Sentient: a personal assistant's memory and tasks over MCP
Sentient
A local assistant with an MCP bridge: Claude Code and Cursor search its memory, create tasks, send notifications and ask it questions
- 719 ★ on the repository
- Last commit
Broad access
Why this level
- Through `ask_sentient` an agent can drive an assistant that has access to mail, calendar, browser and code: each dangerous action needs your confirmation in the window
- Memory writes from an outside app could plant false facts in the assistant, so by default they wait for your review
- The build is alpha and runs from source, and the bridge access key sits in the system keychain
Install
- Works in
Manual install
git clone https://github.com/existence-master/Sentient.git && cd Sentient && uv venv .venv --python 3.12 && uv pip install --python .venv/bin/python -e ".[voice]" && cd desktop && npm install && npm run devRunning from source on macOS and Linux needs Python 3.12, uv and Node 22. On Windows the Python path in `uv pip install` differs: `.venv/Scripts/python.exe`. Then pick a model, for example `ollama pull qwen3:8b`.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Through `ask_sentient` an agent can drive an assistant that has access to mail, calendar, browser and code: each dangerous action needs your confirmation in the window
- Memory writes from an outside app could plant false facts in the assistant, so by default they wait for your review
- The build is alpha and runs from source, and the bridge access key sits in the system keychain
This is third-party code. Review the repository files before installing.
What it does
Sentient is a personal AI assistant that runs on your computer with no account: long tasks on a schedule and on events, memory of your preferences, voice, and pairing with a phone and smart glasses. For other agents it has a separate MCP server. It is off by default and turns on in settings with the "Let other apps use Sentient" switch. The `sentient mcp` command starts a thin process with no state of its own: every call is passed to the running Sentient engine over REST, so the safety rules stay in one place. Sign-in uses a separate key kept in the system keychain and replaceable in settings. There are eight tools: `memory_search`, `memory_add` and `memory_get` for memory, `task_create`, `task_list` and `task_status` for tasks, `notify_me` for a notification and `ask_sentient`, which sends the assistant a message as one chat turn. A memory write goes to a review list by default and is used only after you approve it. A message through `ask_sentient` counts as outside content: anything that sends, deletes or runs code is confirmed with you in the window and by notification. The app settings show a ready command for Claude Code and a JSON block for other clients.
Who it is for. For people who run the Sentient assistant on their own computer and want Claude Code or Cursor to use its memory and tasks.
Good fit when
- You work in Claude Code and want the agent to find your preferences and project context in Sentient's memory
- An agent finished long work and you want a notification on your computer through `notify_me`
- From Claude Code you need to give Sentient a scheduled background task or check the status of tasks already created
- You want facts the agent learned in a session to reach the assistant's review queue
- You want one shared assistant for several MCP clients that runs locally on your model
Not a fit when
- You use only cloud services and do not want to keep an assistant on your computer: the bridge works only next to a running Sentient
- You need a stable product for work: version 3 is alpha, and the author lists in a document what has been tested and what has not
- You need access from another computer or from the cloud: the bridge works only on this computer and accepts only a local address
- You plan to give an agent free access to mail and files through the assistant: `ask_sentient` still stops at confirmations, and they need you present
Example request
Search Sentient's memory for how I usually write commits, and when you finish the refactoring send me a notification through notify_meMCP
- Transport
- stdio
- Authentication
- API key
Limitations
Sentient version 3 is in alpha. The app runs from source on Windows, macOS and Linux and needs Python 3.12, uv and Node 22. The Windows installer builds from the repository with `npm run package` in the `desktop` folder, and the author promises prebuilt signed downloads with the first public release. The license is AGPL-3.0.
The MCP bridge is off by default, works only on this computer and refuses to connect to any address except a local one. Before every call it re-verifies the engine's identity with a one-time-number challenge, so swapping the port cannot capture the key. The access key opens only a small set of tools and gives no access to settings and approvals.
A memory write becomes active at once only with a separate switch in settings, and otherwise it waits for approval in the review section. The assistant itself needs a model: a free local one through Ollama such as `qwen3:8b`, or your own Anthropic, OpenAI or Gemini key.
How to disable. Turn off the "Let other apps use Sentient" switch in settings and delete the sentient entry from the client's MCP settings: for Claude Code that is `claude mcp remove sentient`. The key replacement button in the same settings invalidates the old key.
FAQ
How do I connect Sentient to Claude Code?
Turn on "Let other apps use Sentient" in the app settings, then run the command shown there, in the form `claude mcp add sentient -- sentient mcp`. For a source install the path to Python and the module are used in place of `sentient`: `-m sentient mcp`.
Which tools does the bridge provide?
Eight: `memory_search`, `memory_add`, `memory_get`, `task_create`, `task_list`, `task_status`, `notify_me` and `ask_sentient`. It is a small safe subset of the assistant's arsenal.
Do new memories become active at once?
No. By default a write waits in Sentient's review list and is used only after you approve it. To save at once, turn on a separate switch in the same settings.
What happens if an agent asks Sentient to send an email?
A message through `ask_sentient` counts as outside content. Anything that sends, deletes or runs code is confirmed with you in the Sentient window and by notification, and without your answer the action does not run.
Do I need an account or a cloud?
No, no account is needed and the data lives in one folder on your computer. On a free local model nothing leaves the machine, and with cloud models requests go to your provider under your key.
Discussion
No questions or reviews yet.
Related
MCP serversA macOS app that gathers your knowledge into a Markdown folder and optionally opens it to AI clients through an encrypted MCP mirror with two tools
A self-hosted task manager for human and agent collaboration through MCP
A local MCP server that lets the agent ask you questions in the console and notify you when a task is done
Cunzhi (寸止)
寸止 (Cunzhi)
MCP server in Rust and Tauri: the agent shows a dialog with options before ending a reply, keeps project rules and searches code
Minutes
minutes
A local-first MCP server for meeting memory: on-device recording and transcription into Markdown, searchable from Claude Code, Codex, Cursor and any MCP client
A local Markdown note base with an MCP server and skills: Claude Code, Cursor, Codex and Gemini CLI all read the same context
README in short
The project page describes the assistant by capability: long tasks, watchers, memory, voice, phones and glasses, and learning new skills with your approval. Then come running from source, picking a model and connections to Gmail, Calendar, GitHub, Slack and other services.