craft-ai-agents/craft-agents-oss v0.15.0

on GitHub

Features

• A new look — The whole app moves to one design language: a navigator canvas with a joined paper card, rows on a shared layout grid, a composer that is one raised card with quiet toolbar chips, a calmer transcript (a quiet work log, plain answers, sunken code and diff wells, inline errors), and settings and detail pages on a single column. Dialogs, toasts and sheets float on their own material, empty states use a serif display heading, panels move with smooth motion, and every theme's colors are solved for contrast. The chat header now carries the session's status, labels and actions, and "+N" on a session row expands its labels in place. The WebUI gets a compact, touch-sized layout without zooming.

• Claude Sonnet 5.5 and Haiku 5.5 — Both are available on Anthropic and Amazon Bedrock connections, with 1M-token context. Haiku 5.5 now writes titles and summaries on Anthropic connections, at about a tenth of the cost of Haiku 4.5. Like Opus 5.5, both always think adaptively: "off" maps to low-effort thinking. GPT-6.1 Sol is listed for OpenAI API keys and ChatGPT accounts.

• Cloudflare AI Gateway — A new API key preset routes Claude, GPT and Workers AI models through your own Cloudflare gateway, with its logging, caching and rate limits. Paste an API token and your gateway URL; Craft Agents reads the account and gateway from it.

• Publish a page as a Claude artifact — With a Claude account connected, a page's header in the desktop app gets a Publish as Claude artifact button: one click puts the page, with its current data, on claude.ai as a private artifact in that account, and clicking again updates the same artifact. Source actions are disabled in the published copy.

• Background tasks you can follow and stop — The chips above the composer for background subagents, commands and workflows show what the task is doing right now, open its real output while it runs, and stop just that task.

• Run shell snippets in the background — Shell code blocks in a reply get a Run button next to Copy, and pasting a console snippet into the composer offers to run it. The agent starts it as a background task under your permission mode.

Improvements

• Claude Agent SDK 0.3.296 and Pi SDK 1.1.0 — The latest Claude Code runtime and Pi agent, with their fixes and model catalogs.

• Cheaper titles and summaries on more connections — ChatGPT accounts now use GPT-6 Luna for titles, summaries and call_llm instead of whichever model was listed last (most recently GPT-6.1 Sol, 20 times the price), and OpenAI keys prefer the Luna tier too. Google connections use Gemini Flash: every Gemini model used to count as "mini", so the first one listed, often 2.5 Pro, did these jobs.

• Failed tool calls show as failures in Pi sessions — A failing source, session tool, web search or web fetch reached the model and the transcript as a success, and ran PostToolUse instead of PostToolUseFailure automations. The model now gets an error result and the call is marked as failed.

• Suggestions in a new chat — An empty chat offers three starter prompts based on your connected sources, such as checking your latest emails or today's calendar, or ones that need no setup, like today's tech news or connecting your email. A click sends the prompt.

• Editable label values — Any label chip that shows a value can now be edited, and a chip in the session list edits only its own entry.

Bug Fixes

• Source tools work again on Claude — Since 0.14.1, every MCP and API source call (Slack, Linear, Notion, Gmail and others) failed with "_intent expected string, received undefined". Fixes #1086.

• Pi sessions no longer load extensions from the project or your Pi setup — A Pi session ran any .pi/extensions code in the working folder, and titles and call_llm loaded your global ~/.pi/agent extensions. Pi sessions now load only what Craft Agents provides.

• Background model calls no longer carry the full tool set — Titles, summaries and call_llm on Claude connections were sent with every Claude Code tool attached and, with the new runtime, would have picked up your own Claude Code permission defaults. They now run without tools.

• Labels stay clean — Auto-label rules stored the matched text as a value on plain labels and kept every spelling (bug, Bug, BUG), so label automations missed or fired repeatedly. Plain labels never carry values, duplicates collapse by meaning, and existing sessions are cleaned up when they load.

• Subagent steps grouped under their subagent — Steps from a background subagent started in an earlier turn were mixed into the next turn's list, each indented under an unrelated row. They now appear in their own group.

• Automation next run in the right timezone — The automations list computed the next run in your timezone instead of the automation's, so a 9:00 New York job viewed from Europe showed 3:00 AM.

• WebUI deep links — A link to a session no longer jumps to the first session after loading, ?ws= opens the right workspace, and the login page is styled before you sign in.

• A failed source test shows in the list — Sources whose test failed looked healthy in the sources list while their detail page showed the error.

• Composer fixes — Cancelling a credential prompt no longer submits the typed credential, and phones no longer rotate keyboard-shortcut tips in the placeholder.

• Windows — Windows 10 and Windows 11 before 22H2 get an opaque window instead of a transparent one, and live task output works when your user folder has a space in its name.

• Azure OpenAI after the Pi update — Azure connections keep working with Pi's renamed Azure provider, and an edited Azure endpoint takes effect without restarting the session.