Aldwin
An open-source personal assistant that works with your files and documents, knows good habits out of the box, and runs on your own computer, in your terminal or your browser, built on Agentling.
Install
Install with uv (Python 3.11+ on macOS or Linux):
uv tool install aldwin
If you don't have uv installed yet:
curl -LsSf https://astral.sh/uv/install.sh | sh
Or install with pip:
pip install aldwin
Then start Aldwin in your terminal:
aldwin
Or open it in your browser:
aldwin web
Already have Aldwin? Upgrade in place and keep your conversations, memory, persona, and settings:
uv tool upgrade aldwin
First run
The first time you run aldwin, it creates a private home folder at ~/.aldwin and looks for a local model (Ollama or LM Studio). If one is running, you can chat with no account. Otherwise it offers a cloud provider (OpenAI, Anthropic, Gemini, OpenRouter, or DeepSeek) or shows how to install a local server. A Tavily key for web search is optional: Aldwin checks it with Tavily before saving it, and Enter skips it. You can connect any search provider later with /search.
If you already have an OpenAI key from an earlier install, Aldwin keeps it and does not force you to reconfigure.
After that, aldwin picks up your latest conversation where you left off. Type /help at the prompt to see what you can do. Switch models any time with /model.
The first aldwin web asks the same thing in the browser: connect a model with an API key, a local model, or another server, then optionally tell Aldwin your name, your work, and how you like answers.
Use it in your browser
aldwin web serves Aldwin from your own computer and opens it in your browser. It comes with the normal install; there is nothing else to set up.
Aldwin web is running at http://127.0.0.1:7841
Open this link to start: http://127.0.0.1:7841/?t=…
Press Ctrl-C to stop.
- Chat with streamed answers, and choose how much of Aldwin's work to see (Clean or Detail).
- Approve file edits, commands, and app actions from a card that shows exactly what will happen.
- Find, rename, pin, and delete past conversations in the sidebar.
- Settings for models, search, skills, connected apps, limits, appearance, and updates, with no files to edit.
The terminal and the browser share one home folder, so conversations, memory, skills, keys, and settings are the same in both. A conversation answers in one place at a time.
The server listens on this computer only, and each run prints a new one-time link. Leave the terminal open while you use it, and press Ctrl-C to stop. Options: --port, --no-open, --workspace, and --host (loopback addresses only).
The full guide is at askaldwin.com/docs/web.
Run it fully local
No cloud account required when Ollama or LM Studio is already serving a model on your machine.
-
Install Ollama (or LM Studio) and pull a tool-capable model, for example:
ollama pull qwen3:8b
-
Install Aldwin (
uv tool install aldwin) and runaldwin. -
When Aldwin finds the local model, pick it and start chatting. Your prompts stay on loopback (
127.0.0.1/localhost).
Built-in providers (bundled catalogue — no online lookup):
| Provider | Kind | Notes |
|---|---|---|
| Ollama | Local | Default http://localhost:11434 |
| LM Studio | Local | Default http://localhost:1234/v1 |
| OpenAI | Cloud | Needs an API key |
| Anthropic | Cloud | Needs an API key |
| Google Gemini | Cloud | Needs an API key |
| OpenRouter | Cloud | Needs an API key |
| DeepSeek | Cloud | Needs an API key |
Other server: use /model → Other server, or add an OpenAI-compatible endpoint under providers: in ~/.aldwin/config.yaml:
providers:
my-local:
title: My local server
shape: openai_chat
base_url: http://127.0.0.1:8080/v1
local: true
Supported shapes are openai_chat, anthropic, and ollama. Non-loopback http:// URLs require local: true (or use https://).
What Aldwin can do
- Files: list, find, search, and read files in your workspace; edit or create them after showing you a diff.
- Documents: read PDF, CSV, and HTML files as markdown, plus Word, PowerPoint, and Excel with the
aldwin[office]extra, and download files from the web. - Notes and memory: remember facts you tell it, search past notes and conversations, and keep a profile you write yourself.
- Persona: change how Aldwin talks by editing
persona.mdor asking it to. - Web: search the web with SearXNG or a hosted key (
/search), and read pages. - Terminal: run shell commands, always with your approval.
- Planning: keep a live checklist for longer jobs and do exact arithmetic and date maths.
- Skills: ten built-in skills, your own skills in
~/.aldwin/skills/, and an@picker to choose one for a message. - Models: switch between local and cloud providers with
/model; see Run it fully local. - Long chats: when a chat gets long, Aldwin writes itself notes on the older part and keeps going, and every message stays on screen.
/compact, or Make room now in the browser, does it on request. - Connectors: connect Notion, Linear, GitHub, and more with
/connectors, or from Settings in the browser; see the Connectors section below.
See the full tool list and the skills guide.
Connectors
Aldwin can connect to your other apps as MCP servers. Add one with /connectors in the terminal, or under Settings, Tools in the browser.
Catalogue: pick from 34 apps in the built-in catalogue, including Notion, Linear, GitHub, and Context7. Sign in through your browser and Aldwin stores the token in ~/.aldwin/credentials (mode 0600).
Any MCP server: paste a URL to add any remote server that isn't in the catalogue. If the server needs a sign-in, Aldwin offers a browser sign-in when the server supports one, and asks for a token when it doesn't.
Local servers: run a server from a command, for example:
npx -y @modelcontextprotocol/server-filesystem /path/to/folder
Local servers start only after you approve the command. They run with a minimal environment in your workspace folder. Local servers are not supported on Windows.
Import: aldwin connectors import brings in servers already configured in other agent apps.
Ask Aldwin to connect: type "connect me to Notion" and Aldwin proposes the connector for you to approve.
Safety
Aldwin runs on your machine without a sandbox, so every tool call goes through a policy engine first:
- Reading is free, changing asks. Read-only tools run straight away. Edits show a diff and default to No. Shell commands and downloads ask every time. In the browser, each request is a card with Allow and Deny, and nothing runs until you click one.
- Secrets are off limits. SSH keys, cloud credentials,
.envfiles and similar are blocked outright, whatever you or the model ask for. - Dangerous commands are refused.
sudo,rm -rf /, disk formatting and similar are refused before you're even asked. - Commands get a clean environment. API keys, tokens, and passwords in your shell are never passed to commands Aldwin runs.
- Everything read is untrusted. Text from files, pages, and command output is treated as data, never as instructions.
- Every decision is logged to
~/.aldwin/logs/approvals.jsonl, with secrets redacted. - Connector tools that only read run on Auto based on the server's own hint. Turn the trust toggle off in
/connectorsand every tool from that server will ask first. - Anything that changes something asks you first. Connector tools the server does not mark read-only always show an approval.
- All connector output is treated as untrusted. The same rule as for files and command output.
- Tokens stay local. Sign-in tokens are kept in
~/.aldwin/credentials(mode 0600) and never pass through the model. - Local servers need approval and run in a clean environment. If you edit the command by hand, Aldwin asks again before re-running it.
- Known limits: a server can mark a write tool read-only, and Aldwin will trust that hint. A local server runs with your user's rights; there is no sandbox.
Read the safety model for the full rules and known limits.
Data and privacy
- No telemetry. Aldwin collects no analytics.
- Sent to a cloud model provider: when you use a cloud model, your prompts, the conversation (or Aldwin's notes on its older part), your profile and recent notes, and anything Aldwin reads for the turn, including connector tool results, go to that provider. When a chat gets long, its older messages go to the provider once more so it can write those notes.
- Fully local: when you use Ollama, LM Studio, or another loopback model, conversation data stays on your machine (except optional web tools, remote connectors, and the PyPI update check).
- Web search: with SearXNG, the query goes to your instance (often localhost). With a hosted key, the query goes to that provider. Connect or switch with
/search. - Stays local: everything else in
~/.aldwin, including your API keys, saved conversations, sign-in tokens, and logs. Connector tokens never go to the model. - In the browser:
aldwin weblistens on this computer only, needs the one-time link it prints, and loads nothing from other sites. - Platforms: macOS 13+ and Linux. Local MCP servers are not supported on Windows.
Read data and privacy for the home folder layout and every network call Aldwin makes.
Commands
| Command | What it does |
|---|---|
aldwin |
Open the chat and resume the latest conversation |
aldwin web |
Open Aldwin in your browser, served from this computer only |
aldwin setup |
Set or update API keys and the default model |
/help |
Show slash commands and keyboard shortcuts inside the chat |
Esc |
Stop the current answer |
See all commands, slash commands, and shortcuts, including how to uninstall.
Web search
Web search works with self-hosted SearXNG or one hosted search key, all behind the same web_search tool. Type /search (or open Settings, Search in the browser), pick a provider, and paste its key: Aldwin checks the key with the provider before saving it, and tells you why if it doesn't work. The same screen switches provider, updates a key, or disconnects one. Step-by-step guides for every provider are at askaldwin.com/docs/search; for SearXNG see also Set up SearXNG.
| Provider | Cost | Key check |
|---|---|---|
| SearXNG | free, runs on your computer | found automatically on :8080, or enter its address |
| Tavily | free monthly credits, then paid | free; can take up to 20 seconds |
| Brave | paid plan | uses one Brave query |
| Exa | paid | free |
| Perplexity | paid | free |
| Parallel | paid | free |
| Firecrawl | paid plan with credits | free |
Keys are saved in ~/.aldwin/secrets.yaml (mode 0600). You can also set them in the environment (TAVILY_API_KEY, BRAVE_API_KEY or BRAVE_SEARCH_API_KEY, EXA_API_KEY, PERPLEXITY_API_KEY, PARALLEL_API_KEY, FIRECRAWL_API_KEY, SEARXNG_URL); an environment variable wins over a saved key. With search.provider: auto (the default), Aldwin uses the first connected provider in the order above, so a Tavily key wins over SearXNG.
SearXNG must expose JSON results (format=json / search.formats). Prefer SearXNG when you want queries to stay on a machine you control; a hosted provider receives the query under its own privacy terms. Brave's key check uses one of your plan's queries.
What's next
Aldwin runs while its terminal window is open, whether you started aldwin or aldwin web; it isn't always-on yet. Closing the terminal stops any active task, and your conversation resumes next time you start Aldwin. The browser interface only opens on the computer running it. Windows is not yet supported.
Contributing
Contributions are welcome. Read the contributing guide to get set up, and see the docs index for the protocol, test, and release docs. Working on the browser interface? Start with the web UI guide.
Have a question or an idea? Start a discussion. Found a bug? Open an issue.
Licence
Metadata
Release files for aldwin 0.0.6
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| aldwin-0.0.6.tar.gz | 3.3 MB | Without provenance | Details for aldwin-0.0.6.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| aldwin-0.0.6-py3-none-any.whl | Python 3 | none | any | Details for aldwin-0.0.6-py3-none-any.whl |
Total release size: 6.8 MB
Release files / aldwin-0.0.6.tar.gz
| Download URL | aldwin-0.0.6.tar.gz |
|---|---|
| Size | 3.3 MB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
7ce3ebbe527adde8ae373d7348208b86614bdc8ece655979630f52adb5cc4197
|
|
BLAKE2b-256 checksum How to use checksums |
14150506b0f1b26ba39958ab0db75a7fbbe9b85495d513ee0b1d3fdbf3c19719
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
uv/0.13.8 {"installer":{"name":"uv","version":"0.13.8","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"Ubuntu","version":"26.04","id":"resolute","libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":true}
|
Release files / aldwin-0.0.6-py3-none-any.whl
| Download URL | aldwin-0.0.6-py3-none-any.whl |
|---|---|
| Size | 3.5 MB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
a08a25e0d27852443922de1dd02ce217f4b9ebab12fc0be6195c7fca58371f49
|
|
BLAKE2b-256 checksum How to use checksums |
bdd076d1b45a131dcaca8ea772043f42484044c6797fe908c4bb323eebe3d1fa
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
uv/0.13.8 {"installer":{"name":"uv","version":"0.13.8","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"Ubuntu","version":"26.04","id":"resolute","libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":true}
|
Security for aldwin 0.0.6
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.