My Brain Is Full, Crew: An Open Source Team of AI Agents That Manages an Obsidian Vault Through Conversation

A second brain is supposed to lighten the mental load, not add to it. In practice, many personal knowledge management systems end up doing the opposite. Folders multiply, tags drift out of control, notes pile up in an inbox that never quite gets emptied, and the very system meant to reduce overwhelm quietly becomes another task competing for attention. An open-source project called My Brain Is Full, Crew was built directly in response to that problem. Rather than offering another clever way to capture notes, it hands the entire operation of an Obsidian vault over to a coordinated team of AI agents that organize, file, connect, search, transcribe, and even triage email, all through ordinary conversation.

Where the Idea Came From

The project traces back to a very human experience rather than a theoretical product plan. Its creator, a doctoral researcher, spent years training a mind to hold an enormous amount of information at once: academic papers, half-formed ideas, deadlines, names, and the kind of fragmented late night theories that rarely survive until morning. For a long stretch of time, that mental juggling act simply worked.

Then it stopped working. Nothing dramatic happened, no single crisis or diagnosis marked the moment. Instead, there was a slow, creeping realization that the available mental bandwidth had run out. Things that used to stick started slipping through the cracks. Conversations got lost. Material that had been read carefully was suddenly hard to recall. A constant, low grade feeling of being behind became the default state.

The search for a fix led toward the popular combination of Obsidian and AI assistants that had already become common online. Most of what turned up, however, were essentially clever note capture tools, glorified search engines layered on top of a personal wiki. Useful for some purposes, certainly, but not enough for someone who needed far more than a better way to find old notes.

What was actually needed was a genuine brain dump system, something capable of organizing not just scattered knowledge but an entire overwhelmed life: physical health that had been neglected, an avalanche of unanswered emails, and a backlog of commitments that kept growing faster than they could be addressed. That gap between what existing tools offered and what an overwhelmed mind actually required became the starting point for building this project from scratch.

Who This Project Is Actually Built For

Most tools combining artificial intelligence with Obsidian are aimed at people who already have reasonably organized lives and simply want to optimize further. My Brain Is Full, Crew takes the opposite starting assumption. It is built for people who feel like they are drowning and need something closer to a lifeline than a productivity upgrade.

That audience includes doctoral students, researchers, and academics buried under papers and competing commitments. It includes anyone dealing with brain fog or an overloaded working memory, regardless of the underlying cause. It includes non native English speakers who want a system that genuinely operates in their own language rather than forcing every interaction through English first. And it includes anyone who has tried Obsidian before, admired its potential, and eventually gave up because maintaining it started to feel like an unpaid second job. Put simply, anyone who has thought “organization is clearly needed here, but there is no energy left to actually do it” is exactly who this system was designed for.

What Sets This Approach Apart

Several deliberate design choices separate this project from the more common note capture assistants already available.

The most fundamental one is that the chat interface is the entire interface. There is no browsing through Obsidian folders by hand, no dragging files between directories, and no manually maintained folder taxonomy to keep updated. A person simply talks, describing what happened, what needs remembering, or what needs to be found, and everything else happens automatically behind that conversation.

The system also genuinely operates across languages rather than defaulting to English with a translation layer bolted on top. Whatever language feels natural, whether Italian, French, German, Spanish, Japanese, or anything else, the agents respond in kind. There should be no requirement to think in a second language just to keep a personal knowledge system running.

Coordination is another defining feature. The agents are not a loose collection of isolated single purpose tools; they operate as a genuine crew, routed through a central dispatcher. When a transcription agent processes a meeting recording and discovers that it describes a brand new project, the dispatcher automatically chains in the agent responsible for vault structure to build the appropriate folders, without anyone needing to issue a second instruction.

Finally, the eight included agents are explicitly framed as a starting point rather than a fixed ceiling. Since no two lives look the same, the system supports building entirely custom agents through conversation alone. Saying something like “create a new agent” launches a guided design conversation, with no code to write, no configuration files to edit, and no templates to fill in manually. A person simply describes the problem, and the system builds an agent to solve it.

That flexibility becomes clear through a few illustrative examples. Someone struggling to track a monthly grocery budget can end up with a dedicated agent that monitors spending notes and flags whenever the limit is close. Someone whose partner jokes about a wardrobe chosen with eyes closed can get an agent that tracks owned clothing, suggests outfit combinations from existing notes, and gently discourages yet another repeat outfit. Someone who keeps buying duplicate items at a furniture store because inventory memory has failed can get an agent that catalogs household belongings room by room. Someone who repeatedly starts side projects and quietly abandons them can get a weekly check in agent that flags anything gone stale. Someone juggling several freelance clients can get an agent that aggregates deadlines across every one of them automatically.

Custom agents built this way still coordinate fully with the core crew, get discovered automatically by whichever agent platform is in use, and respond in whichever language feels natural, all while solving problems specific to one particular life rather than a generic template. It is worth noting explicitly that custom agents are created by the person using them and operate on that person’s own data; the project itself provides no warranty regarding their behavior, and responsibility for how a custom agent is built and used rests with its creator, consistent with the project’s published Terms of Use.

Meeting the Core Crew

Eight agents ship with the project by default, and each one has a clearly scoped role.

The Architect handles vault structure and setup. It designs the entire vault layout, runs the onboarding conversation for new users, and establishes the underlying rules that every other agent follows.

The Scribe focuses on text capture. Its particular strength lies in taking messy, typo filled, stream of consciousness input and transforming it into clean, well structured notes.

The Sorter is responsible for inbox triage. It empties the inbox on a regular basis and routes every captured note toward its proper home inside the vault.

The Seeker handles search and broader intelligence tasks. It can locate anything already stored in the vault and synthesize answers that draw across multiple notes, complete with citations back to the original source material.

The Connector builds and maintains the underlying knowledge graph, surfacing hidden relationships between notes that might never have been noticed manually.

The Librarian is responsible for ongoing vault maintenance, running regular health checks, removing duplicate content, repairing broken internal links, and tracking growth analytics over time.

The Transcriber converts audio recordings and raw meeting transcripts into rich, well structured meeting notes.

The Postman bridges email and calendar services, specifically Gmail or Hey.com alongside Google Calendar, with the vault itself, powering functions such as deadline tracking and meeting preparation.

Together, agents and skills form the complete system. Agents are optimized for quick, reactive, single shot tasks: capturing a note, running a search, creating a folder. More involved, multi step workflows that genuinely require back and forth conversation, such as full onboarding, email triage, or a complete vault audit, are instead handled by one of fourteen specialized skills.

Understanding the Skills Layer

Skills exist because some operations simply cannot be completed in a single exchange. They run inside the main conversation itself, which allows them to ask follow up questions, wait for responses, and maintain state naturally across an extended dialogue, something a quick reactive agent is not built to do.

The fourteen included skills cover a wide range of workflows. Onboarding walks through a full vault setup conversation from scratch. Creating an agent guides the step by step design of a brand new custom agent. Managing an agent allows existing custom agents to be edited, removed, or simply listed. Defragmentation runs a five phase weekly cleanup of the entire vault. Email triage scans and prioritizes unread messages. Meeting preparation produces a comprehensive briefing ahead of an upcoming meeting. A weekly agenda skill lays out a day by day overview of the week ahead. Deadline radar consolidates every known deadline into a single unified timeline. Transcription processes recordings directly into structured notes. A full vault audit runs a comprehensive seven phase review of the entire system. Deep cleaning offers an even more extensive cleanup pass. Tag garden analyzes and tidies up tag usage across the vault. Inbox triage processes and routes notes waiting to be filed. Contact synchronization keeps contact records aligned with Apple Contacts.

Choosing between a skill and an agent is not something the person using the system needs to think about. A central dispatcher checks incoming messages against the list of available skills first, since skills cover the most involved workflows. If a message clearly matches one of those complex conversational flows, the relevant skill takes over. If nothing matches, the dispatcher falls through to the appropriate reactive agent instead. The distinction happens entirely behind the scenes.

Each individual crew member operates as an isolated AI instance, complete with its own system prompt, its own restricted set of available tools, and its own assigned underlying model. Getting started involves cloning the project repository directly into an existing vault and running a setup script once; from that point forward, everything is managed purely through conversation, with no separate graphical interface, no manual dragging of files, and no hand maintained folder management required.

Getting the Crew Onto Different Platforms

The project is built to work across several different agentic platforms rather than locking a person into a single ecosystem. A single shared codebase gets translated and deployed to whichever platform is preferred, currently covering Claude Code in both its command line and desktop forms, Gemini CLI, OpenCode, and Codex CLI.

Installation on any of them follows the same general pattern, running a single installer script with a platform flag. For Claude Code, that looks like:

bash scripts/launchme.sh --platform claude-code

Gemini CLI, OpenCode, and Codex CLI each follow the identical pattern, simply swapping the platform name. If the flag gets left out entirely, the installer simply asks which platform to target interactively. Behind the scenes, each platform receives its own translated set of agents, skills, reference documentation, hooks, and any required MCP server connections, all handled automatically by the installer rather than requiring manual configuration for each target. Windows users planning to run Codex CLI specifically should note that native Windows support remains experimental, and running inside the Windows Subsystem for Linux is strongly recommended for a smoother experience.

How the Vault Gets Organized

Rather than inventing an entirely new organizational philosophy, the project relies on a hybrid structure combining the widely used PARA method with Zettelkasten style note linking. The default folder set includes an inbox for capturing everything before it gets sorted, a projects folder for active work with real deadlines, an areas folder for ongoing responsibilities without a fixed end date, a resources folder for reference material and how to guides, an archive for completed or historical content, a people folder acting as a lightweight personal relationship tracker, a meetings folder for timestamped notes, a daily folder for journals and daily notes, a maps of content folder for thematic indexes tying related notes together, a templates folder for reusable note formats, and a meta folder holding vault configuration, agent activity logs, and periodic health reports.

Importantly, these are simply the default folder names rather than a rigid requirement. Anyone who already maintains a vault using a different naming convention does not need to rename anything; during onboarding, the system scans the existing structure and adapts to it automatically, mapping the crew’s internal expectations onto whatever names are already in place.

Setting Everything Up From Scratch

Getting started requires an existing Obsidian installation, which is free, along with one of the supported agentic platforms already mentioned. The first step is simply creating or opening an Obsidian vault as usual.

From there, the project repository gets cloned directly inside that vault folder:

cd /path/to/your-vault
git clone https://github.com/gnekt/My-Brain-Is-Full-Crew.git

The installer script is then run from inside the cloned folder:

cd My-Brain-Is-Full-Crew
bash scripts/launchme.sh

Running the script without specifying a platform simply triggers an interactive prompt asking which one to target, after which the agents and skills get built and installed directly into the vault. Once the chosen agent platform is opened while pointed at that vault folder, the crew activates automatically. Open any other, unrelated project instead, and the agents remain dormant, since activation is tied specifically to the vault folder itself. For anyone who has never touched a command line terminal before, a dedicated step by step beginner guide walks through the entire process, and the whole setup is short enough that a tech savvy friend could complete it in roughly a minute if asked.

The final setup step is initialization. Opening the chosen agent platform inside the vault folder and simply saying “Initialize my vault” triggers the onboarding skill, which runs a friendly guided conversation covering three broad areas: who the person is, including name, preferred language, professional role, and what brought them to the system; what they actually need, meaning which of the available agents should be activated and which areas of life need managing; and which integrations, such as Gmail or Google Calendar, should be connected.

Once that conversation wraps up, the Architect generates a vault map file that records how the folder names in that particular vault correspond to the crew’s internal structure, builds out the full folder hierarchy if it does not already exist, saves a personal profile, and leaves behind a short welcome note before handing control over to everyday use.

What Daily Use Actually Looks Like

Once initialized, interacting with the crew becomes a matter of describing what is needed in plain language and letting the dispatcher route the request appropriately.

Saying something like “Save this: meeting with Marco about the Q3 budget, he wants the report by Friday” triggers the Scribe agent, which turns that raw statement into a clean note complete with extracted tasks, relevant internal links, and a recorded deadline. Asking to “triage my inbox” invokes the corresponding skill, which files every waiting note into its proper location, updates any relevant thematic index, and returns a short summary of what happened.

A question like “what did we decide about the pricing strategy” sends the Seeker agent searching across the vault, returning a synthesized answer supported by citations pointing back to the original notes. Saying “check my email” triggers the email triage skill, scanning Gmail, saving anything important as a note, and flagging upcoming deadlines. Requesting a “weekly review” runs a full vault audit, checking for broken links, duplicate content, and an overall health score.

Asking the system to “find connections for my latest note” activates the Connector agent, which surfaces relationships to other existing notes that might otherwise have gone unnoticed.

Working in Any Language, and From Any Device

Although the underlying system is built in English, every agent responds in whichever language a person actually writes in, whether that is Italian, French, Spanish, German, Portuguese, Japanese, or something else entirely. There are no separate language packs to install and nothing extra to configure; the matching happens automatically based on the language of each individual message.

For anyone using Claude Code specifically, the crew can also be reached remotely through its Remote Control feature. The underlying computer continues running Claude Code locally, with full access to the vault and every installed agent, while a phone acts purely as a remote interface through either a browser or the Claude mobile app. That makes it possible to capture a quick thought while out on a walk, check email from a couch, or search the vault while standing in a supermarket, with all of the actual processing still happening back on the original machine.

How the Agents Coordinate Behind the Scenes

Coordination between agents happens through a dispatcher driven orchestration system. Whenever an agent or skill finishes a task and detects that follow up work belongs to a different agent, it signals that intent through a specific section in its output, which the dispatcher reads and acts on automatically, chaining in the next appropriate agent without requiring a fresh instruction from the person using the system.

A few concrete examples illustrate how this plays out in practice. When the transcription skill processes a meeting recording that turns out to introduce a brand new project, the dispatcher automatically chains in the Architect to build the necessary folder structure. When the email triage skill identifies messages related to upcoming deadlines, the dispatcher chains in the Sorter to file them appropriately.

When the Connector discovers notes that appear disconnected from the rest of the vault, the dispatcher chains in the Librarian to investigate further. When the Sorter notices notes that seem to belong to an entirely new area of responsibility, the dispatcher chains in the Architect once again to build that area out properly. No single agent operates in true isolation, and the resulting crew ends up genuinely more capable than the sum of its individual parts.

Connecting Email, Calendar, and Contacts

The Postman agent, along with its related skills covering email triage, meeting preparation, weekly agendas, and deadline tracking, requires at least one supporting integration to function fully. Direct read and write access to Gmail and Google Calendar is available through a dedicated Google Workspace command line tool, supporting operations such as searching, reading, archiving, labeling, and sending email, alongside creating, updating, and deleting calendar events.

For anyone using a Hey.com account instead, a separate command line tool provides reading, replying, and composing capabilities while taking advantage of Hey’s own pre sorted mailbox categories, though calendar operations still rely on the Google Workspace tool regardless. As a lighter weight fallback, the installer also sets up MCP server connections automatically, though this option is limited to reading emails and calendar events along with creating draft messages, without full read and write access. Both the Google Workspace and Hey integrations can be used simultaneously for anyone maintaining accounts on both services at once.

An optional contact synchronization skill is also available for macOS users, keeping contact records aligned with Apple Contacts through a dedicated MCP server offering full create, read, update, and delete access to both contacts and contact groups. Once connected, that synchronization happens automatically whenever an email reply is sent, or on direct request. Every other agent and skill in the system, it is worth emphasizing, works perfectly well using nothing but a local Obsidian vault, with no external integrations required at all.

Keeping the System Updated

Staying current with new releases involves pulling the latest changes and running the included update script from inside the cloned project folder:

cd /path/to/your-vault/My-Brain-Is-Full-Crew
git pull
bash scripts/updateme.sh

The updater automatically detects which platform is currently installed by checking for its corresponding configuration folder, and if more than one platform happens to be installed at once, it simply asks which one should be updated, or a platform can be specified directly using the same flag pattern used during installation. Only files that have actually changed get touched during an update, and personal vault notes are never modified in the process.

Complementary Obsidian Plugins

While the crew itself handles the bulk of the organizational work, a handful of standard Obsidian plugins round out the experience nicely. Templater, Dataview, Calendar, and Tasks are considered essential companions. Beyond that core set, QuickAdd, Folder Notes, Tag Wrangler, Natural Language Dates, Periodic Notes, Omnisearch, and Linter are all recommended additions that pair naturally with the way the crew structures and maintains a vault over time.

Conclusion

My Brain Is Full, Crew reframes what an AI powered second brain can actually be. Rather than layering a smarter search bar on top of an already demanding note taking system, it hands the day to day operation of an entire Obsidian vault over to a coordinated team of specialized agents and guided conversational skills, all reachable through nothing more than plain, natural language, in whatever language feels most comfortable.

Its origin as a solution built out of genuine personal necessity, rather than a purely commercial roadmap, shows clearly in its design choices: a chat first interface with no manual file management, a dispatcher that lets agents chain into one another without extra prompting, and the ability to design entirely new custom agents through conversation alone.

The repository is available at: https://github.com/gnekt/My-Brain-Is-Full-Crew