Overview
- Tool: MagiCrew (超级麦吉)
- Developer: Lighthouse Engine
- Website: https://www.magicrew.ai/
- China site: https://www.letsmagic.cn/
- Type: Cloud web app, also open source and self-hostable
- Repository: https://github.com/dtyq/magic (TypeScript, ~5.1k stars)
- Category: AI agent platform
MagiCrew is an enterprise AI agent platform built on a plain premise: an assistant should hand you a file you can use, not a paragraph you still have to format. Ask for a deck and you get a deck. Ask for a dashboard and you get an HTML dashboard you can open and click through.
Lighthouse Engine, founded in 2023 and headquartered in Shenzhen, ships it through two sites: magicrew.ai internationally and letsmagic.cn in China. The code is public on GitHub under the Magic Open Source License, a modified Apache 2.0, so teams that cannot send work to a third-party SaaS can run the whole platform in-house.
What it is used for
- One-line decks: give it a sentence, a topic or a reference document and it builds the storyline, writes the outline, matches images page by page and exports a slide deck you can keep editing.
- Dashboards: turn spreadsheets or business data into a visual dashboard delivered as an HTML file you can open directly.
- Deep research reports: multi-round research with live web retrieval, producing long structured reports with supporting data.
- Meeting notes: live transcription, contested points flagged, action items tracked, and a summary with owners and deadlines pushed to attendees afterwards.
- Digital worker marketplace: wrap ERP, CRM, internal databases and job expertise into agents that anyone in the company can call.
- Multi-agent parallelism: an orchestrator agent splits the task and dispatches specialist agents to work simultaneously instead of answering one prompt at a time.
- Custom skills: define skills conversationally with Skill Creator, compatible with the Anthropic Skills and OpenClaw Skills ecosystems.
- Cost control: three-tier budgets by department, user and task, so every model call is attributed to a person and a job.
Pricing
Cloud plans are point-based subscriptions. Figures below are from the official pricing pages, checked in October 2026.
- Free: 1,000 points on sign-up, plus a limited-time 200 points for daily login. 1 parallel task, 1GB storage, limited features and a subset of models.
- Plus: $9.99/mo (list price $29.00/mo, 66% off). 2,900 points per month (1,400 base + 1,500 bonus), 4 parallel tasks, 20GB storage, $0.35 per 100 points.
- Pro: $24.99/mo (list $72.50). 7,250 points (3,500 + 3,750), 6 parallel tasks, 40GB storage.
- Max: $49.99/mo (list $150.00). 15,000 points (7,500 + 7,500), 10 parallel tasks, 160GB storage.
- Ultra: $99.99/mo (list $300.00). 30,000 points (15,000 + 15,000), 20 parallel tasks, 320GB storage.
All four paid tiers include the same rights: unlimited use of every feature, access to all current text, image and video models, purchasable point top-ups, and priority scheduling at peak hours. Unused points do not roll over - the pricing page states this explicitly.
The China site lists the same tiers at ¥58, ¥128, ¥268 and ¥668 per month.
Self-hosting costs nothing for the code itself; servers, model API calls and operations are yours to carry. Enterprise terms - private deployment, dedicated model integration, custom integration with internal systems - are quoted individually via bd@dtyq.com.
Who it is for
- Small teams that need big-team output: market research, industry reports, social copy and campaign plans handled by several agents at once, so headcount stays flat while output rises.
- Companies with scattered internal systems: wrap ERP, CRM and databases into digital workers so staff can ask a question instead of learning an API.
- Organisations with strict data boundaries: if data cannot leave your network or enter a third-party SaaS, self-host it.
- Anyone producing a recurring report: weekly updates, monthly reviews, quarterly retrospectives, from data pull to finished document in one chain.
- Teams about to lose a retiring expert: encode the diagnostic path of a senior engineer into an agent so juniors get senior-level judgement.
- Heavy individual users: one person running like a small team, producing decks, reports and dashboards without help.
How to use MagiCrew
There are two routes. Start with the cloud one.
1. Open https://www.magicrew.ai/ (or https://www.letsmagic.cn/ for China) and sign up with email or phone.
2. The account starts with 1,000 points and one parallel task, enough to test real work before paying.
3. Type a request, or attach source material - documents, images and voice input are all accepted.
4. Review the plan it proposes; the orchestrator agent then splits the work and dispatches specialist agents in parallel.
5. Results arrive as a deck, dashboard, report or spreadsheet that you can preview and edit in the browser.
6. Put standing context - your industry, house style, reporting rules - into long-term memory so you do not repeat it every time.
7. When volume grows, upgrade the plan or buy point top-ups.
Self-hosted:
1. You need a machine with Docker and curl. macOS and Linux are supported; Windows is listed as coming soon.
2. Run `curl -fsSL https://getmagicrew.sh | bash`.
3. The script creates the cluster, provisions the infrastructure and deploys the services.
4. When it finishes, open http://localhost:38080 for the web UI.
5. Tear everything down with `magicrew teardown`.
6. You supply your own model API keys; see the deployment guide under docs in the repository.
Practical advice
- Give it a goal, not a sequence of steps. A complete brief ("a Q3 review for management, 8 pages, lead with growth and the problems") works far better than asking for paragraphs one at a time.
- Start small, then commit. Run two or three real tasks on the free points to see how it performs in your industry before choosing a tier.
- Watch the point budget. A deck and a deep research run are not in the same consumption class; check the balance before heavy tasks.
- Invest in long-term memory. Industry, definitions and writing style written once pay off in every later output.
- Keep the approval loop on. Deleting data or sending email triggers human approval by default - do not switch it off to save clicks.
- Do the maths before self-hosting. Open source is not zero cost; model API spend dominates, especially for teams running long tasks often.
- Do not parallelise a vague task. If the split is fuzzy, several agents will fight each other. Think through the decomposition first.
Limitations and things to know
- The licence is not plain Apache 2.0. The repository uses the Magic Open Source License: Apache 2.0 plus two restrictions - you may not run a multi-tenant SaaS on the source code without written permission, and you may not remove logos, trademarks or copyright notices from the console or applications. Read the LICENSE before commercial redistribution.
- No Windows support for self-hosting yet. The README lists macOS and Linux only, with Windows marked as coming soon.
- Points expire monthly. Unused points are cleared, so buy the tier you actually use.
- The free tier is an evaluation tier. One parallel task, limited features and a subset of models - fine for testing, not for daily production.
- It depends heavily on model APIs. Quality is bounded by the models you connect, and cost scales with call volume.
- Headline numbers come from the vendor. "Hundreds of experts" and "100k+ skills" are the company's own claims; the star count you can verify on GitHub, and it sat around 5.1k when this page was written.
- Web only. There is no desktop client, so fully offline work is out of scope.
- Enterprise features are not self-serve. Private deployment, dedicated models and system integration all require a conversation.
FAQ
Is MagiCrew free?
Sign-up comes with 1,000 points, plus a limited-time 200 points for daily login, which is enough to evaluate it. Ongoing use needs a subscription starting at $9.99/month. Self-hosting has no licence fee, but you pay for servers and model API usage.
How is this different from a chat assistant?
A chat assistant returns text that you still have to lay out, tabulate and export. MagiCrew returns the deck, dashboard, report or spreadsheet itself, so there is no manual conversion step.
Do I need to write code?
No. The cloud version is a web interface where you type a request. Coding only comes in if you choose to self-host.
Can I run it on my own servers?
Yes. The code is at github.com/dtyq/magic and installs with a single script; you need Docker. macOS and Linux are supported today.
Which models does it use?
The pricing page says paid tiers get "all latest models" across text, image and video, but it does not publish a model-by-model list, so the exact selection is whatever you see after signing in. Self-hosted installs use whatever models you configure.
How is data security handled?
Per the vendor: each agent runs in an isolated sandbox container in a separate VPC from the main system, connected over private endpoints; each user gets an independent sidecar network proxy; plugins pass a security review before listing. Teams with strict boundaries can deploy privately.
Can I use it commercially?
Yes, with two caveats: you cannot run a multi-tenant SaaS on it without written permission, and you cannot strip the branding. The full terms are in the repository's LICENSE file.
*This page is not affiliated with MagiCrew. Details were compiled from letsmagic.cn, magicrew.ai and the github.com/dtyq/magic repository in October 2026. Prices and quotas change - check the official pages before purchasing.*
