MagiCrew Review (2026): Pricing, Features & Honest Verdict

MagiCrew is an AI agent platform that deploys specialised digital workers instead of a chat box, sold on points and marketed as open source. Our verdict: capable and cheap, but the licence is not the Apache 2.0 it advertises. Best for: teams that want to self-host. Price: From $9.99/mo (free plan: yes). Rating: 6.5/10

What is MagiCrew?

MagiCrew is an agent platform built around the idea that you deploy workers rather than chat with a model. Each one is scoped to a job: research, analysis, report writing, presentation building. The company behind it is Guangdong Lighthouse Engine Technology Co., Ltd., which also publishes the codebase on GitHub as dtyq/magic, where it carries just over 5,000 stars and 557 forks. It arrived at number two on Product Hunt this week.

The product has two halves. There is a hosted cloud service with a points-based subscription, and there is a self-hosted community edition you deploy yourself. The homepage sells the second one hard: “Fully open-source under Apache 2.0. Complete data sovereignty, customizable, ideal for enterprises and developers with technical teams.” We checked that claim against the licence file and it does not survive contact. More on that below, because it is the most consequential thing about this product.

What are MagiCrew’s key features?

A digital employee marketplace instead of prompt templates

The core abstraction is a configured worker rather than a saved prompt. MagiCrew’s framing is that you encapsulate business knowledge, ERP and CRM context into reusable roles that any employee can deploy. That is a meaningfully different product shape from a chat assistant, and it maps to how work is actually assigned inside a company. It also means the setup cost is real: an unconfigured digital employee is worse than a chat box.

Both text and vision models in every paid tier

Every paid plan carries the same model access, listed as Gemini 3.5 Flash, Gemini 3.1 Pro, Gemini 3 Flash, Claude Opus 4.8, three GPT-5.6 variants, GPT 5.5, Kimi K3, DeepSeek V4 Pro and MiniMax M3, plus a vision set covering Nano Banana 2 and Pro, GPT Image 2, Seedream 5.0 Pro, Veo 3.1 in two speeds, Seedance 2.0 and Kling 3.0. Not gating models by tier is the right call and it is rarer than it should be. One note for anyone choosing on model quality: the newest Anthropic model on that list is Opus 4.8, a generation behind the current Opus 5, and the free tier gets “Select Models” rather than all of them.

Parallel task slots as the real capacity limit

Points buy work, but concurrency decides throughput. Free runs one parallel task, Plus two, Pro three, and both Max and Ultra five. If you are running long research jobs, the parallel slot count constrains you before the points do, and it stops improving at the $49.99 tier. Nothing on the pricing page frames it this way, so read that row before choosing a plan on point volume alone.

A workflow engine, IM and sandbox in the same codebase

The GitHub description calls it a generalist AI agent plus a workflow engine plus instant messaging plus an online sandbox, and the repo topics list low-code, no-code, MCP and sandbox. That is a wide surface for one project. It is also the reason self-hosting is attractive here: you get the orchestration layer, the execution sandbox and the chat interface as one deployment rather than three vendors.

How much does MagiCrew cost?

Four individual plans, all currently discounted. Plus is $9.99 a month against a listed $29.00, marked 66% off, with 500 monthly points, two parallel tasks and 100GB of storage. Pro is $24.99 against $72.50 with 2,500 points, three parallel tasks and 1TB. Max is $49.99 against $150.00 with 13,000 points, five parallel tasks and 2TB. Ultra is $99.99 against $300.00 with 25,000 points, five parallel tasks and 2TB. A free tier exists on sign-up with 200 daily login points, one parallel task, 1GB of storage, limited feature access and no ability to buy top-ups. Team plans are sold per seat, and the self-hosted enterprise plan is contact-only.

Every discount reconciles. $29.00 to $9.99 is 65.6%, $150.00 to $49.99 is 66.7%, and the cards round those to 66% and 67% correctly. Unused points do not carry over on any tier.

Against the field, this is aggressive. Dify’s pricing page defaults to annual billing and renders $590 per workspace per year for its developer tier and $1,590 for its team tier, with the monthly rates hidden behind a toggle that does not appear in the page source. MagiCrew’s Max tier is $600 a year at the promotional price.

Divide the price by the points and the ladder inverts

Run the one calculation the pricing page does not. Plus is $9.99 for 500 points, which is $2.00 per 100 points. Pro is $24.99 for 2,500, which is $1.00 per 100. Max is $49.99 for 13,000, which is $0.38 per 100. Ultra is $99.99 for 25,000, which is $0.40 per 100.

Ultra is 4% worse per point than the tier below it. Doubling your spend from $49.99 to $99.99 buys 12,000 more points at a worse rate and changes nothing else: the comparison table gives Max and Ultra the same five parallel tasks, the same 2TB of storage, the same model access and the same feature set, and the two cards even carry the identical description, “Maximum capacity for demanding workloads”. Two Max subscriptions cost $99.98, one cent less than Ultra, for 26,000 points instead of 25,000. The inversion is not an artifact of the promotion either: at list price Max is $1.15 per 100 points and Ultra is $1.20.

At the other end, Plus at $2.00 per 100 points costs 5.2 times what Max charges for the same unit. The cheapest plan is by far the worst value per unit of work, which is the normal shape, but a 5.2x spread across four self-serve tiers is not.

One more thing we could not resolve. Each card prints a small per-unit figure beside the discount badge: “$0.35/100 Points” on Plus and Pro, “$0.34/100 Points” on Max and Ultra. Those match none of the four plans’ own arithmetic, and nothing on the page explains what they refer to. Our best reading is that it is the rate for the purchasable point bundles, which the page lists as “Available” without publishing a price. Either way, the number a buyer sees next to the plan is not the rate that plan charges. Meter on the free tier before you commit.

Is MagiCrew actually open source?

The site says so in three places: “AI Agents, Open Source, Enterprise” in the hero, “Fully open-source under Apache 2.0” on the deployment card, and “enterprise-grade open-source AI Agent platform” in the footer. GitHub disagrees. The API reports the licence as NOASSERTION and the name as “Other”, which is how GitHub records a licence it cannot classify.

The file itself explains why. It is headed “Magic is licensed under a modified version of the Apache License 2.0, supplemented by the following terms”, and the supplements are substantive. Operating a multi-tenant SaaS service using the source code is prohibited without explicit written consent. Logos, trademarks and copyright notices may not be removed or altered from the console or applications. And the contributor agreement states that the producer can adjust the licence to be stricter or more relaxed as it sees fit, and that contributed code may be used commercially in the company’s own cloud business. Licensing enquiries go to a business development address.

Those terms may be perfectly reasonable commercial choices. They are not Apache 2.0, which permits exactly the hosted competition this licence forbids, and they are not what “fully open-source under Apache 2.0” tells a reader. The distinction matters most to the audience the copy targets: an enterprise or an agency planning to run MagiCrew for multiple client organisations needs written consent, and the licence defines a tenant as an independent organisation to make sure of it.

Two related signals from the same repository. There are zero tagged releases, so anyone self-hosting is pinning to a moving branch rather than a version. And commits are concentrated in one contributor with 891 of them, which is a bus-factor risk on a platform being sold to enterprises. The last push we saw was three weeks before this launch.

Who is MagiCrew best for?

Use MagiCrew if you want a self-hosted agent platform with a workflow engine and sandbox in one deployment, you are running it for a single organisation, and data residency is why you are shopping. That use case is squarely inside the licence and the codebase is substantial. Use the cloud version if you want broad model access cheaply, in which case go straight to Max at $49.99, which is the best rate on the ladder by a distance.

Skip MagiCrew if you intend to host it for clients or tenants, because you need written permission and the site does not tell you that. Skip it if you need a tagged release to pin against. Skip it if Chinese governing law is a problem for your data, since the terms specify it. And skip Ultra outright: at $99.99 it is the worst per-point deal on the page and adds no capability over Max.

Best MagiCrew alternatives

Dify is the closest comparison and the more established open-source LLM application platform, with a large plugin ecosystem and a self-hosted community edition. Its cloud pricing page defaults to annual and shows $590 and $1,590 per workspace per year for its two self-serve paid tiers. It is more expensive and its licensing position is clearer.

n8n is the workflow-automation route for teams who want agents as nodes in a broader pipeline rather than as a standalone product. Its pricing page geolocates and returned no dollar figures from our location, so we are not quoting one. Its fair-code licence carries its own restrictions, which is the same conversation as MagiCrew’s, just conducted more openly.

Manus is the general-purpose agent that made the credits-for-tasks model popular. Its pricing renders client-side and we could not read a figure first-hand, so treat it as the hosted-only, higher-priced option with no self-host path at all.

Final verdict: is MagiCrew worth it?

MagiCrew is a real platform with a real company behind it, five thousand stars, a wide feature surface and the cheapest broad model access we have priced this month. At $49.99 for 13,000 points, five parallel tasks, 2TB of storage and every model on the list, the Max tier is a strong deal, and the free tier is generous enough to evaluate on. If you want an agent platform you can run yourself for one organisation, it belongs on your shortlist.

It lands at 6.5 because of what the marketing says rather than what the software does. “Fully open-source under Apache 2.0” is not accurate for a licence that forbids multi-tenant hosting, mandates branding retention and reserves the right to be tightened later, and the audience most likely to act on that sentence is the one the restriction is aimed at. Add a ladder that gets worse per point at the top rung, a per-unit figure printed on every card that matches no plan, zero tagged releases and a single dominant contributor, and the picture is a good product sold with copy that needs correcting. Read the licence file before you deploy it, and buy Max rather than Ultra.