首次启动需等待几分钟才能访问 localhost:3000；自托管需自备模型凭证并自行管理更新运维；autogpt_platform 目录采用 Polyform Shield 许可，不能作为竞品托管服务出售。
AutoGPT
Significant-Gravitas活跃维护让大模型自主拆解目标、调用工具并连续执行任务的智能体平台，提供可视化搭建与运行环境，帮你把重复性工作交给 AI 自动完成。
AutoGPT is the vision of accessible AI for everyone, to use and to build on. Our mission is to provide the tools, so that you can focus on what matters.
autogpt-platform-beta-v0.11.1发布于 2026-11-06 03:36 UTC187,977star +44 / 7 天45,881fork同步于 2026-11-10 11:17 UTC
docker run -d --name autogpt --restart unless-stopped --shm-size 2g --ulimit nofile=65536:65536 -p 127.0.0.1:3000:3000 -e AUTOGPT_PUBLIC_URL=http://localhost:3000 -v autogpt-data:/data significantgravitas/autogpt:latest
docker exec autogpt autogpt-admin promote [email protected] 硬件要求按 README 说明
README 未说明具体显存或 GPU 型号要求；自托管以 Docker 单容器方式运行，模型可使用自带凭证或本地模型。
有 Docker 镜像
把内容、销售、研究等 AI 专家组成团队，由 Otto 统一协调分工
让大模型自主拆解目标、调用工具并连续执行任务的智能体平台，提供可视化搭建与运行环境，帮你把重复性工作交给 AI 自动完成。
核心能力
- 由 Otto 统筹，把任务分派给各领域 AI 专家
- 用自然语言给专家下指令，并配置上下文与技能
- 可视化工作流构建器，支持分支逻辑与工具连接
- 提供 API 与 MCP server，可从自有代码调用工作流
- 支持自托管，模型凭证由自己提供或接本地模型
容易踩的坑
内容由模型根据 README 生成，可能不完整或不准确，以官方文档为准。看原文 README
- autogpt-platform-beta-v0.11.1🚀 AutoGPT Platform v0.11.1最新正式版2026-11-06 · 2026-11-06 03:36 UTC
🔥 What's New?
⚠️ Breaking Changes
- #16416 - Self-hosted installs:
.env.defaultno longer ships a Web Push (VAPID) key pair, and the backend no longer sends push notifications with the old shipped pair. Runmake init-envto generate a pair for your install; the single-container image generates its own on first boot. Replacing the pair invalidates existing push subscriptions.NEXT_PUBLIC_VAPID_PUBLIC_KEYcan stay blank, and the frontend then uses the backend's key (by @ntindle)
New Features
- #11220 - Support input nodes alongside trigger nodes (by @Pwuts)
- #14991 - Run skills as /commands with arguments in the copilot chat (by @ntindle)
- #16582 - Make background subtasks real work that returns to the chat (by @ntindle)
- #16506 - Warn canceled trials 3 days and 1 day before access ends (by @Torantulino)
- #16622 - Make the email-change approval email clear about both addresses (by @Torantulino)
- #15488 - Redesign the k6 load tests around real user workloads (by @ntindle)
- #16605 - Add Tencent Hy4 to the LLM catalog (by @ntindle)
New Blocks
- #16561 - Add Firecrawl Alexandria data tool blocks (by @Torantulino)
- #16395 - Bring the Exa integration up to date with Exa's current API (by @Torantulino)
- #16393 - Add Treg blocks for people, company and web data (by @Torantulino)
Enhancements
- #16446 - Creating a GitHub repository asks first in Auto mode (by @Pwuts)
- #14377 - Drop redundant schema titles and window every digest node (by @Pwuts)
- #12017 - Correct uploaded image types without guessing containers (by @Otto-AGPT)
UI/UX Improvements
- #16612 - Remove the Tailwind palette; roles are the only colours (SECRT-3097 C4) (by @Abhi1992002)
- #16620 - Left-align the bullet points on the Meet Otto screen (SECRT-3380) (by @Torantulino)
- #16623 - Polish the first onboarding screen (SECRT-3382) (by @Torantulino)
- #16584 - App links in chat open in a new tab on middle-click; external links always ask first (by @kcze)
- #16617 - The run card in the chat follows the run until it ends (by @kcze)
- #16587 - The chat run card shows when a run fails or completes (by @kcze)
- #16640 - The side panel follows a file Otto overwrites (by @kcze)
Bug Fixes
- #16331 - A failed database connect at start-up is retried, not reported as connected (by @Pwuts)
- #16343 - A run's output no longer deletes an earlier output with the same file name (by @Pwuts)
- #16475 - Bound the cost of user regex patterns and Sentry event scrubbing (by @Pwuts)
- #16559 - Keep shell text like "$(…)" and "$HOME" out of chat math (by @Pwuts)
- #16586 - Otto names a failure's cause only when the error states it (by @kcze)
- #16392 - Numbers too large for the database are refused with a 4xx before they reach it (by @Pwuts)
- #16585 - Normalise chat input before it is stored or queued (by @ntindle)
- #16613 - Trim whitespace around prompt-value keys (by @kcze)
- #16614 - Count only listed workflows on the creator card (by @kcze)
- #16449 - The PostHog consent test stops each page's capture interval (by @ntindle)
- #16356 - Otto's workflow validator accepts links into list items (by @Pwuts)
- #16361 - Stop context_test leaving an executing session behind (by @Pwuts)
- #16566 - Give the shared auth mock the session exports actions.ts calls (by @Pwuts)
- #16276 - Values linked into MCP tool arguments reach the tool (by @Pwuts)
- #16357 - Turn numpy scalars in block outputs into plain Python values (by @Pwuts)
- #16426 - Chat messages keep text that looks like an HTML tag (by @Pwuts)
- #16367 - Keep interval system jobs' next run across scheduler restarts (by @Pwuts)
- #16565 - Give the external API concurrency test room for a slow Redis (by @Pwuts)
- #16364 - Send a changed email address its own verification email (by @Torantulino)
- #16333 - read_workspace_file retu
- #16416 - Self-hosted installs:
- autogpt-platform-beta-v0.11.0🚀 AutoGPT Platform v0.11.02026-11-04 · 2026-11-04 05:48 UTC
查看更新说明
🔥 What's New?
⚠️ Breaking Changes
- #16459 - Payment enforcement is now the backend setting
ENABLE_PLATFORM_PAYMENT(default off) instead of theenable-platform-paymentfeature flag. A self-hosted install that charges must set it;FORCE_FLAG_ENABLE_PLATFORM_PAYMENTandNEXT_PUBLIC_FORCE_FLAG_ENABLE_PLATFORM_PAYMENTno longer do anything (by @Pwuts) - #16474 - The platform API answers
413to a request body overMAX_REQUEST_BODY_MB(default 10 MiB) on any route that is not an upload route or webhook ingress. Raise it, or configure media storage, if you pass larger files inline (by @Pwuts) - #16522 - Behind a proxy chain whose
X-Forwarded-Forcarries more than one address, list your proxies in the newTRUSTED_PROXIESsetting (IP addresses or CIDR ranges). Left unset, per-visitor limits get no client IP instead of the header's leftmost entry. An entry that is not an IP address or CIDR range stops the frontend from starting (by @Pwuts) - #15853 - LaunchDarkly support is removed and every feature flag is read from PostHog.
LAUNCH_DARKLY_SDK_KEY,FEATURE_FLAG_BACKEND,NEXT_PUBLIC_LAUNCHDARKLY_ENABLED,NEXT_PUBLIC_LAUNCHDARKLY_CLIENT_IDandNEXT_PUBLIC_FEATURE_FLAG_BACKENDare ignored; without PostHog every flag is its default (by @kcze) - #16026 - AutoPilot conversations are no longer traced to Langfuse, which is now used only for prompt management. AutoPilot traces go to PostHog (by @kcze)
New Features
- #16310 - Add Parallel Decisions blocks (web-grounded yes/no, choice, score) (by @Torantulino)
- #16326 - Show real changelog entries in the login What's new panel (by @Torantulino)
- #16316 - Show popular services first in the Connect a service dialog (by @Torantulino)
- #16306 - Offer Connect an integration from the @ mention list (by @Torantulino)
- #16279 - Redesign the "Confirm your email address" email (by @Torantulino)
- #16268 - One Microsoft connection for mail, calendar, files, Teams, Copilot and Work IQ (by @ntindle)
- #15087 - Add Gemini 3.5 Flash, 3.5 Flash Lite and 3.1 Flash Lite to the LLM catalog (by @Bentlybro)
- #16407 - Link feedback to the person's PostHog replay (by @kcze)
- #16398 - Keep the onboarding role when its step is answered, not at the end (by @kcze)
- #16527 - Ask self-hosted users for a GitHub star after five answered prompts (by @ntindle)
- #14545 - Rate limit the v1 external API (by @Jia-ben00)
- #16459 - Payment enforcement is the ENABLE_PLATFORM_PAYMENT setting (by @Pwuts)
- #16522 - Resolve the client IP through operator-listed trusted proxies (by @Pwuts)
- #16554 - Seed previews with a small synthetic marketplace (by @ntindle)
- #16569 - Route load-test accounts to the dummy engine or a scripted model (by @ntindle)
- #15733 - Wake Otto after opt-in background subtasks (by @ntindle)
Enhancements
- #16352 - Compact spend approval card with a positive tone (by @0ubbe)
- #16280 - Clarify app connections in Otto (by @0ubbe)
- #16314 - Sign-up panel describes a team of AI experts, not agent building (by @Torantulino)
- #16394 - Design-system guardrails as lint rules with a per-file baseline (SECRT-3097 wave 0) (by @Abhi1992002)
- #13676 - Skip the truncate() probe grid when the payload already fits (by @dexhunter)
- #16225 - Say "Done" at the end of the latest turn (by @kcze)
- #16226 - Plain-language compaction copy and a clock on long-running steps (by @kcze)
- #16227 - Turn notifications on from the chat banner in one click (by @kcze)
- #16228 - Thumbs up/down work on a reply that just finished (by @kcze)
- #16250 - Give a Discord channel mention the recent channel messages as context (by @Bentlybro)
UI/UX Improvements
- #16401 - Design tokens, block-menu chips, LegacyButton refactor (by @Abhi1992002)
- #16404 - Design tokens, Text and type scale; zero strict violations in copilot (by @Abhi1992002)
- #16399 - Shadc
- #16459 - Payment enforcement is now the backend setting
- autogpt-platform-beta-v0.10.0🚀 AutoGPT Platform v0.10.02026-10-27 · 2026-10-27 16:25 UTC
查看更新说明
🔥 What's New?
⚠️ Breaking Changes
- #15970 - Self-hosted Docker Compose generates its own service passwords per install and won't start without them: run
make init-env(or the installer) beforedocker compose up. PostgreSQL, RabbitMQ, Redis, FalkorDB and ClamAV now listen on127.0.0.1only. Existing installs: see Upgrading: service passwords are generated per install (by @ntindle) - #12206 - External API v1 is deprecated in favour of the new v2, and keeps working until 2027-01-31 (by @Pwuts)
New Features
- #15990 - Record when each API key was last used (by @ntindle)
- #15862 - Tell chat-platform bots where each message came from (by @Bentlybro)
- #16023 - Role and text design tokens at today's values (DS1) (by @Abhi1992002)
- #16035 - Send product mail only to subscribers, trialists and accounts that lapsed in the last 7 days (by @Torantulino)
- #16037 - Molecules read only role tokens, same look (DS2b) (by @Abhi1992002)
- #16044 - Move contextual and renderers onto the design tokens (DS5c) (by @Abhi1992002)
- #16045 - Atoms read only design tokens, same look (DS2a) (by @Abhi1992002)
- #16078 - Move library, artifacts and (platform)/components onto design tokens (DS5e) (by @Abhi1992002)
- #16081 - Data-slot on every element an atom renders (DS7-1) (by @Abhi1992002)
- #16084 - Data-slot on every element the molecules render (DS7-2) (by @Abhi1992002)
- #15549 - The read-injection check gets its own flag and a shadow mode that judges every read and holds nothing (by @Pwuts)
- #16105 - Slot-object AttentionCard as the card and notice composition reference (DS8) (by @Abhi1992002)
- #16072 - The paywall tells users their scheduled workflows are paused (by @Pwuts)
- #16096 - Add Kimi K3.1 Pro model (by @ntindle)
- #16156 - Run Tailwind CSS 4 with Tailwind 3's theme and output (SECRT-3102) (by @Abhi1992002)
- #16176 - Stop the runs a removed member has going in the org (by @ntindle)
- #12206 - External API v2 (by @Pwuts)
- #15970 - Per-install service passwords and localhost-only data ports for Docker Compose (by @ntindle)
- #15941 - Replace the pixel-grid thinking loader with a drawn line (by @0ubbe)
- #15939 - Show places as an interactive Google Map in chat (by @0ubbe)
- #15863 - Pick an expert and fast or smart mode from the chat bots (by @Bentlybro)
- #16036 - Redesign the onboarding paywall as one calm trial welcome (by @Torantulino)
- #16242 - Read HTTP response bodies under a size limit (by @ntindle)
- #16241 - Cap the field sizes a template can format (by @ntindle)
- #16240 - Run every user-supplied pattern under a time limit (by @ntindle)
- #16217 - Add Cloudflare to the MCP catalog (by @ntindle)
- #16261 - Add GPT-6.1 Terra model (by @ntindle)
- #16221 - Tag every Otto model call with what started the turn (by @kcze)
- #16236 - Polish onboarding and Otto greeting cards (by @0ubbe)
- #15952 - Warn before sending something that looks like a secret in chat (by @kcze)
- #16094 - Pick Airtable bases from a card, preview table changes (by @Abhi1992002)
- #16095 - Explain what pays for what in the wallet and chat (by @Abhi1992002)
- #16107 - Guided first run over the real UI, with one-time tips (by @Abhi1992002)
- #16101 - Hire an expert and hand off the task in one click (by @Abhi1992002)
- #16252 - Role-based prompts, guidance, and a way back from a blocked mic on the brain dump step (by @itsababseh)
Enhancements
- #15817 - Delete what the old navigation left behind (by @Pwuts)
- #15763 - Compile the copilot with the React Compiler (by @Abhi1992002)
- #16040 - Move copilot (rest) onto colour and text tokens (DS5g-3) (by @Abhi1992002)
- #16041
- #15970 - Self-hosted Docker Compose generates its own service passwords per install and won't start without them: run
- autogpt-platform-beta-v0.9.0🚀 Release `autogpt-platform-beta-v0.9.0`2026-10-23 · 2026-10-23 01:28 UTC
查看更新说明
🔥 What's New?
New Features
- #15824 - Add measures and admin tools against multi-account abuse (by @Pwuts)
- #14553 - Self-host from the Docker Hub image, with a Windows installer (by @ntindle)
- #14981 - Every block declares its effect; Auto judges cheap outward actions and asks first on costly ones (by @Pwuts)
- #15675 - Add admin users page to look up, ban and impersonate a user, with rate limits folded in (by @Pwuts)
- #15859 - Send the browser tools' traffic through an egress proxy (by @ntindle)
- #15215 - Redesign usage limits and review upgrade charges (by @Torantulino)
- #15388 - AutoPilot turns pause at their limits instead of dying, and Keep going picks them up (by @ntindle)
- #15390 - Resume or send a chat's turn when the budget resets (by @ntindle)
- #15403 - Add send at reset, the out-of-budget panel for Max, and a running-low hint (by @ntindle)
- #15585 - Add voice mode to the + menu; model and token usage open as dialogs (by @0ubbe)
- #15115 - Add per-expert default approval mode for new web chats (by @Abhi1992002)
- #14551 - Distill private account and Expert skills during dreaming (by @ntindle)
- #14960 - Add durable per-scope schedule registry for dream and community jobs with lifecycle hooks and backfill (by @ntindle)
- #14961 - Add durable DreamPass record written from the sync and batch dream paths (by @ntindle)
- #14970 - Add one guard against overlapping dream passes and a stop path built on the DreamPass record (by @ntindle)
- #15374 - A scheduled chat says once why its run was skipped (by @ntindle)
- #15375 - Pause a scheduled chat that keeps failing the same way, and say why (by @ntindle)
- #15377 - Alert the owner when a scheduled chat pauses, list it on Home, resume it on reconnect (by @ntindle)
- #15400 - Run a scheduled chat less often from its usage-limit note (by @ntindle)
- #14865 - Serve the swap proxy from a service that exposes only its two calls (by @kcze)
- #14868 - Hand boxes behind the swap proxy placeholders, not tokens (by @kcze)
- #15874 - Add multi-choice, answer review, full keyboard use and touch sizing for expert onboarding cards (by @0ubbe)
- #15836 - Hire the whole recommended team in one click on the hire step (by @Pwuts)
- #12652 - Show a changelog notification for new releases (by @ntindle)
- #15850 - Send Postmark email deliveries, opens, clicks and bounces to PostHog (by @kcze)
- #15307 - Add Claude Haiku 5.5 to the LLM catalog (by @ntindle)
- #15566 - Add Meta Muse Spark Open model (by @ntindle)
- #15023 - Add a Cookie settings link to signed-out pages (by @kcze)
- #15483 - Read PostHog flags from one project while dev reports to its own (by @kcze)
Enhancements
- #15564 - The chat's approval gate answers the review pauses of workflows it started (by @Pwuts)
- #15701 - A banned account is refused by the API and its scheduled runs stop (by @Pwuts)
- #15906 - Show people who don't own a workflow the shared version of it (by @ntindle)
- #15982 - Keep each run in an org its runner belongs to (by @ntindle)
- #15949 - Require a shared key on calls between the backend's internal services (by @ntindle)
- #15951 - Start every Claude Code CLI from an allow-listed environment (by @ntindle)
- #15926 - Apply the outbound address rules to every IPv6 form of a blocked address (by @ntindle)
- #15986 - Send the Ollama blocks' traffic through the egress proxy (by @ntindle)
- #15959 - Pre-screen email addresses before sign-up and auth sends (by @Bentlybro)
- #14869 - Declare each provider's content hosts so the swap proxy scrubs them (by @kcze)
- #14872 - Hold each box to the run's ceiling on connected accounts (by @kcze)
- #14874 - Cap the swap proxy's credentialed requests per box and per user (by @kcze)
- #15725 - Let Otto view workspace and rendered images (by
- preview-seed-fixturePreview seed fixture (rolling)pre-release2026-08-06 · 2026-08-06 04:50 UTC
查看更新说明
Rolling, fully SYNTHETIC preview-database seed fixture baked from dev@a8af1a855687 (migration head: 20261104120000_store_creator_counts_visible_listings). Preview environments restore fixture.dump.gz (custom-format, gzipped, platform schema only, including _prisma_migrations) BEFORE running a PR's own migrations, so schema changes run against populated tables. RESTORE: apply restore-preamble.sql to the target database first (it creates the platform schema and installs the vector + pg_trgm extensions into it, which pg_dump --schema=platform omits), then run 'gunzip -c fixture.dump.gz | pg_restore -d --no-owner' WITHOUT --exit-on-error (the dump's own CREATE SCHEMA collides benignly with the preamble's). SECURITY INVARIANT: this bake holds no cloud credentials and has no read path to any real database — it runs entirely against a throwaway Postgres container and contains only synthetic rows (Faker-generated, plus the hand-written preview marketplace) and the already-public marketplace agent exports checked into autogpt_platform/backend/agents/. This asset is regenerated on every eligible dev push and weekly; treat it as disposable.
用户点评
登录用户的真实使用记录，可按版本和硬件筛选共 0 条
还没有人点评。第一个说出真实体验的人最有价值。
One conversation. The right expertise.
Tell Otto what's on your plate. It brings in the right Experts, hands over the work, and keeps you up to date. Talk to a specialist directly when you want to get into the details.
Each Expert has a role you can shape with your instructions, context, skills, and connected apps. Start with a ready-made specialist or create your own.
| Meet your team | Put them to work |
|---|---|
| Otto · Head of AI | Coordinate a job across specialists and bring their work back to you. |
| Maria · SEO Content Manager | Turn a target keyword into a content brief and article draft. |
| Max · Account Executive | Research prospects, draft outreach, and prepare the next conversation. |
| Frankie · Executive Assistant | Prepare meeting context, email drafts, and a digest for the day. |
See the team at work
One launch brief to Otto. Maria prepares the article, Max researches prospects, and Frankie prepares the launch review.
https://github.com/user-attachments/assets/83cddd83-00f8-46e7-be9e-e0a7237c7eea
Watch the full 45-second demo → · Sign up for a trial
A 15-second highlight of a real run, edited and sped up. In this example, the team prepares the work for review without sending or publishing it.
From a first brief to a routine
- Choose your Expert. Tell Otto what you need, pick a specialist, or create one for your own way of working.
- Brief them on the job. Share the context, connect the apps they need, and agree on the first outcome.
- Review and refine. Check the work and adjust the instructions. Activate routines for jobs that keep coming back.
You choose which accounts each Expert can access, set spending controls, and follow progress in their activity.
Get started
AutoGPT Platform
Use the hosted Platform to get your team running without managing the infrastructure or supplying model API keys.
The hosted Platform is a paid service. See current plans, trial terms, and usage pricing.
Self-host AutoGPT
Run AutoGPT on your own infrastructure. You provide model credentials, configure the deployment, and manage its updates and operation.
The quickest way to self-host is the single-container image on
Docker Hub, significantgravitas/autogpt:
docker run -d \
--name autogpt \
--restart unless-stopped \
--shm-size 2g \
--ulimit nofile=65536:65536 \
--log-driver json-file \
--log-opt max-size=50m \
--log-opt max-file=5 \
-p 127.0.0.1:3000:3000 \
-e AUTOGPT_PUBLIC_URL=http://localhost:3000 \
-v autogpt-data:/data \
significantgravitas/autogpt:latest
Give the first boot a few minutes, then open http://localhost:3000, create your account, and make it the administrator:
docker exec autogpt autogpt-admin promote [email protected]
Requirements, closing signup, ports, model providers, backups, and upgrades are covered in the guide:
Read the single-container guide →
On Linux, macOS, and Windows, the installer sets up the same image with one command.
To run each service in its own container with Docker Compose, or to develop from source, use the manual self-hosting guide.
| Hosted Platform | Self-hosted | |
|---|---|---|
| Setup and operations | Managed by AutoGPT | Managed by you |
| Model access | Built in | Bring your own model credentials or a local model |
| Cost | Paid plan and usage | Your infrastructure and model usage |
| Data and infrastructure | Hosted by AutoGPT | Runs on your infrastructure |
See the license for the repository's usage terms.
Workflows when you need more control
Experts can use skills and workflows to do their jobs. The visual builder lets you define repeatable steps, connect tools, and inspect how work runs. Start with a workflow from the marketplace or build your own.
- Visual builder: connect blocks, branch logic, and control the steps.
- Marketplace: discover workflows you can adapt to your work.
- Integrations: connect the tools your team uses.
- API and MCP server: run and build workflows from your own code, or from an AI coding agent.
Build with us
Explore the code, report a problem, or help improve AutoGPT.
| Resource | Where to go |
|---|---|
| Platform source and development setup | autogpt_platform/ |
| Documentation | agpt.co/docs |
| Contributing | CONTRIBUTING.md |
| Bugs and feature requests | GitHub Issues |
| Community and support | Discord |
License
| Component | License | What it means |
|---|---|---|
autogpt_platform/ | Polyform Shield | Free for personal and internal business use; cannot be sold as a competing hosted service |
classic/ and everything else | MIT | Permissive open-source use |
AutoGPT Classic
Looking for the original standalone AutoGPT? Explore Classic, Forge, or agbenchmark.
Contributors
Deutsch | Español | français | 日本語 | 한국어 | Português | Русский | 中文