← newsletter

DevLog #17: Connect Your Provider on One Screen, Switch Models While You Chat, and Run Any Model ID

Connect Your Provider on One Screen

• Choose a provider and sign in without leaving the screen. Onboarding opens on a single “Connect your provider” screen with tabs for Claude, Codex, Cursor, and Antigravity. Claude and Codex connect today; Cursor and Antigravity show a “Coming soon” tooltip on hover. A provider you have already configured carries a check badge.
• Pick the sign-in method that matches how you buy your model. Each connectable provider offers three: a subscription sign-in (Codex device login or Claude setup token, recommended and preselected), an API key, and a credentials JSON file for people who already have one on disk.
• Connect both providers in one pass. The screen saves every provider you connected when you press Continue, so you can set up Claude and Codex together at the start.
• Sign in to Codex before your workspace exists. Codex subscription sign-in now runs as a device login from the login page, so you can approve it in a browser while your workspace is still provisioning and the credential is waiting when the workspace is ready. The flow expires after 15 minutes if it is left unapproved, and provider tokens are never returned to the browser.
• See why Continue is disabled. A token provider is required to run Nori, so Continue stays disabled until at least one is connected. Hovering it explains that: “Connect at least one token provider.”
• Land on a confirmation page after connecting Slack. Finishing the Slack install takes you to a page that shows the bot handle you got and a “Continue setup” link back into Nori. Cancelling the install now shows an actionable error page. Both paths used to end somewhere less useful, and one of them showed raw JSON.

Change Your Model, Skillset, and Provider While You Chat

• Switch your model from the browser. The web chat composer row carries Model, Skillset, and Provider buttons, each labeled with the value you are running (for example Model: Opus 4.8). Clicking one opens a list with the active option marked “Current.” Choosing a model from the UI is new; provider and skillset moved here from the settings drawer so all three live in one place.
• Confirm a provider switch before it happens. Choosing a different provider opens a “Switch provider?” dialog that tells you a fresh session starts with the new provider, your visible conversation carries over, and the current agent is released. Reselecting the value you are already on does nothing.
• Read the switcher menus. The popover text rendered black on the dark theme and was invisible until you hovered it, and the menu was clipped by the top of the composer bar. Both are fixed, and on phone-width screens the three buttons wrap onto their own row so their menus stay on screen.

Manage Billing and Ownership

• Open Billing as an org admin. Admins can now reach the Billing page and the Stripe customer portal, including updating a payment method and cancelling a subscription. This was owner-only, which meant routine billing work waited on one account. Transferring ownership still requires the owner.
• Transfer ownership from the Users page. Owners get a “Make Owner” button on any other member’s row, behind a “Transfer Ownership” confirmation. The previous owner becomes an admin, since an org has exactly one owner. There was no way to set an owner from the UI at all before this.

Find Out What Happened When Work Stops

• Read a real explanation when you hit your model provider’s usage limit. A session-limit or rate-limit error from Claude or Codex used to collapse into “Something went wrong, but your progress is safe.” It now says the limit is on your provider plan, points you at the error details for when access resets, and on Slack suggests switching providers with !provider.
• Get an email when provisioning halts. If repeated startup failures stop your fleet from bringing up new machines, the org owner now receives an email about it in addition to the dashboard notification. Previously that alert only appeared inside Nori, so nobody found out until they looked.

Fewer Interruptions

• Read a Slack or Discord reply that is just the answer. Posted replies no longer open with the model’s reasoning as a “Thinking:” blockquote, and an answer split across several chunks is rejoined without the blank lines that used to break a sentence in half. Reasoning is still on the live working card and available with !thoughts.
• Write through your MCP integrations. Nori now requests the scopes the authorization server advertises, so servers that default to a minimal grant issue a token with write access. Airtable was the case that surfaced it: you could read records and not create them. Reconnect an existing integration to pick up the new token.
• See one control card per Slack thread. When a thread’s agent provider resolved during startup, the thread got a second control card. Now the first card is posted once and updated in place.
• Get the Reconnect button when a web chat connection keeps dropping. A machine that accepted a socket and immediately dropped it pinned the retry at two seconds forever, so the retries never ran out and “Reconnect now” never appeared. Socket recovery and turn recovery are now tracked separately, the backoff finishes, and the button shows up.
• Tell which session an “End idle session” button will end. When a task is queued for capacity, Slack now lists each idle candidate with its title, how long it has been idle, who started it, and a link to it, with the End button beside that description. It used to be a row of bare buttons labeled only with a duration.

Sessions Ask Before Changing Production

• Get asked before an agent touches anything outward-facing. Sessions now stop and request explicit confirmation before landing code on main or another protected branch, creating or deleting live data, sending email, contacting a customer, or writing to a third-party or payments API. When a session cannot tell whether an action is production-facing, it treats it as production-facing and asks. An approval earlier in the conversation does not carry forward to the next action. This applies to sessions started after the next platform update; conversations already running are unaffected.

Run Any Model ID

• Type a model ID your agent’s picker does not list. /model now ends with a “Use custom model...” entry that opens a text field for any model ID, including dated snapshots. Setting a custom ID in [default_models] in your config also works now; those entries used to be dropped at session start when the agent had not advertised them.
• Have that model actually run. Nori passes the model to the coding agent when it starts the process (ANTHROPIC_MODEL for Claude Code, GEMINI_MODEL for Gemini, the model key in the Codex config), which is what makes out-of-catalog IDs work. Agents reject these over the live switch, so if you pick one from /model, Nori saves it as your default for that agent and restarts the session to apply it, telling you it did. The same model is reapplied when you resume.
• Set the channel for your own agent. Custom agents declare how they take a model in config.toml: model_override = { env = "MY_AGENT_MODEL" } or { arg = "--model" }. Nori also stopped printing a duplicate raw “Internal error” cell alongside the friendly /model message. One caveat: a model ID is not validated before you use it, so a typo surfaces at your first prompt and you fix it by picking another with /model.

Two Terminal Fixes for nori-handroll

• Open the command overlay with Ctrl-B on more terminals. Terminals that use the enhanced keyboard protocol send Ctrl-B as an escape sequence. Handroll only recognized the plain byte, so the overlay never opened on those terminals. It now recognizes both encodings, including when the sequence arrives split across two reads.
• Get a usable terminal back after you detach. Detaching from the overlay outside tmux could leave your terminal reporting every keystroke as an escape sequence. Handroll now resets the enhanced keyboard protocol on every teardown path, including detach, child exit, handoff, and signals.