Mads: Building a Personal AI Operating System

Most assistants answer one question at a time. Mads is designed to stay aware of the work around the conversation: the active project, pending tasks, earlier decisions, repository state, files, and preferred workflows.

The goal is simple. Instead of opening ten applications and rebuilding context in each one, a user should be able to say, “Continue working on Mads,” and receive a useful next action.

Knowledge work is distributed across GitHub, local files, Google Workspace, search engines, media services, maps, terminals, and multiple AI interfaces. Every switch loses time, attention, and context.

Conventional chat interfaces make the problem worse by starting each session from zero. They may understand the current prompt, but they do not understand the surrounding project or what should happen next.

Mads deliberately avoids keyword routing and multi-agent orchestration. Gemini receives natural-language tool descriptions, decides which capabilities are needed, executes independent calls concurrently when appropriate, and turns the results into one response.

Beneath the agent, an MCP manager resolves every requested tool to a provider exposing the same small interface. Whether a capability uses a stdio MCP server, a REST API, SQLite, or a native system command is invisible to the reasoning layer.

On startup, Mads walks upward from the current directory looking for project markers such as a Git repository, package manifest, or language configuration. Known workspaces activate silently; new ones can be registered with their name and GitHub remote detected automatically.

Once active, filesystem access, tasks, prompts, dashboard data, and planning are scoped to that project. Switching projects rebuilds the tool session around the new workspace.

The planner combines project awareness, task management, a reusable prompt library, and deterministic prioritization. It ranks work by overdue state, due date, priority, and recency, then surfaces a recommended next task without spending another model call.

A live dashboard joins this plan with the current branch, recent commits, documents, and project progress so recommendations reflect the actual state of development.

Imported conversation history moves through an incremental pipeline: read, deduplicate, gate, classify, compress, extract, score, and enqueue. The system identifies durable preferences, decisions, workflows, constraints, and identity facts without injecting raw history into every runtime conversation.

A content ledger makes repeated imports free, while atomic file writes prevent concurrent approvals from corrupting the profile.

Filesystem and document operations are restricted to configured roots. Destructive system actions require a second, confirmed tool call. OAuth tokens are encrypted at rest, and imported conversations are discarded after candidate extraction.

These controls remain ordinary parts of the tool contract rather than special cases hidden inside the agent loop, keeping the architecture understandable and testable.

Mads demonstrates how a personal assistant can move beyond isolated chat interactions. Project context, planning, memory, adaptive learning, and a broad tool ecosystem become parts of one persistent workflow.

The result is not another interface for generating answers. It is an evolving operating system for software engineering, research, automation, and daily knowledge work.