Bits: An Open-Source Personal Agent That Runs Locally and Handles Everything for You

In an era where cloud-based AI assistants are everywhere, an open-source product called Bits has chosen a different path: it runs entirely on your local machine, yet can work around the clock without interruption. Recently, Bits gained 77 upvotes and 5 comments on Product Hunt, reaching #14 on the daily leaderboard. Its combination of "open source, local, and memory-capable" has drawn attention from developers and productivity tool enthusiasts.

A Local Agent That Never Rests

At its core, Bits is a small background daemon. Unlike typical chatbots, Bits is designed to run continuously: routines wake it on a schedule, and tasks keep executing even after the app is closed. This means you can have Bits handle recurring work in the background—such as organizing your inbox, monitoring feeds, or generating reports—without keeping a chat window open.

This design clearly draws on the reliability principles of server-side task queues. Bits persists the state of every execution step, so even if the program crashes, tasks can resume from where they left off rather than starting over. For long-running or multi-step tasks, this greatly reduces the risk of duplicated work.

Developer Harsh Patel emphasizes in the description that Bits "runs on your own machine, and never sleeps." This is both a selling point and a hint at its fundamental difference from cloud agents: data never leaves your local environment, and computation relies on your own hardware.

Daemon is a background resident process in an operating system that runs independently of any user interaction interface. It is typically started at system boot and continuously listens for events or performs periodic work in the background until explicitly terminated. Windows services, macOS launchd jobs, and Linux systemd units are all typical forms of daemons. Bits adopting this model means its task scheduling does not depend on whether a chat window is open, and it also makes its reliability design—state persistence and crash recovery—comparable to server-side task queues.

Task crash recovery relies on the "checkpoint" approach: writing the intermediate state of long tasks to persistent storage so that when a task fails, it resumes from the most recent checkpoint rather than rerunning the entire flow. This pattern is common in distributed task queues but rare in personal local agents, because it requires additional state management overhead.

Memory in Markdown: Controllable and Editable

Bits stores its long-term memory as plain Markdown files. Users can directly open the files to read, modify, or even delete memory content without going through a complex settings interface. This choice lowers the barrier to understanding and gives users full control—you can adjust the AI's understanding of you just like editing notes.

For users who care about privacy or want deep customization of AI behavior, plain-text memory is a transparent and portable solution. You don't need to export a proprietary format, and you don't have to worry about data being locked into a platform. Bits essentially turns memory into a set of files managed by users themselves, which is highly consistent with its open-source positioning.

Sentinel Review: Trust Only Your Own Words

When agents execute tasks automatically, risk control is an unavoidable issue. Bits's solution is a built-in reviewer called Sentinel: when risky operations are involved, Sentinel trusts only the user's own explicit words, rather than letting the model make its own judgment.

This design point is worth pondering. In many agent systems, whether to execute sensitive operations is often decided by the model based on context, which introduces security risks such as prompt injection. Bits returns the final decision-making power to the user's explicit instructions, effectively adding a "human language" verification layer between the model and real-world operations. Although the details have not been fully disclosed, this approach reflects a pragmatic consideration of agent security.

Prompt injection refers to attackers hiding malicious instructions in content that the agent will read—web pages, emails, documents, tool return results—inducing the model to treat "data" as "instructions." For agents that can automatically operate files, send messages, and call APIs, the consequences of such attacks are far greater than in chat scenarios.

Bits's Sentinel strategy is to shift the basis for deciding "whether to execute" from the semantics of the model's output back to the explicit words the user wrote in advance. This is similar to the industry's "human-in-the-loop" approval approach, but stricter—it does not require manual confirmation for every operation, but requires that the authorization text itself be traceable to the user. The limitation lies here as well: the user's original words may be ambiguous or misinterpreted in different contexts, so the design of expression standards and permission boundaries remains critical.

Live Cards, Voice, and Global Hotkey

On the interaction side, Bits does not follow the pure text chat format. It presents results using live cards and mini-apps, making information display more structured. It also supports voice conversation and provides a global hotkey ⌥Space for quick access, clearly optimized for daily high-frequency use.

Regarding model compatibility, Bits claims it can work with any model, including locally running models. This keeps it usable in offline or privacy-sensitive scenarios and makes it easy for users to flexibly switch backends based on cost and privacy needs.

Live cards are a structured output interface: the agent renders results as updatable, clickable components rather than a block of static text. Compared to pure conversation, cards can carry progress, status, action buttons, and dynamic data, and refresh in real time as tasks execute—suitable for monitoring, checklists, and reporting scenarios. Mini-apps go a step further, essentially lightweight interactive interfaces generated or configured by the model that solidify the common operations of a task.

This hybrid "conversation + cards" format is becoming a common paradigm for agent products: conversation handles intent expression, while cards handle result presentation and follow-up actions—the two complement each other.

Differentiation in the Open-Source Agent Space

The personal AI agent track is becoming increasingly crowded. Cloud solutions have advantages in computing power and ecosystem, but local solutions have unique value in privacy, cost, and controllability. Bits's differentiators can be summarized in three points: continuous execution enabled by a local daemon, transparency enabled by Markdown memory, and controllable security enabled by Sentinel.

Of course, as an early-stage project, Bits's ecosystem and integration richness cannot yet compare with mature platforms. The engagement of 77 upvotes and 5 comments on Product Hunt also indicates that it is still in the community validation stage. But the combination it proposes—open source, local, memory-capable, and reviewable—provides a valuable reference sample for personal agent design. For users willing to get their hands dirty and who care about data sovereignty, Bits may be a starting point worth watching.

One-line Summary

Bits is a locally running 24/7 open-source AI agent that achieves continuous execution and controllable security through a daemon, Markdown memory, and Sentinel review.

Paragraph Summary

Bits is an open-source AI agent that runs entirely on a local machine. It recently received 77 upvotes on Product Hunt and ranked #14 for the day. Its core is a background daemon that wakes up via scheduled routines to execute tasks, continues running even after the app is closed, and achieves crash recovery through state persistence. Long-term memory is stored as plain Markdown files that users can directly read, modify, and delete, balancing transparency and portability. For security, it includes the Sentinel reviewer, which for risky operations trusts only the user's explicit words rather than the model's autonomous judgment. For interaction, it supports live cards, voice, and a global hotkey, and is compatible with any model, including local ones. As an early-stage project, its ecosystem integrations are still limited, but it provides a differentiated design sample for local personal agents.