Personal Model MCP
Local memory for Claude Code, Codex and Cursor Agent: learns from your activity on a Mac and gives agents context with source receipts over MCP
- 1,337 ★ on the repository
- Last commit
- Apache-2.0
Broad access
Why this level
- Continuously reads what is on the Mac screen: interface text, window titles and, when enabled, screenshots with text recognition
- Passes derived personal context to the chosen language model provider or through a Codex, Claude Code or Cursor Agent subscription
- Any connected MCP client gets access to a model with personal data, and the author asks you to connect only trusted clients
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add mcp/personal-modelDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Continuously reads what is on the Mac screen: interface text, window titles and, when enabled, screenshots with text recognition
- Passes derived personal context to the chosen language model provider or through a Codex, Claude Code or Cursor Agent subscription
- Any connected MCP client gets access to a model with personal data, and the author asks you to connect only trusted clients
This is third-party code. Review the repository files before installing.
What it does
Personal Model is a macOS program that continuously gathers the context of your work, builds a model from it and gives it to agents through an MCP server. It installs with `uv tool install personal-model`, and the package provides the `persome` command. The first run, `persome onboard`, explains each macOS request before it appears: Accessibility access goes to the `mac-ax-helper` helper, and Screen Recording is requested only when policy needs screenshots or text recognition. Full Disk Access is not needed. Captured data is kept in `~/.persome`. By default every five minutes activity is folded into layers: a point (a sourced observation), a line (a link or change over time), a face (a pattern), a volume (structure across projects) and a root (the current model of you). A text projection of the model lies in the `HUMAN.md` file. Semantic processing needs a language model; capture and keyword search work without one. Agents are connected with `persome install claude-code`, `persome install codex`, `persome install cursor-agent`, `persome install claude-desktop` and `persome install opencode`. Through MCP an agent calls `search`, opens a source with `read_receipt`, gets a model overview with `get_model_snapshot` and sees which records a conclusion rests on.
Who it is for. For Mac owners who work in Claude Code, Codex or Cursor Agent and want the agent to remember their projects, priorities and unfinished work between sessions. You must be ready to give the program access to what is on your screen.
Good fit when
- An agent in a new session does not know where you stopped: through `search` it finds recent project context and shows the records behind a proposed step
- You want one memory for several agents: Claude Code, Codex and Cursor Agent connect to the same model
- You want to verify conclusions: `read_receipt` and `resolve_evidence` show the path to a source, a time and a record number
- You have notes in Markdown, Obsidian or Notion and want the first model built from them: `persome import-data` reads them without writing
- An agent should propose postponed tasks in line with your priorities and leave actions outside the computer for your approval
Not a fit when
- You use Windows or Linux: the program runs on macOS 13 and newer and needs Xcode Command Line Tools
- Screen content must not go to a language model under employer or client rules: prompts of the semantic stages may include screen text, window titles and URLs
- You need shared team memory on a server: the author describes the project as an owner-local program, not a hosted multi-user service
- You need simple project memory without screen capture: `projectmem` fits, which keeps issues, attempts and decisions per project
Example request
Search Personal Model for where I left off in the payments project, show the sources and propose the next stepMCP
- Transport
- stdio, http
- Authentication
- not required
Environment variables
PERSOME_ROOT- Data directory instead of ~/.persome
Limitations
The project is in alpha and needs Python 3.12 or 3.13. It watches the screen: the `~/.persome` data directory holds interface text, window titles and, when enabled, screenshots and text recognition output. Capture is kept seven days by default, and accumulated data can be erased with `persome clean memory` and `persome clean all`.
The semantic stages send requests to the chosen model, and derived personal data goes into them: memory text, screen content, window titles and URLs. The project has no telemetry or update check. The mode where background stages use a Codex, Claude Code or Cursor Agent subscription is enabled separately with `--fund-model` and a daily cap, 50 calls by default.
A first useful result is expected in about ten minutes with working capture and a model, but the author gives no guarantee. Before stdio clients use a fresh data directory for the first time, run `persome start` once. License: Apache-2.0.
How to disable. Remove the `persome` entry from the client's MCP settings (in JSON it is `mcpServers.persome`) so the agent loses access to the model. Accumulated data is erased with `persome clean all`. The package is removed with `uv tool uninstall personal-model`, and the Accessibility and Screen Recording permissions are revoked in macOS system settings.
FAQ
How do I give Claude Code long-term memory of my work?
Install `uv tool install personal-model`, go through `persome onboard` and connect the agent with `persome install claude-code`. Claude Code then gets an MCP server started on demand over stdio, so a constantly running daemon is not required after the first setup.
Which macOS permissions does Personal Model ask for?
Accessibility for `mac-ax-helper` and, when event-driven capture is on, for `mac-ax-watcher`. Screen Recording is requested only for a policy with screenshots or text recognition. The project does not require Full Disk Access.
Where is Personal Model data stored and for how long?
In `~/.persome`, changeable with the `PERSOME_ROOT` variable. Capture is kept seven days and screenshots are removed sooner. The directory has mode 0700 and databases and logs mode 0600. A model export is redacted by default, while the `HUMAN.md` file is not redacted for sharing.
Does Personal Model need a paid language model?
Not for capture and keyword search. Semantic model building needs a provider set up with `persome llm setup`, or an existing agent subscription: `persome llm agent setup --client codex --daily-call-limit 50 --check`. `claude-code` and `cursor-agent` are supported too.
How do I load Obsidian or Notion notes into Personal Model?
During `persome onboard` a read-only import from local folders is offered, and Obsidian and Notion appear if they are found on the Mac. Later `persome import-data` does the same: the first model is built from existing Markdown history.
Which MCP tools does Personal Model provide?
Among them `search`, `read_receipt`, `resolve_evidence`, `get_model_snapshot` and `process_pending_model_work`. The last one hands a client that supports Sampling a batch of 1–10 sessions for semantic processing without passing on a sign-in token.
Discussion
No questions or reviews yet.
Related
MCP serversLocal memory for agents: stores conversations and project files verbatim and retrieves them with semantic search over MCP
Local persistent memory for coding agents: hooks record every session, and Claude Code, Codex or Cursor recall past decisions through 54 MCP tools
Self-hosted long-term memory for agents: semantic search, a knowledge graph and autonomous consolidation over MCP and a REST API
A local memory MCP server for coding agents: records issues, attempts, fixes and decisions and warns before repeating a failed approach
Long-term memory for agents: a remote MCP server, plugins for Claude Code, Cursor and Codex, and an API for your own apps
Shared local memory for AI agents: Claude Code, Codex, Cursor and others find the agreements of past sessions in it and record new ones
Comparison with similar tools
Entries that share the main search query, and how they differ.
| Entry | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| Personal Model (this entry) | no | not required | 1,337 | |
| projectmem | yes | not required | 869 | keeps issues, attempts and decisions per project and does not capture the screen |
| agent-memory | no | 4,459 | shared agent memory in Markdown files with a SQLite index, without capturing activity on a Mac |
README in short
The front page explains the idea of the `HUMAN.md` file, a table of model layers and a table of clients with install commands. Next come illustrated scenarios, a four-step install and verify section, sample agent prompts, updates and per-client recipes. The page ends with product boundaries, a privacy summary and links to the MCP, model format and security documents. The repository holds the core sources, a Mac helper, a companion client for iPhone, tests and benchmark notes.