β οΈ Breaking change
The frontend no longer reads TRUST_X_FORWARDED_FOR (#16530, @Pwuts). A self-hosted frontend that set it by hand gets no client IP from X-Forwarded-For for the sign-in IP log, IP blocks and per-visitor store search until TRUSTED_PROXIES lists the proxy in front of it. The single-container image and Vercel need nothing.
What to do: if your frontend sets TRUST_X_FORWARDED_FOR, remove it and set TRUSTED_PROXIES to a comma-separated list of the addresses or CIDR ranges of the proxies in front of it. Until you do, the frontend logs a warning at start that the old setting is ignored. The same TRUSTED_PROXIES setting now also names the proxies for the backend and the External API's per-address limits, replacing FORWARDED_ALLOW_IPS and TRUSTED_PROXY_COUNT, which keep working but are deprecated; behind your own reverse proxy, the single-container image takes it too. See Client IP Addresses.
β¨ Highlights
- AutoGPT in your pocket. Our new native mobile app puts your team on your phone: sign in, chat with Otto and your experts, hire experts from the Marketplace, connect your services right from the chat, and talk instead of typing with live voice dictation. (#16715, #16869, #16923, #17058)
- Get started faster. Signing up takes fewer steps (no more typing your password twice), the sign-up and login pages show up right away, and setup no longer gets stuck on "Loading your setupβ¦". Sign up with Google and you can connect Gmail, Drive and Calendar in the same step. (#16829, #16831, #16968, #16845, #16877)
- Safer sign-ins. Adding a new way to sign in to your account (Google, Microsoft, GitHub or Discord) now takes a six-digit code from your inbox, unless that provider runs your mailbox (Gmail through Google, Outlook.com through Microsoft). You also get an email whenever a new sign-in method is added. (#16860)
- Never lose a message. If Otto can't confirm your message arrived, one click on Send again delivers it, files included, without sending it twice. And "Thinkingβ¦" now stays up until Otto starts answering, so you always know it's working. (#17041, #16962, #17038)
- Always see where your runs really are. After a brief connection drop, your Library, run history, Activity and chats catch up on their own, so a run that already finished no longer looks like it's still going. Live updates also keep flowing on slow connections. (#16898, #16915, #16932, #16903, #16865, #16960)
- Schedules you can count on. During brief maintenance your schedules show as temporarily unavailable instead of disappearing, a run that comes due during a restart still runs (exactly once), and retrying a schedule you just created no longer makes a duplicate. (#16937, #16866, #16919, #16904)
- More models to choose from. Gemini 4 Neon, GPT-6.2 Astra, GPT-6.2 Sol and Qwen 4 Max are now available in your workflows, and Otto's advanced Codex tier now runs on GPT-6.2 Astra. (#17023, #16870, #16899, #16836, #16852, #16902)
- More tools for Otto. Otto can now connect to AnySearch, Supabase, Klaviyo, AI/ML API and CompanyEnrich, reads from Supabase no longer need your approval each time, and Otto now finds the custom MCP servers you've connected by URL. (#16913, #16659, #16850, #15202, #16771, #16447, #15168)
π Security
This release fixes several privately reported security issues. We'll publish the advisories once the fixes are live in production.
π§© Also in this release
Otto and AutoPilot
- Experts hired from chat or the API introduce themselves (#16687, @Bentlybro)
- Measure first answer timing separately from model activity (#15533, @ntindle)
- Measure send-to-answer timing in chat (#15541, @ntindle)
- Let the greeting's Sonnet call think, so OpenRouter stops refusing it (#16781, @Pwuts)
- Reword the Founder plan's third task reason (#16851, @kcze)
- Tell Otto a refused tool call did not happen (#17036, @kcze)
- Judge a failed tool's error text like its output (#16884, @ntindle)
- bash_exec says when a command killed its own shell (#16347, @Pwuts)
- List nested block output objects by field name in describe_capability (#16679, @Bentlybro)
- Keep saved file names in the bot's reply text (#16680, @Bentlybro)
- Don't create a second schedule for work already scheduled (#16678, @Bentlybro)
- Consume an approval at most once (#16863, @ntindle)
- Refund a block run whose saved review can't be read (#16895, @ntindle)
- Back off sub-agent catch-ups while stale, and close two pub/sub gaps (#16888, @ntindle)
- Scope find_session and message_session to the calling chat's organization (#16966, @ntindle)
- find_session and message_session read sessions over RPC (#16858, @ntindle)
- Keep a chat's skill index for 15 minutes so follow-ups stop rebuilding it (#15985, @ntindle)
- Start a message's paywall and in-flight reads beside its session checks (#15779, @ntindle)
- Read an expert's row beside its owner's scope in the turn's check (#15854, @ntindle)
- Save a chat message without re-reading its session or its next sequence (#15486, @ntindle)
- Save a chat message and its session's fields in one transaction (#15987, @ntindle)
Builder and workflows
- Ctrl/Cmd+Z in a builder text field undoes the text, not the workflow (#17035, @kcze)
- Save the builder graph from one loaded snapshot after a save (#17021, @ntindle)
- A stored review decision wins over safe mode (#16874, @ntindle)
- Look up a node's waiting run with a correlated query, so a busy pin name can't time it out (#16835, @Pwuts)
- Scope four workflow and run routes to the request's organization (#16931, @Pwuts)
- Expand Unix cron day-of-week tokens to day sets before APS mapping (#15321, @ntindle)
- Bound custom cron intervals so APScheduler accepts them (#15322, @ntindle)
Integrations and blocks
- Group Parallel Search MCP with native Parallel provider (#16572, @Iitian001)
- Surface regex errors in Extract Text Information; forward falsy data in Match Text Pattern (#15313, @ntindle)
- Read Spreadsheet handles short, ragged, duplicate-header and blank-line CSVs (#15317, @ntindle)
- Short block input error messages (#17000, @kcze)
Platform and self-hosting
- Boxes behind the swap proxy trust its signing CA (#16627, @kcze)
- Refuse uploads clamd can't scan whole instead of scanning them in pieces; the bundled ClamAV's limits rise to 256M (#16694, @ntindle)
- Resume a streamed download where storage ended early, and stop gzip holding it back (#16838, @Pwuts)
- Harden credential masking in validation logs and error reports (#16956, @ntindle)
- Keep request credentials out of error reports and logs (#16908, @ntindle)
- Admin requeue publishes stuck QUEUED executions (#15328, @ntindle)
- Report memory graph truncation only when something was cut (#16762, @67-luck)
Other fixes
- Show a failed trial's Pay now button by reading its subscription with the card expanded (#16834, @Pwuts)
- Prevent translation extensions from breaking suggested tasks on Home (#16776, @sentry[bot])
- Stop the avatar image hang and drop networkidle waits from e2e (#17013, @ntindle)
Dependencies: 5 bumps (#16945, #16943, #16944, #16946, #16947).
CI, tests and developer tooling: 11 changes (#17012, #16799, #16832, #16918, #16842, #16924, #16914, #16930, #16876, #16859, #16928).
Documentation: 3 changes (#17002, #16892, #16519).
π Thanks to our contributors
@ntindle, @Bentlybro, @67-luck, @Pwuts, @kcze, @Lookoff-AIMLAPI, @Iitian001, @namirali, @ayush-singh-0601, @0ubbe
Full Changelog: autogpt-platform-beta-v0.11.3...autogpt-platform-beta-v0.12.0
π Feedback and issues
If you encounter any issues or have suggestions, please join our Discord and let us know!