Skip to content

Changelog

What shipped, release by release. Releases are tagged on GitHub.

v0.0.6

AddedChangedRemovedFixedSecurity

Aldwin opens in your browser, and long chats keep going

Aldwin opens in your browser. Run aldwin web and you get the same assistant as the terminal, with buttons and menus instead of commands. It still runs on your own computer. Long chats keep going too: when a chat gets long, Aldwin writes itself notes on the older part instead of forgetting it, and every message stays on your screen.

Added#

  • Web app: aldwin web serves Aldwin from your computer and opens it in your browser. It comes with the normal install, with nothing extra to set up. Start with the Quickstart or Install and open.
  • First run in the browser: connect a model with an API key, a free local model, or your own server, then tell Aldwin your name, work, and answer style. Each of those is optional. See First run.
  • Chat: answers stream as they are written, with highlighted code, tables, diagrams, and maths. Stop an answer, retry it, or continue one that was cut short. The message box switches model and search provider, and a ring shows how full the model's context is.
  • Clean and Detail views: Clean folds Aldwin's steps into one line such as "Worked for 12s". Detail shows every search, file read, and app action as it happens.
  • Conversations: a sidebar lists past conversations, with a See all page grouped by date. Search finds a conversation by its title or by something said in it. Rename, pin, or delete a conversation from its menu. See Conversations.
  • Approval cards: when Aldwin wants to run a command, change a file, or act in an app, a card shows exactly what it will do, with Allow, Deny with an optional reason, and sometimes Always allow in this chat. See Approvals.
  • Skills and apps in a message: type @ to add a skill and / to add a connected app, up to three of each per message.
  • Settings: seven sections for Models, Search, Skills, Tools, Advanced, Appearance, and About. Advanced shows each limit with its unit and default, and won't save a value outside its range. About shows the version you're running. Keys are checked before they are saved and never shown in full again. See Settings.
  • Tools page: connected apps on top, then a library of apps grouped by kind, with one search across both. Connecting an app happens in one window, from checking the server to signing in. See Connect apps.
  • A bigger app catalogue: 34 apps, built from the MCP Registry, including monday.com, GitLab, Vercel, Supabase, Stripe, Canva, Miro, Zapier, and Airtable alongside Notion, Linear, and GitHub.
  • Custom apps that need a sign-in: an MCP server added by its address now gets a browser sign-in when the server supports one, and a token box when it does not. The app keeps its own icon.
  • Light and dark themes, and layouts that work from phone width to a wide screen.
  • Long chats keep going: when a chat fills 60% of what the model can read, Aldwin makes room before your next message. It shortens long results from earlier steps, then writes itself notes on the older messages and keeps the recent ones word for word. Every message stays on screen and in search. See Long chats.
  • Make room now: open the ring in the web app and select Make room now, or type /compact in the terminal, to make room yourself.
  • Looking back: Aldwin can search the current chat for the exact words of an earlier message or result, even after it has made room.
  • When to make room: Summarise older messages at, under Settings, Advanced, sets how full a chat gets first, from 60% to 90%.
  • Plain skill descriptions: each skill has a short line saying what it does, shown in the @ menu, in Settings, and by /skills.
  • Protocol: requests to list, open, start, rename, pin, and delete conversations, a pinned field on conversations, typed displays on tool results, a compact request for making room, and REST routes under /v1 for the web app, including conversation search. All are additive to protocol v2.

Changed#

  • One Aldwin, two apps: the web app and the terminal share conversations, memory, skills, settings, and keys. A model or search provider picked in one is used by both.
  • One answer at a time per conversation: while Aldwin is answering in a conversation in one app, the other app waits for that conversation.
  • Settings apply without a restart: Aldwin rereads its settings before each answer, so a change made in the other app takes effect on the next message.
  • Approvals survive a dropped connection: a waiting approval can still be answered after the page reconnects.
  • Base instructions: Aldwin is told to finish the job, try another approach after a failure, respect a declined approval, and look for a connected app's tools first.
  • Steps per answer: the default is now 50 instead of 15. A number you already set is kept, and local models still stop at 15.
  • No more 20-message memory: Aldwin used to forget everything older than the last 20 exchanges (8 on local models). It now makes room instead.
  • Hosted models: OpenAI-compatible providers count as 128,000 tokens and Anthropic as 200,000 from the start, so Aldwin can make room before it has fetched the model list.
  • Local models: Aldwin measures room against the size it asks Ollama for, because Ollama cuts longer requests without saying so.
  • Docs: a new Get started section with a Quickstart, a Web app section, shared guides for memory and files, web steps in every model and search guide, and a Help section.

Removed#

  • The terminal's "Context nearly full" notice. Aldwin now makes room on its own.

Fixed#

  • Long chats lost messages: when a model refused a request as too long, Aldwin deleted the oldest exchange from the saved chat, so it vanished from the screen and from search. Nothing is deleted now.

Security#

  • The web app listens on this computer only. Addresses other than 127.0.0.1, localhost, and ::1 are refused.
  • Each run prints a new private link. Only a browser that opened that link can use Aldwin, and requests from other websites are refused.
  • Only one browser tab is connected at a time.
  • Pages load nothing from other websites. Fonts and logos ship with Aldwin.
  • A custom app's icon is fetched without your sign-in, only from the app's own site, and served so that it cannot run scripts.
  • Aldwin's notes on a long chat are written without tools, treated as background rather than instructions, and have keys and tokens removed. A file read only in the older part of a chat must be read again before Aldwin edits it.

Upgrade notes#

  • Upgrading from 0.0.5 keeps conversations, notes, memory, approvals, apps, persona, settings, and keys. No home migration runs.
  • Run uv tool upgrade aldwin, then aldwin web.
  • If you set Aldwin up with 0.0.5, your settings file keeps 15 steps per answer. To use the new default of 50, change Steps per answer under Settings, Advanced.
  • The terminal app works as before.

Known limitations#

  • The web app only opens on the computer running it. It cannot be reached from a phone or another computer.
  • Attaching files to a message, voice input, and editing a sent message are not available in the web app.
  • The workspace folder cannot be changed from the browser. Restart with aldwin web --workspace <folder>.
  • After a page reload, past tool steps show as plain rows without their detailed views.
  • Windows is not supported. Local MCP servers remain unsupported on Windows.
  • Aldwin makes room before each message, not in the middle of one long task.
  • A hosted model that holds fewer than 128,000 tokens can turn down a long chat before Aldwin makes room for it.

v0.0.5

AddedChangedSecurity

Aldwin can search the web with seven providers, including free self-hosted SearXNG

Aldwin can search the web with seven providers, including free self-hosted SearXNG. You connect them from /search without editing any files.

Added#

  • Seven search providers: web search works with free self-hosted SearXNG, or with a key from Tavily, Brave, Exa, Perplexity, Parallel or Firecrawl. Each has a setup guide under Search.
  • /search screen: pick a provider with the arrow keys and paste its key. Aldwin checks the key with the provider before saving it, and tells you why if it fails, such as a wrong key, no credits or a rate limit. Pick a connected provider to use it, update its key or disconnect it, or choose Automatic. Typed shortcuts are /search <name>, /search auto and /search status.
  • Free key checks: keys are checked against each provider's free account endpoint, so connecting never runs a paid search. Brave has no free check, so its check uses one Brave query. Exa uses one query only when its free check isn't available for your plan.
  • SearXNG found automatically: a SearXNG instance with JSON output on http://127.0.0.1:8080 or http://localhost:8080 is picked up with no setup. You can also set search.searxng_url in config.yaml or the SEARXNG_URL environment variable. Aldwin spaces out calls to a local SearXNG so its upstream engines don't block it.
  • Search settings: search.provider in config.yaml chooses a provider, or auto (the default) uses the first one set up, in this order: Tavily, SearXNG, Brave, Exa, Perplexity, Parallel, Firecrawl. limits.web_searches caps searches per message (default 5).
  • Search provider in the status line: the bottom toolbar shows which search provider is in use.
  • Manage model keys: Manage keys… in /model updates or removes a saved cloud model key. Aldwin won't remove the key for the model you are using, and explains how to remove a key that comes from an environment variable.
  • /copy: copies Aldwin's last answer to the clipboard as markdown, without the terminal's indentation and line wrapping. /copy code copies just the last code block. It uses OSC 52 where the terminal supports it, then pbcopy, wl-copy or xclip.
  • Protocol doors: search.list, search.connect, search.use, search.disconnect and provider.disconnect requests, a search.providers event, and a search_unavailable error code. All are additive to protocol v2.

Changed#

  • Keys take effect straight away: connecting or disconnecting a search key applies to the next search, with no restart.
  • OpenRouter keys are checked properly: OpenRouter's model list is public, so a wrong key used to pass the check. Aldwin now checks the key itself.
  • First run checks the Tavily key: a Tavily key pasted during setup is checked before it is saved, with three attempts. Skipping it still works.
  • Clearer "not set up" messages: when no search provider is set up, the model is told to ask you to run /search.
  • Docs: README, commands, tools and privacy docs cover /search and Manage keys…. New guides for every search provider and every model provider.

Security#

  • Search keys travel only in each provider's auth header, never in a URL or request body, and are removed from any error message Aldwin shows.
  • secrets.yaml is written atomically and stays readable only by you (mode 0600).
  • A private or localhost SearXNG address is allowed on purpose. Only the configured search address gets this exception; web_fetch still blocks private addresses.
  • /copy runs clipboard programs with a fixed argument list, a 5 second timeout and a minimal environment, so secrets in your environment aren't passed on.

Upgrade notes#

  • Upgrading from 0.0.4 keeps sessions, notes, memory index, approvals, connectors, persona, settings and keys. No home migration runs.
  • Tavily users keep working with no change: Tavily stays first in the automatic order.
  • Run uv tool upgrade aldwin.

Known limitations#

  • Search result snippets are cut to 200 characters before the model sees them.
  • Model key checks list your models, so an account with no credit can still connect. The first message then fails with a rate limit or HTTP 402.
  • DuckDuckGo was considered and is not included: it was too unreliable to ship.
  • Windows is not supported. Local MCP servers remain unsupported on Windows.

v0.0.4

AddedChangedRemovedSecurity

Aldwin can use any first-ship model provider, including local servers - no Op...

Aldwin can use any first-ship model provider, including local servers - no OpenAI lock-in.

Added#

  • Any model, including local: switch between cloud and local providers with /model. Built-in catalogue (bundled providers.yaml, no online lookup): Ollama, LM Studio, OpenAI, Anthropic, Google Gemini, OpenRouter, and DeepSeek.
  • Local-first first run: on launch Aldwin probes Ollama and LM Studio. Path A uses a found local model with no account; Path B offers cloud or install steps; Path C keeps an existing OpenAI setup with no reconfigure.
  • Three HTTP translators (no vendor SDKs): OpenAI-compatible chat, Anthropic Messages, and Ollama native - plain httpx only.
  • Key store: environment variables win and are never written to disk; pasted cloud keys are checked against the provider before save; secrets stay in ~/.aldwin/secrets.yaml (0600).
  • Friendly provider errors: agentling error kinds map to copy that names the provider and what to do next, with branded retry status lines.
  • Protocol doors: models.list, provider.key, and provider.server for listing models, saving keys, and registering Other server entries under providers: in config.yaml.
  • Other server: add any OpenAI-compatible (or Anthropic/Ollama-shaped) endpoint via /model or providers: (shape, local, base_url, secret, title).

Changed#

  • agentling 0.2: depends on agentling>=0.2,<0.3 from PyPI; the OpenAI Python SDK is not a dependency.
  • What's-new and tip: upgrade panel explains any model including local; tip says "Switch between cloud and local models any time with /model."
  • README and docs: "Run it fully local" quickstart, first-ship provider table, Fully local privacy, provider-neutral protocol error wording.

Removed#

  • OpenAI-only product wording from the README and release messaging (OpenAI remains a first-ship cloud option).

Security#

  • Provider keys follow env-wins and check-then-save; pasted keys are validated before persist.
  • Local model traffic on loopback does not leave the machine (except optional web tools, remote connectors, and the PyPI update check).
  • No openai / anthropic vendor SDK packages in the install (size and supply-chain).

Upgrade notes#

  • Upgrading from 0.0.3 keeps sessions, notes, memory index, approvals, connectors, persona, and settings. No home migration runs; bare model ids still resolve as openai/<id>.
  • Run uv tool upgrade aldwin. OpenAI users keep working (Path C).

Known limitations#

  • First-ship providers only in the built-in catalogue; other backends use Other server / providers:.
  • Windows is not supported. Local MCP servers remain unsupported on Windows.

Sizes#

  • Wheel and installed size: record after TestPyPI / CI (budget ≤120 MB installed, no extras).

v0.0.3

AddedFixedSecurity

Aldwin can now connect to your other apps through MCP servers, and warns you ...

Aldwin can now connect to your other apps through MCP servers, and warns you when it detects a secret in your message.

Added#

  • Connectors: connect Aldwin to Notion, Linear, GitHub, Context7, and other apps with /connectors. Each connector is an MCP server that adds tools to the current session.
    • Catalogue: pick a service from the built-in list and sign in through your browser; Aldwin handles the OAuth flow and stores the token at ~/.aldwin/credentials/<name>.json (mode 0600).
    • Custom URL: paste any remote MCP server URL that is not in the catalogue.
    • Local servers: run any MCP server from a shell command (for example npx -y @modelcontextprotocol/server-filesystem /path); the command is shown for approval before it starts, and re-approved after any hand edit. Local servers are not supported on Windows.
    • Import: aldwin connectors import reads connector configs from other agent apps on your machine and adds compatible entries.
    • Ask Aldwin to connect: type "connect me to Notion" and Aldwin proposes the right connector for you to approve.
    • On-demand tool loading: tools are loaded from the server only when the model needs them, so the context window is not pre-filled with hundreds of tool definitions. Up to 40 connector tools can be active at once; /new clears the loaded set.
    • Trust toggle: in /connectors, toggle trust on a server to control whether its read-only tools run on Auto or always ask first.
  • Pasted-secret warning: Aldwin detects patterns that look like API keys, tokens, or passwords in your message before it is sent to the model. It shows the detected pattern and asks you to confirm or cancel. If you confirm, the text is redacted from the session file and from logs/approvals.jsonl so it is not stored at rest.
  • mcp version in --version: aldwin --version now reports the installed mcp SDK version alongside agentling, for example aldwin 0.0.3 (protocol v2, agentling 0.1.1, mcp 2.3.0).
  • Connectors tip: the rotating welcome-card tip now includes "Try: /connectors to connect Notion, Linear or GitHub" (15 tips total).

Fixed#

  • The set_persona tool always told the model the persona limit was 2,000 characters, even when limits.persona_chars was set to something else. With a lower limit the model wrote text that was then rejected. The tool now states the configured limit.
  • /memory always said "Notes in the last 7 days" and counted over a fixed seven days, while the prompt uses limits.notes_days. The count and label now follow notes_days.
  • /tools labelled every tool that needs approval "(always asks y/n)". That was wrong for edit_file and write_file, which can be allowed per file for the session with a, and for the diff-approved tools, which don't ask y/n. Each tool now shows a note that matches its approval tier. run_command is described as running in the workspace, not the launch folder.
  • Setting theme: light or theme: colour_blind in ~/.aldwin/config.yaml had no effect: chat and aldwin setup always used the dark theme. The configured theme is now applied. theme: plain already worked and is unchanged.
  • The "Stopped after N steps" notice always said 15, even when limits.max_steps in ~/.aldwin/config.yaml was set to something else. It now shows the configured limit. DoneEvent carries the limit in a new optional max_steps field when reason is max_steps.
  • Some invalid config.yaml values produced garbled errors such as 'check_for_updates' must be Input should be a valid boolean…. check_for_updates and last_seen_version now have proper messages, and a bad item in run_command.env_allow is named by its position (run_command.env_allow[0]).
  • Running aldwin (or aldwin skills) from a folder that had been deleted after you cd-ed into it crashed with a FileNotFoundError traceback. Aldwin now falls back to ~/aldwin_workspace, the same as when it is launched from the home directory.

Security#

  • Aldwin warns and asks for confirmation when it detects a secret-looking pattern in your message before sending it to the model. Confirmed messages have the detected text redacted in the session file and in logs/approvals.jsonl.
  • Connector sign-in tokens are stored at ~/.aldwin/credentials/<name>.json (mode 0600) and never passed through the model or written to logs.
  • Local MCP server commands are shown for approval before they start and are re-approved after any hand edit. They run with a minimal environment (clean_env).
  • Connector tool output is wrapped in <untrusted> the same as file and command output, and is never trusted as instructions.
  • The ~/.aldwin/credentials/ path is blocked from every tool.

Upgrade notes#

  • Upgrading from 0.0.2 or 0.0.2.1 keeps all your sessions, notes, memory index, persona, and settings. No migration runs on first launch; just uv tool upgrade aldwin.
  • There is no connectors.yaml or credentials/ folder until you add a connector with /connectors.
  • The mcp package (~18-27 MB) is now a production dependency. Total installed size is higher than 0.0.2; see Sizes below.

Known limitations#

  • Local MCP servers run with your user's rights; there is no sandbox. A server can mark a write tool as readOnlyHint: true, and Aldwin will trust that hint.
  • Connector tools are loaded on demand and counted against limits.connector_tools (default 40). /new clears the loaded set if you need a fresh slate.
  • Only macOS 13+ and Linux are supported. Local MCP servers are not supported on Windows.

Sizes#

  • Wheel: ~0.34 MB. Installed (no extras): to be confirmed in CI (Linux x86_64, Python 3.11-3.13). macOS arm64: see CI run.

v0.0.2.1

ChangedFixedSecurity

Aldwin 0.0.2.1

Fixed#

  • Running aldwin from the home directory, or from any folder that cannot be a workspace (/, or inside ~/.aldwin), crashed with a traceback. Aldwin now creates and uses ~/aldwin_workspace for that session, shown in the welcome panel as workspace: ~/aldwin_workspace. Pass --workspace PATH or start Aldwin inside a project folder to choose a different root.

Changed#

  • The one-time What's new panel is keyed to the feature release (major.minor.micro), so a patch release such as this one does not show the 0.0.2 panel a second time.

Security#

  • Upgraded the pinned certifi to 2026.10.23. Earlier releases bundled emSign root CAs whose intermediate CA keys were compromised (GHSA-6q48-6c7w-9gw2); web_fetch and download_file verify TLS against this bundle.

v0.0.2

AddedChangedSecurity

Aldwin now works with your files and documents, knows good habits out of the ...

Aldwin now works with your files and documents, knows good habits out of the box, and lets you pick a skill with @.

Added#

  • File tools with diff approval: list_dir, glob, grep, read_file (read-only); edit_file and write_file, each shown as a folded diff before anything is written to disk.
  • Document reading: read_document converts PDF (page-range selection with pages=), Word, PowerPoint, Excel (sheet= by name or index), CSV/TSV, and HTML to markdown. download_file saves a public URL into the workspace after approval.
  • Planning tools: todo keeps a checklist for jobs of three or more steps, shown as a live view in the CLI; calculate does exact arithmetic and ISO date maths without a shell.
  • Upgraded memory: recall now does ranked search across notes and past conversations; notes can be corrected in place (replace=) or removed (forget, after diff approval) instead of only growing.
  • Nine packaged default skills (using-aldwin, files, markdown-notes, documents, making-documents, web-research, memory, terminal, skill-creator) loaded from the install via importlib.resources; user skills under ~/.aldwin/skills/<name>/ override by name.
  • @ skill picker: type @ to insert a skill into the turn; /skills and aldwin skills list name · source · description with default / user / user override labels, and /skills show <name> prints a skill's effective body.
  • create_skill tool (diff approval, never session-grantable) to write or update a user skill; the host rebuilds the agent after the turn and shows skills reloaded.
  • aldwin[docs] optional extra (python-docx, openpyxl, reportlab) for the making-documents skill to write Office files via run_command.
  • aldwin[office] optional extra (mammoth, python-pptx, openpyxl) for reading .docx, .pptx, and .xlsx. It is left out of the default install to keep the installed size within budget; without it, read_document says how to enable those formats. PDF, CSV/TSV, and HTML always work.
  • Protocol v2 (hello(v_min, v_max) negotiation): a v2 door gets Diff-tier approvals with digests; a v1 door still completes a turn with plain y/n approvals, and any Diff-tier call from a v1 door is refused as unapprovable rather than silently downgraded.
  • New slash commands: /todo (checklist view), /skills (catalogue and show), /verbose (toggle file-tool previews), /workspace (inspect or change the workspace root).
  • One-time What's new panel on the first everyday launch after an upgrade, and three new idle tips pointing at the file, document, and skill-picker workflows.
  • PyPI update-available nudge: a muted one-line notice under the welcome card when a newer aldwin is published, opt out with check_for_updates: false.

Changed#

  • Policy engine: every tool call now goes through a deterministic three-tier approval pipeline (session-grantable, per-call y/n, and Diff approval that can never be session-granted) with a hard floor that denies outright before any tier is even offered.
  • run_command runs with a clean, minimal environment. Secrets and API keys are no longer visible to spawned commands by default; extra variable names to pass through are opt-in via run_command.env_allow in config.yaml (secret-looking names are always ignored).
  • run_command always asks, with No as the default answer, and refuses outright (no prompt) a short list of privilege-escalation, interactive-program, and bare-REPL commands, and anything naming a blocked path (see Security).
  • Default skills are no longer copied into ~/.aldwin on first run; upgrades improve packaged skills automatically.
  • Skills never load code: a tools: frontmatter key is stripped (flagged as tools removed for user skills).

Security#

  • Hard floor, fail closed: a fixed set of paths that commonly hold secrets (.ssh/, .aws/, .gnupg/, .kube/, .docker/config.json, *.pem/*.key/*.p12/*.pfx, SSH private keys, .netrc, .pgpass, .npmrc, .pypirc, .git-credentials, .env*, and ~/.aldwin/ itself) is never readable, writable, or passed to run_command, with no config override (except read-only ~/.aldwin/tmp/** for read_file). This check runs before any approval tier and cannot be session-granted away.
  • Approvals bound to what was shown: a Diff-tier approval carries a sha256 digest over the tool name, resolved arguments, and the diff text; if the file or digest changed since the diff was shown, Aldwin re-confirms once rather than applying a stale approval.
  • Every tool result is wrapped as <untrusted> before it reaches the model, so text read from a file, page, or command output is never treated as an instruction.
  • Approvals log: every policy decision (approved, denied, blocked, refused) is appended as one redacted JSON line to ~/.aldwin/logs/approvals.jsonl, rotated at 10 MB.
  • Clean command environment: see run_command above.

Upgrade notes#

  • Upgrading from 0.0.1 runs an automatic, idempotent migration on first launch: config.yaml is backed up to config.yaml.bak-0.0.1 (once), gains run_command.env_allow: [] with every existing value and comment preserved, tmp/, state/, and logs/ are created or repaired to 0700, and memory/index.sqlite is built from existing notes and sessions. Nothing in ~/.aldwin is deleted. A second launch is a no-op.
  • A command that used to see $OPENAI_API_KEY (or any other secret-looking variable) under 0.0.1 no longer does; add its name to run_command.env_allow if a command genuinely needs a non-secret variable passed through.
  • An unmodified 0.0.1 starter copy of ~/.aldwin/skills/using-aldwin/ is renamed to using-aldwin.old-0.0.1/ so the new packaged identity block takes effect; an edited copy is kept as a user override.
  • The one-time What's new panel shows once on the first launch after upgrading, then never again for that version.

Known issues#

  • A workspace file's contents could, in principle, be read by Aldwin and included in a URL it is asked to fetch. Review run_command and web_fetch approvals for anything that forwards file contents externally.
  • A hard link inside the workspace pointing to a file outside it can resolve as if it were inside, letting a file read skip the outside-workspace approval step; its own filename is still checked against the blocked-path list.
  • run_command's refusal checks parse the command as text, not a real shell grammar; unusual quoting or encoding could in principle evade a check meant to catch it.

Sizes#

  • Wheel: ~0.23 MB. Installed (no extras): ~116 MB on macOS arm64; confirmed within the 120 MB budget on Linux x86_64 (Python 3.11, 3.12, and 3.13) in CI. With aldwin[office]: ~131 MB on macOS arm64.

v0.0.1

Added

First public release of Aldwin: an open-source personal assistant that runs o...

First public release of Aldwin: an open-source personal assistant that runs on your own computer, in your terminal, built on agentling.

Added#

  • One-command install: uv tool install aldwin installs the aldwin CLI (aldwin, aldwin setup, aldwin uninstall [--purge], --version, --plain, --debug).
  • Private local home folder (~/.aldwin): created on first run with mode 0700 (overridable via ALDWIN_HOME), atomic file writes (fsync + os.replace), rotating file logs (logs/aldwin.log, 1 MB × 3) with secret redaction and no message content logged, and separate secrets.yaml (0600) with environment-variable priority (OPENAI_API_KEY, TAVILY_API_KEY).
  • First-run onboarding & setup: guided terminal setup that prompts for an OpenAI API key with hidden input, verifies it against OpenAI, optionally configures a Tavily web search key, and saves a default chat model.
  • One brain, many doors architecture: typed Pydantic v2 protocol (aldwin.protocol), in-process asyncio.Queue host (aldwin.host), client SDK (aldwin.client), and CLI door (aldwin.cli) with strict architectural isolation (cli never imports core or agentling).
  • agentling + OpenAI core: streaming responses, automatic exponential-backoff retries on transient OpenAI errors, dynamic context manager keeping the last 20 exchanges on clean turn boundaries, and user-extensible Markdown skills in ~/.aldwin/skills/<name>/SKILL.md (shipping with the using-aldwin starter skill).
  • Five built-in tools:
    • web_search: web search via Tavily behind a swappable SearchProvider interface.
    • web_fetch: HTTP/HTTPS page extraction via httpx + trafilatura (capped at 20,000 chars, wrapped as <untrusted> data, with DNS/IP SSRF blocking against localhost and private networks).
    • run_command: shell command execution in the launch directory with mandatory y/n user approval, pre-execution refusal of sudo and interactive programs, secret environment stripping (*_KEY, *_TOKEN, *_SECRET), 60-second timeout, and 10,000-char output cap.
    • remember: appends a single-line dated note with metadata to ~/.aldwin/memory/notes/YYYY-MM-DD.md (last 7 days injected into context, up to 4,000 chars).
    • recall: keyword search across daily notes.
  • User-controlled memory and persona: ~/.aldwin/memory/profile.md (user-editable only, up to 4,000 chars) and ~/.aldwin/persona.md (up to 2,000 chars, updated via set_persona only after explicit y/n user approval).
  • Session persistence & recovery: conversations saved atomically to ~/.aldwin/sessions/<YYYY-MM-DD>_<id>.json after every step; reopening aldwin displays the last 20 messages and resumes where you left off, with an interrupted-turn notice if the previous terminal closed mid-task.
  • Inline terminal UI: built with Typer, Rich, and prompt_toolkit in native terminal scrollback, featuring fence-aware streaming Markdown, live spinner and status bar, tool cards, bordered approval prompts with terminal bell, Esc mid-turn stop within 1 second, ? keyboard shortcut guide, plain mode (--plain / NO_COLOR), and slash commands (/help, /new, /model, /persona, /memory, /skills, /tools, /exit).

Known limitations#

  • Terminal only, not always-on yet: Aldwin 0.0.1 runs inside your terminal process. Closing the terminal stops the active task (your conversation is saved after every step and resumes next time you run aldwin). Background daemon mode over ~/.aldwin/aldwin.sock arrives in 0.1.
  • OpenAI API key only: requires your own OpenAI API key (OPENAI_API_KEY or configured on first run). Other model providers and account OAuth are not included in 0.0.1.
  • Fixed built-in toolset: includes the five built-in tools (web_search, web_fetch, run_command, remember, recall); custom code tools cannot be added in 0.0.1 (drop-in Markdown skills in ~/.aldwin/skills/ are supported).
  • macOS and Linux only: requires Python 3.11+ on macOS or Linux. Windows is not supported yet.