Starts Nov 23Live Cohort 3 of Vibe Coding & AI Agents (Nov 23, 18:30 MSK): 3 weeks, 9 live hands-on sessionsCohort 3 (Nov 23) →
ReleaseCode / SDK · ActionableX: @gengdaJ→ original

Linger Runs an AI Agent on a 16 GB Mac With No Token Fees and Promises a 5x Speedup

Linger is an open-source personal agent that runs on your computer rather than in the cloud, so there are no token fees. It requires a Mac with at least 16 GB of memory: the installer picks one of 12 open models, such as Qwen or Gemma, for you. The authors claim the PolyStrata engine speeds up responses by up to 5x, but there are no independent benchmarks. Private tasks can stay local, while complex ones can be routed to Claude Code or Codex.

AI-processed from X: @gengdaJ; edited by Hamidun News
Linger Runs an AI Agent on a 16 GB Mac With No Token Fees and Promises a 5x Speedup
Source: X: @gengdaJ. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceX: @gengdaJ ↗
Editor-in-Chief
Agent memory is a wikidaily reports, projects, topics, people, and companies
Plugins via Composio10,000 free calls per day
Key metric #3Local and cloud agents work together in one room on a task

Typically, an AI agent works like this: your requests go to someone else's cloud, and you pay for every token, the small chunk of text the model reads or writes. Bills grow, and private data leaves your computer. The open-source project Linger proposes the reverse: a personal agent lives directly on your laptop and runs open models locally, so you don't pay for cloud tokens.

The minimum requirement is a Mac with 16 GB of RAM, and the developers claim response speeds up to 5x thanks to their own PolyStrata engine. The project was announced by the Chinese developer @gengdaJ, who says the team behind Linger comes from the DeepSeek circle. We are relaying this connection based on his word; the post contains no separate confirmation.

How Linger Works on Your Hardware

By default, Linger uses the Pi framework. A built-in "quick install" checks your hardware and selects a suitable model on its own. You don't have to guess whether your laptop can handle a 7-billion-parameter version or a 30-billion-parameter one: the installer decides for you.

It supports 12 popular open models, including Qwen, DeepSeek, and Gemma. Speed is handled by the PolyStrata inference engine, which is credited with the "up to 5x" speedup. This is the authors' own claim: the post includes no independent measurements, and the result depends on the model and the machine. It's worth verifying the figure on your own hardware before building plans around it.

Local Models Plus Cloud Subscriptions

While local models still trail the top cloud models, Linger doesn't force you to choose one or the other. If you already pay for Claude Code or Codex, you can import them through the CLI and use them as the agent's "brain." The subscription doesn't go to waste, and Linger provides a convenient wrapper around it.

There is also a collaborative mode: local and cloud agents can be placed in one "room" where they solve a task together. The post's author tested this on an everyday example. He recently moved to Shenzhen and asked Codex, together with a local model, to pick a neighborhood to live in. By his assessment, the interaction went smoothly, and the final report looked tidy.

For people building agents, this is a practical setup. Cheap and sensitive operations, such as sorting email or notes, run on the local model. Complex reasoning goes to the cloud model. Costs drop, and extra data stays on the machine.

Integrations and Long-Term Memory

Connections to external services run through Composio. The free tier is 10,000 calls per day. The advertised set covers email and calendar (Gmail, Outlook, Google Calendar, Outlook Calendar), work tools (GitHub, Notion, Google Meet), and creative services (Figma, Canva, Higgsfield, Fal).

There is also a practical caveat. The author writes that Composio blocks sign-ups from the IP addresses of some hosting providers, such as Dmit. In that case, the key can be obtained by asking Grok Bot. It's a small detail, but it shows that part of the setup still depends on external cloud services.

Agent memory is organized as a wiki. Conversations are sorted into sections: daily reports, projects, topics, people, and companies. The agent remembers details from past tasks and becomes more accurate over time. For developers, this format is convenient because it consists of ordinary structured notes: they can be read, edited, and stored in version control.

What to Keep in Mind

The only source is a post on X written by an enthusiast of the project. The claimed 5x speedup and the product's "maturity" are not yet confirmed by testing. The quality of local models remains a constraint: the author himself acknowledges that for serious tasks, Claude Code and Codex are still needed. Running larger models comfortably requires a powerful computer or an Nvidia graphics card, which also costs money.

Nonetheless, the direction is clear. In the announcement, the authors recall that in the 1960s computing was centralized until the personal computer gave each user their own power; now AI is trained in data centers and rented to users by the token. If on-device models become faster and smarter, token costs could be replaced by an electricity bill, and data would never leave the computer.

Linger is one of the first open projects to assemble this approach into a finished product: model selection, integrations, memory, and a bridge to cloud subscriptions. For teams that care about privacy and cost control, it makes sense to install it on a test machine and compare it with a cloud agent on their own tasks.

Story

Key Questions Answered

How Linger Works on Your Hardware?

By default, Linger uses the Pi framework. A built-in "quick install" checks your hardware and selects a suitable model on its own. You don't have to guess whether your laptop can handle a 7-billion-parameter version or a 30-billion-parameter one: the installer decides for you. It supports 12 popular open models, including Qwen, DeepSeek, and Gemma. Speed is handled by the PolyStrata inference engine, which is credited with the "up to 5x" speedup.

What are the key details on Local Models Plus Cloud Subscriptions?

While local models still trail the top cloud models, Linger doesn't force you to choose one or the other. If you already pay for Claude Code or Codex, you can import them through the CLI and use them as the agent's "brain." The subscription doesn't go to waste, and Linger provides a convenient wrapper around it. There is also a collaborative mode: local and cloud agents can be placed in one "room" where they solve a task together. The post's author tested this on an everyday example.

What are the key details on Integrations and Long-Term Memory?

Connections to external services run through Composio. The free tier is 10,000 calls per day. The advertised set covers email and calendar (Gmail, Outlook, Google Calendar, Outlook Calendar), work tools (GitHub, Notion, Google Meet), and creative services (Figma, Canva, Higgsfield, Fal). There is also a practical caveat. The author writes that Composio blocks sign-ups from the IP addresses of some hosting providers, such as Dmit. In that case, the key can be obtained by asking Grok Bot.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…