v0.1.54 — More models, portable agents, screenshots in chat, and an end to the random sign-outs
Released 2026-06-10. Multi-platform signed/notarized DMG (Mac Apple Silicon + Intel) + EXE + AppImage. Auto-update from v0.1.53 happens on next launch.
The biggest model news since dual-lab routing: open-weight coders Kimi K2 and Qwen (via Fireworks) are now first-class everywhere a model runs — SMART-Code, Cmd+K inline edits, autocomplete, and the multi-model Studio. Alongside that, the Agent Foundry learned two tricks (drag an agent into another project, and an auto-generated run walkthrough), SMART-Code can now read screenshots and pick up where your other AI agents left off, there’s a light theme, and a cluster of reliability fixes — including the one a lot of you flagged: getting signed out every few hours is fixed.
Kimi K2 + Qwen coders, everywhere
Three open-weight coding models (served by Fireworks) are now selectable across the whole product:
- Kimi K2 — strong general coder
- Qwen3-Coder-480B — large, capable
- Qwen2.5-Coder-32B — fast and cheap
Where they show up:
| Surface | How |
|---|---|
| SMART-Code | ”Kimi” lab pill in the cloud picker — CodeNow-served and metered in credits (no key needed), or use your own key |
| Cmd+K inline edit | New inline model picker — click the model label to switch (your choice persists) |
| Autocomplete | Opt-in alternative to the default Haiku fast-lane |
| Studio (multi-model) | Run them side-by-side against Claude / GPT / Gemini |
Bring your own key anytime: set FIREWORKS_API_KEY in AI provider keys and every surface above uses it directly (BYOK is always free).
Agent Foundry — drag agents between projects, and an auto walkthrough
Drag-portable agents (P1). Built an agent in one project and want it in another? Drag its row onto the drop bar and CodeNow copies the agent’s manifest into the target project — no re-scaffolding, no copy-paste.
Auto walkthrough + summary card (P2). After a run, the Foundry now generates a ✨ Summary card — a plain-English walkthrough of what the agent actually did, digested from the latest recorded run. A fast way to see “what happened” without scrubbing the full replay.
SMART-Code can read screenshots
Paste or drop an image straight into the SMART-Code chat — a screenshot of an error, a design, a diagram — and the agent reads it as part of the turn (Haiku 4.5 + GPT-5-mini both see it). Dropped images become vision attachments instead of useless file-path chips.
SMART-Code picks up where your other agents left off
New cross-agent context continuity: SMART-Code now reads your accumulated project context — ~/.claude memory, recent transcripts, nested CLAUDE.md files, and the claude-mem ledger — so it continues the thread instead of starting cold. If you’ve been working with Claude Code, Codex, or Gemini on a project, SMART-Code arrives already knowing the history. Toggle it with the Continuity pill in the model picker (on by default).
A light theme
Prefer a bright editor? There’s now a white/light theme — switch it from the Theme picker in Settings.
Fixes that matter
| Fix | What changed |
|---|---|
| No more random sign-outs | Running the dev build and the installed app side-by-side (or just unlucky timing) could sign you out every few hours as the two raced on a rotating refresh token. The session now survives that race — you’re only sent to sign-in when the session is genuinely unrecoverable. |
| SMART-Code chat: copy/paste, scrollbar, drag | The transcript was rebuilt on every update, which wiped your text selection mid-copy and yanked you to the bottom while reading. It now re-renders without destroying selection or scroll, the scrollbar stays put, and drops are owned cleanly. |
| File drops land in the right project | A file dropped on a terminal no longer ends up in a different project’s terminal. |
| Screenshot 400s floored | A SMART-Code edge case that returned an “empty content blocks” API error is fixed. |
| In-app signup honors invites + promo codes | Admin invite links and promo codes now grant the right plan on direct signup. |
Upgrade notes
| Change | Who notices | What to expect |
|---|---|---|
| Kimi / Qwen models | Everyone | Needs the v0.1.54 build. Pick the Kimi lab pill (metered) or set FIREWORKS_API_KEY for BYOK. |
| Drag-portable Foundry agents + Summary card | Agent Foundry users | Needs the v0.1.54 build. |
| Vision / screenshots | SMART-Code users | Needs the v0.1.54 build. Paste or drop an image into the chat. |
| Cross-agent context | SMART-Code users | On by default; toggle with the Continuity pill. |
| Light theme | Anyone | Settings → Theme picker. |
| Sign-out / chat fixes | Everyone | Automatic with the v0.1.54 build. |
Auto-update from v0.1.53 happens on next launch. No migration steps.
Under the hood
The release pipeline now gates every build on the full test suite — smoke (240+ cases) plus both end-to-end journeys must pass on the tagged commit before any platform binary is built, signed, or published. A version can no longer ship red.
What’s next
- MCP HTTP/SSE import sources — stdio + CLI/stdin landed in v0.1.49; hosted transports are still queued.
- Stripe billing + tier enforcement in the desktop app.