Repository navigation
Aldwin opens in your browser, and long chats keep going: aldwin web serves the same assistant as the terminal from your own computer, and when a chat gets long, Aldwin writes itself notes on the older part instead of forgetting it.
Added
aldwin web: a loopback server that opens Aldwin in your browser. Listens on127.0.0.1port 7841 (or a free port), prints a one-time link, and takes--host,--port,--no-openand--workspace. The UI ships inside the wheel, so no Node is needed to run it.- Browser first run: connect a model with an API key, a local model, or another server, then an optional profile step (first name, occupation, answer styles).
- Chat in the browser: streamed answers with highlighted code, tables, Mermaid diagrams and maths; Stop, Retry and Continue; model and search pickers; a context ring; Clean and Detail views; light and dark themes.
- Conversations: a sidebar with search, Pinned and Recents, a See all page grouped by date, and rename, pin and delete from each conversation's menu. Search finds chats by the words in their messages as well as their titles, and shows the matching words.
- Approval cards: Allow, Always allow in this chat (where grantable) and Deny with an optional reason, shown on the step they belong to.
- Tool displays: typed displays for terminal output, diffs, file lists, search results, pages, tables, memory and denials.
- Skills and apps per message:
@picks up to three skills and/up to three connected apps in the browser's message box. - Settings in the browser: Models, Search, Skills, Tools, Advanced, Appearance and About. Advanced groups the limits with their units and defaults, checks each value against its allowed range, and keeps Save and Discard in a bar while there are edits; Reset to defaults puts every field back. About shows the running version and the update check.
- Tools page: connected apps first, then a library by category with one search across both, and a single dialog for connecting, signing in and retrying.
- Larger connector catalogue: 34 apps, built from the MCP Registry by
scripts/connectors/import_catalog.py, with categories. - Custom connector sign-in: a server added by URL that answers 401 gets browser sign-in when it publishes OAuth metadata, and a token prompt otherwise. Its own icon is fetched from its site and kept in
~/.aldwin/state/connector-icons/. - Long conversations: when a chat fills 60% of what the model can read, Aldwin makes room instead of forgetting. It shortens old tool results first, then writes itself notes on older messages with the model in use, and keeps the recent part word for word. After making room, the chat takes about 10% of what the model can read: the notes up to 4% (at most 4,000 tokens), the recent part the rest (at most 40,000 tokens). If the notes can't be written, it leaves the oldest messages out of the model's view and says so. Every message stays on screen, in
/historyand in search. - Make room now:
/compactin the terminal, and Make room now in the browser's context ring, do it on request. If the notes can't be written, it says so and changes nothing. recallin this chat:recalltakesscope="this_conversation"to find the exact words of earlier messages and tool results in the current chat.- When to make room:
limits.context_compact_percent(60 to 90, default 60), also under Settings, Advanced, as "Summarise older messages at". - Skill summaries: skills can carry a plain one-line
summary, which the@menu, Settings and/skillsshow instead of the description written for the model. The built-in skills all have one. - Protocol:
sessions.list,sessions.open,sessions.new,sessions.rename,sessions.pinandsessions.delete; apinnedfield on conversations;turn.appsandapps.used; typeddisplayon tool results; approvalcall_id;done.contextandhistory.context;compact; atopicon notices (roomwhen Aldwin made room); asummaryon skills; and REST routes under/v1, including conversation search andlimit_rulesin the settings. All additive to protocol v2.
Changed
- Shared home, two doors: the terminal and the browser run as separate processes on one home folder. A conversation is locked while a turn runs, so it answers in one place at a time.
- Settings reload before each turn, and writes to
config.yamlandsecrets.yamlare locked, so a change in one door applies in the other without a restart. - Approvals survive a dropped connection: a waiting approval can be answered after the browser reconnects; one asked while no page is attached is denied.
- Base instructions: finish the job, retry or change approach after a failure, respect a declined approval, and look up connector tools first.
- No 20-message memory: Aldwin no longer forgets everything older than the last 20 exchanges (8 on local models).
limits.context_exchangesis still accepted and ignored. - Steps per answer:
limits.max_stepsnow defaults to 50 instead of 15, with no upper limit. Aconfig.yamlthat already sets a value keeps it, and local models still stop at 15. - Hosted models: OpenAI-compatible providers (OpenAI, DeepSeek, OpenRouter) count as 128,000 tokens and Anthropic as 200,000 from the start of a run, so Aldwin can make room before the model list has been fetched.
- Local models: room is measured against the window Aldwin actually asks Ollama for, because Ollama cuts over-long requests silently.
- Docs: README, contributing, safety, privacy, commands, releasing, manual test and tests docs cover the browser.
SECURITY.mdnow says how to report a vulnerability.
Removed
- The terminal's "Context nearly full" notice: Aldwin now makes room on its own.
Fixed
- Long chats lost messages: when a provider refused a request as too long, Aldwin deleted the oldest exchange from the saved chat, so it vanished from the screen and from search, and the typed text of later messages shifted. Nothing is deleted now.
Security
aldwin webrefuses any non-loopback host. Every/v1route checks the session cookie,Host,OriginandSec-Fetch-Site.- The session cookie is HttpOnly and SameSite=Strict, and the token is removed from the address bar after sign-in.
- The WebSocket needs a single-use ticket that lasts 30 seconds, and only one browser tab is connected at a time.
- API keys saved from the browser are checked with the provider first and only ever returned masked.
- Connector icons are fetched without credentials, from the app's own host or parent domain only, and served with a sandboxing CSP.
- Aldwin's notes on a long chat are written with no tools, framed as background rather than instructions, and have secrets redacted. A file read only in the older part of a chat must be read again before Aldwin edits it.
Upgrade notes
- Upgrading from
0.0.5keeps sessions, notes, memory index, approvals, connectors, persona, keys, and settings. No home migration runs. - Run
uv tool upgrade aldwin, thenaldwin web. - Homes made by
0.0.5havemax_steps: 15inconfig.yaml, and it is kept. To use the new default of 50, change Steps per answer under Settings, Advanced.
Known limitations
- The browser interface only opens on the computer running
aldwin web. - No attachments, dictation, or editing of sent messages in the browser.
- The workspace cannot be changed from the browser; restart with
--workspace. - Tool displays are live only: after a reload, past tool rows are plain.
- 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 OpenAI-compatible model that holds fewer than 128,000 tokens can refuse a long chat before Aldwin makes room for it.