News2026-11-20Updated 2026-11-20

Linger Personal AI Agent: A Grok Bot Alternative That Runs Locally

TL;DRLinger is a personal AI agent developed by Simplex AI, the company founded by former DeepSeek contributor Tony Zhang. Its code is available on GitHub, and it can use local models without cloud inference token fees. It can also connect to cloud services, which have their own costs and data policies. Linger is an independent project, not a DeepSeek release.

By James Carter and Olivia Bennett

Simplex AI has released Linger, a personal AI agent designed to run on a user’s own computer. The project pairs an agent interface with local model inference, persistent memory and tools for working with desktop applications and the browser.

Founder Tony Zhang announced the release on X, arguing that users should have more control over the context and data used by their assistants. His connection to DeepSeek has drawn attention to the project: an AWS case study describes Zhang as a core contributor to DeepSeek V1 and VL. Linger itself is built by Simplex AI, not presented as an official DeepSeek product.

Laptop on a desk, illustrating a personal computer workspace
Illustrative image: a laptop workspace, not a screenshot of Linger. Photo by Radek Grzybowski, CC0.

What does Linger do?

Linger is designed to keep context between conversations and help with tasks across applications. Its repository describes a local knowledge base, agent teams, background tasks and a proactive assistant called “Another Me.” With access to a user’s services, the assistant is intended to surface relevant information and follow up on work.

The project also describes desktop and browser control, including the ability to work in a browser where a user is already signed in. Zhang says actions such as sending email or other operations that leave the machine require a confirmation card and that users can take over. These are project descriptions, not a guarantee that every workflow will behave identically on every setup.

How does local inference work?

Local inference means the model generates its response on your own hardware. Linger points to PolyStrata, an inference project in the Simplex AI GitHub organization, to run supported models. For mixture-of-experts models, its approach keeps frequently used components in graphics memory while using system RAM or storage for others.

Simplex AI gives one example: a 125-billion-parameter Qwen3.8-Flash-Next model running with a 12 GB GPU and 64 GB of system RAM. That example should not be read as a 12 GB total-memory requirement. Model choice, quantization, context length and available storage all affect what a machine can run and how responsive it feels. The project’s own speed figures are lab measurements, and some hardware tiers are simulated.

Close view of computer processor hardware
Illustrative image: computer hardware. Photo by RemazteredStudio, CC0.
QuestionWhat the project says
Supported computersLinger lists desktop packages for Apple silicon Macs, Windows x64 and Linux x64; requirements vary by installation and model.
Local token chargesLocal model inference does not create a cloud inference token bill. Hardware, electricity and storage still cost money.
Cloud modelsLinger also lists Claude Code, Codex and Pi as runtime options. External providers’ terms and charges apply when their services are used.
DeepSeek V4 FlashPolyStrata lists experimental Linux support and a 91 GB model file. Its documented configuration calls for a 32 GB GPU and about 72 GB of system RAM alongside it.

These specifications come from the project’s documentation as reviewed on November 20, 2026. They can change, and the DeepSeek V4 Flash figures are not interchangeable with the smaller GPU example for Qwen.

Does data always stay on the computer?

Local inference can keep model processing on the machine, and Linger says its knowledge base and memory are stored on the user’s device. That does not make every task offline. A cloud model receives the information sent to it, and connected email, calendar or other web services communicate with their respective providers. The exact data path depends on the runtime, integrations and permissions selected.

Anyone considering the tool for sensitive work should check those settings, understand what each connected service receives and review permissions before granting access to accounts or local files. A local agent with browser and desktop access can also make consequential changes, so review its requested actions.

Is Linger a Grok Bot alternative?

The repository describes Linger as offering a Grok Bot experience on a PC. The comparison is a statement of product positioning: both are presented as assistants that can perform tasks on a user’s behalf. It does not establish that their capabilities, integrations, security controls or output quality are equivalent. We have not independently tested them against each other.

Can businesses use the code?

Simplex AI calls Linger open source, but the repository uses the Linger Open Source License, based on Apache 2.0 with additional conditions. The license allows commercial uses such as internal use and use as a backend for an application, while restricting certain hosted or managed offerings to external users and requiring retention of specified branding. Businesses should read the current license and obtain permission where required. It is not the standard Apache 2.0 license.

Why this matters for DeepSeek users

For developers following DeepSeek, Linger is an example of a separate team building an agent around locally runnable models. PolyStrata lists DeepSeek V4 Flash among its supported experimental models, but running it locally currently demands substantial resources. Users with less powerful machines may need to choose a smaller supported model or a cloud runtime, each with different performance and privacy trade-offs.

The practical question is whether the combination of model quality, available memory and agent permissions fits the job. The repository provides a starting point; the experience still needs to be evaluated on the intended hardware and tasks.