or an 8 GB NVIDIA card on Windows, and Qwen3.6-35B-A3B on an 8 GB NVIDIA card on Linux; 8 GB Macs and cards with less than about 7.5 GB of free VRAM get no automatic plan. If nothing fits, you can run Linger on a Claude Code or Codex subscription instead.

Does Linger really run a 125B model on a 12 GB GPU?

SimplexAI reports running the 125B Qwen3.8-Flash-Next on a 12 GB card with system RAM using its PolyStrata engine, but its published measurements were taken on a larger card limited to a 12 GB budget, and the default plan its installer documents for a 12 GB Linux card is the smaller Qwen3.6-35B-A3B. The docs say acceptance on physical 8 and 12 GB cards is still pending.

Does my data stay on my computer?

With the desktop app's Quick Install, the server and the computer are the same machine, so your wiki, memory, sessions and files stay on your disk, and a local model keeps model calls local too. Data leaves through services you turn on: a remote model or subscription, Composio for connected apps such as Gmail and Slack, and Soniox or xAI for voice. Background memory work that runs on a remote Claude Code or Codex subscription sends the conversations it reads to that provider. Agents also have your account's network access through their shell and browser.

Integrations & fit

GmailOutlookGoogle CalendarSlackNotionGoogle DriveFigmaGitHubGoogle ChromeComposioClaude CodeOpenAI CodexPiSonioxxAIDocker
Good fit forSolo / individual
Pricing modelFree· No cost to start
See pricing on Linger →

Alternatives to consider

About Linger

Linger pairs a proactive personal agent with its own local-model stack. You install the desktop app on an Apple silicon Mac, Windows x64 or Linux x64 machine (or run the server with Docker Compose so several computers can share it), and a guided installer reads your GPU, memory and bandwidth, proposes the strongest local model that fits with at least a 64K context, and counts it as installed only after it answers a real question. The documented default plans are Spark-X2.6-4B on 16 GB Macs and 8 to 12 GB Windows cards, Qwen3.6-35B-A3B on 8 to 12 GB Linux cards, Qwen3.8-27B on 16 GB and larger NVIDIA cards and 24 to 64 GB Macs, and Qwen3.8-Flash-Next on Macs with 96 GB or more. For mixture-of-experts models, SimplexAI's MIT-licensed PolyStrata engine keeps the most-used experts on the GPU and the rest in system RAM (or on the SSD on a Mac), which is how the Linux plans for 8 and 12 GB cards run a 35B model. Local models run through the Pi runtime; you can instead point Claude Code or Codex at the subscription you already pay for, or bring your own OpenAI-compatible or Anthropic endpoint. The agent at the center is Another Me. Once you connect apps (through Composio, which holds the app sign-ins), it polls Gmail and Outlook inboxes, Google Calendar changes and Slack mentions every 10 minutes, follows up on what it finds, and decides whether to tell you now, later or not at all. Each night a Dreaming pass files your mail, chats and meetings into wiki pages that every agent can search, and rewrites Another Me's notification rules from how you reacted to its messages, so pushing back on the same kind of message twice turns into a rule. Around it sit agent teams that share a team.md plan, background task delegation, desktop control on macOS, Windows and Linux, and work inside your own signed-in Chrome on Mac and Windows. Sending mail, posting to Slack and other writes that reach people wait for a confirmation card. The tradeoffs are real. The project was first released on 2026-11-10 and is at 0.1.x, the macOS and Windows packages are unsigned, and the docs say full acceptance testing on physical 8 and 12 GB NVIDIA cards and 16 GB Macs is still pending. Headline claims such as a 125B model on a 12 GB GPU rest on lab runs on simulated hardware tiers, and the documented default plan for a 12 GB Linux card is the smaller Qwen3.6-35B-A3B. Connected apps need a Composio key, voice needs Soniox and xAI keys, and the Guardrail's model review of desktop, browser and shell actions is off by default and does not apply to agents on Pi. The docs also differ on whether the nightly memory and wiki work can run on a local model through Pi or needs a signed-in Claude Code or Codex. The code is public under the Linger Open Source License, a modified Apache 2.0 that bars offering Linger as a hosted service to others without permission, so it is source-available rather than OSI open source.

Updates from Linger

New FeatureDesktop v0.1.6 released

Linger desktop v0.1.6 shipped on GitHub with builds for macOS (Apple silicon), Windows x64 and Linux x64, ten days after the first public release on 2026-11-10.

LaunchFirst public release

SimplexAI Labs published the Linger repository, under its source-available Linger Open Source License, and desktop v0.1.0 for macOS (Apple silicon), Windows x64 and Linux x64.

Are you the founder? Claim this listing →