Justin launched Breadcrumb, a local encrypted Mac app that records screen, meetings, and AI transcripts to create searchable memory for AI. It exposes over 30 MCP tools and works with Claude Code, Codex, Cursor, and opencode. The tool aims to provide observability and context for AI-assisted development workflows.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
A free, no-account, locally encrypted Mac tool that records your screen, meetings, and AI sessions, letting Claude, ChatGPT, Cursor, and opencode search your digital life. It backs its business model with $299 founding seats and a future team edition, but every privacy and localization claim has no third-party verification.
Breadcrumb's core claim is that AI shouldn't start from scratch every conversation. It records the user's screen, meetings, AI sessions, and the decisions the user reaches with AI on the Mac, turning this content into AI-searchable memory. In its launch post on Hacker News, the company said tools like Claude, ChatGPT, Cursor, and opencode can retrieve this memory, letting new conversations pick up old context.
Based on the official demo, this memory is organized into queryable fragments: AI can answer "what did I owe from yesterday's standup" or "what did I do yesterday," and can even identify speakers in a meeting based on screen content. The company says transcription, speaker identification, and schedule organization are all handled by local models running on the Mac, without relying on the cloud.
Breadcrumb's privacy design is its most prominent selling point. The company says it runs entirely locally on the Mac, with history encrypted and stored using the user's own key, which is kept in the Keychain, and the company never gets a copy. The personal edition is free, with no account, email, credit card, subscription, or telemetry. Users can pause recording instantly from the menu bar; nothing is recorded while paused, and specific apps, websites, words, and phrases can be excluded, with incognito sites never recorded.
The company also says Breadcrumb identifies speakers through screen content rather than voice, and does not build anyone's voiceprint. It uses a 12-word BIP39 recovery phrase, which the company says belongs to the user. If these claims hold up, they do constitute a complete local privacy scheme. But it must be made clear: all privacy and localization claims come from the official page, with no third-party verification. In the absence of an independent audit, these can only be regarded as source_claim, not confirmed fact.
Breadcrumb currently supports four AI tools—Claude Code, ChatGPT, Cursor, and opencode—and the company says other tools will be supported soon. It offers 30+ MCP tools for AI to call, but the source does not list the specific functions and number of these tools item by item. Meeting transcription currently only supports Zoom; transcription for Google Meet, Teams, and others, as well as Windows and Linux versions, are listed as future plans with no timeline.
The hardware threshold is clear: it requires a Mac with at least 16 GB of memory, an Apple M1 chip or newer, and macOS 15 or later. This means older Intel Macs and devices with insufficient memory cannot use it. The specific models and performance of the local models are not explained in the source, so the actual accuracy of transcription and speaker identification cannot be judged.
Breadcrumb's scope is a folder with specific rules. When the user enters that folder to work, the rules take effect automatically. In the official example, /acme has 28 rules plus 22 global rules, /myapp has 91 rules, and /personal has 0 rules. After the user corrects the AI, the AI proposes solidifying the correction into a rule; once the user confirms, all AI tools follow that rule.
This design turns "context" from a one-off prompt into persistent configuration. In the official demo, the user says "never push main without my consent," and the AI proposes making it a permanent rule for /myapp, which takes effect once the user confirms. In addition, any action can become a one-word shortcut—for example, \fb-hours can generate a Friendbuy hours report and works across all AIs.
The company says the personal edition of Breadcrumb is free, and it will make money through a team edition in the future. To support development, users can pay a one-time $299 for a founding seat, with 1,024 available. Each seat includes a numbered seat, a private forum, early versions, a $2,500 credit toward the future team edition, and attribution, and is refundable.
The company also promises that no matter what happens to the company, the free version of Breadcrumb will never be removed, disabled, or restricted. This promise is very strong, but it likewise comes from the company's own statements, with no third-party guarantee. The actual sales of the 1,024 founding seats at the $299 price are unknown, and the specific form, pricing, and release date of the team edition have not been disclosed.
Breadcrumb's narrative rests on two unverified pillars: the performance of the local models and the verifiability of its privacy promises. The source does not explain the specific models and performance of the local models, so it cannot be judged whether the accuracy of transcription, speaker identification, and schedule organization reaches a usable level. The specific functions and number of the 30+ MCP tools are also not listed item by item.
On privacy, the company says "nothing is sent to us," but it also acknowledges that if users use Claude or ChatGPT, those AIs' privacy policies apply. This means that once data is retrieved by an AI, it enters that AI's privacy framework. The company suggests using local AI to avoid data outflow, but the capability boundaries of local AI are not explained in the source.
In addition, Google Meet and Teams transcription and Windows and Linux versions are only described as "coming soon," with no timeline. For non-Zoom users or non-Mac users, Breadcrumb currently cannot cover them. These limitations determine that it is currently only an early product aimed at a specific group, not a general solution.
All facts in this article come from the official Breadcrumb page published by InnerLoop on Hacker News. This source is a product launch post and constitutes first-party statements, but all privacy, localization, encryption, and permanent-free promises have no third-party verification. The specific functions of the 30+ MCP tools, the models and performance of the local models, the actual sales of founding seats, and the form and timeline of the team edition are all undisclosed in the source. Therefore, content involving privacy and future promises should be regarded as source_claim, not confirmed fact.
Breadcrumb tries to solve AI's context fragmentation with local encryption and a free model, and its privacy architecture and scope design have clear product logic. But all key promises—local operation, encryption, no telemetry, permanently free—come from the company's own statements, with no third-party verification. Until an independent audit appears, it is better seen as an attractive early product than a verified privacy solution.
Primary report
Primary source