A self-hosted alternative to Manus
Manus suits people who want a finished app that researches, writes, designs and builds for them, with nothing to install or run. Agenhood suits developers and teams who want general-purpose agents on their own server, driven by an API and schedules, with the model they pick.
Last updated
If you want a self-hosted Manus alternative, it helps to be clear about what you would give up and what you would get. Manus is a hosted general AI agent: you describe a task, and it works in its own virtual computer until it hands back a report, a deck or a website. Agenhood is an open-source platform you run on your own server, where each agent is a long-lived, sandboxed container you send work to from a console, an API or a schedule. Manus is the more finished product. Agenhood is the one you own.
The short answer
Pick Manus if
- You want a polished app on the web, desktop and phone.
- You don't want to run or patch a server.
- You're not a developer, or don't want to touch an API.
- You want slides, video, websites and games built for you.
Pick Agenhood if
- Your prompts, files and task history must stay on your own server.
- You want agents you drive from a REST API, schedules and workflows you host yourself.
- You want to pick the model for each agent, from Anthropic or OpenAI.
- You want MIT-licensed code you can read and change.
Side by side
| Topic | Manus | Agenhood |
|---|---|---|
| Who runs it | Manus. You use it on the web, desktop or phone, or through its API. source | You, on one VM with Docker Compose. No hosted version. |
| Where tasks run | A sandbox that Manus provides: a virtual computer with internet access and a persistent file system. source | One hardened container per agent on your server, with a filtering egress proxy as the only way out. |
| Web browsing | A cloud browser that clicks, fills forms and can use accounts you log into, with a take-over mode. source | Web search through a SearXNG instance in the stack, and page fetching with headless Chromium. No clicking or logins. |
| Models | Agent profiles (Standard, Lite, Max). Model versions are not chosen separately. Media models can be picked per task. sourcesource | Anthropic and OpenAI models, chosen per agent, plus Opencode's own model plans. |
| Memory between tasks | Projects: a master instruction and a knowledge base of files applied to every new task. source | Each agent keeps its /workspace volume across tasks and weeks of idle. Without a linked Git repo, it is snapshotted after every task. |
| Scheduling and triggers | Automations: schedules, tasks triggered by events in connected apps, and advanced automations (beta). source | One-time or recurring schedules in your timezone, workflows across agents, and GitHub and Slack triggers. |
| API | REST API for tasks, projects, files, skills and webhooks. source | REST API plus server-sent events for everything the console does. Output can be schema-validated JSON. |
| Integrations | Prebuilt and custom MCP connectors, Slack, Zapier and more. source | MCP servers by HTTPS URL, skills from Git repos, and GitHub and Slack connectors. |
| Teams | Team plan with a shared credit pool and admin controls. Collab lets several people work in one task. sourcesource | Multi-tenant workspaces with owner, admin and member roles. |
| What it produces | Documents, spreadsheets, PDFs, slides, websites, code, games and video, in Manus Studio. source | Any file the agent can write with bash and Python, saved to its workspace for you to download. |
| Pricing model | Credits, with Free, Pro and Team plans. See their pricing page for current prices. source | Free and MIT licensed. You pay for your server and for model usage on your own keys. |
Where your data lives
Manus runs the agent, the sandbox and the cloud browser for you. That is why there is nothing to set up. It also means your prompts, files and any accounts you log into in the cloud browser are handled by Manus's service.
With Agenhood, the control plane, the database, every agent's container and its
/workspace volume run on a machine you choose, from make prod on one VM.
Task history, files and logs stay there. Provider keys are stored encrypted (AES-GCM) and never sent
to the browser. The part that still leaves your server is each model call, which goes to Anthropic
or OpenAI with the prompt and context it needs.
A research report, end to end
This is the kind of job people use Manus for: research a topic, write it up, get a file back. Here is
the same job on Agenhood, using only the API. It assumes make dev is running and an
Anthropic key is saved under Settings, Credentials.
# Create a research agent with web search, page fetching and file tools
curl -X POST http://localhost:5173/v1/containers \
-H "Authorization: Bearer tk_live_seedkey" \
-H "Content-Type: application/json" \
-d '{"name": "researcher", "config": {"driver": "vanilla", "model": "claude-sonnet-4-6", "tools": ["web_search", "web_fetch", "read_file", "write_file"]}}'
# Ask for the report (agents are "containers" in the API)
curl -X POST http://localhost:5173/v1/containers/$AGENT_ID/tasks \
-H "Authorization: Bearer tk_live_seedkey" \
-H "Content-Type: application/json" \
-d '{"prompt": "Research how five EU countries subsidise home heat pumps in 2026. Compare amounts, eligibility and how to apply, cite every source with its URL, and save the report to report.md"}'
# Check the task until its status is "succeeded"
curl http://localhost:5173/v1/containers/$AGENT_ID/tasks/$TASK_ID \
-H "Authorization: Bearer tk_live_seedkey"
# Download the report from the agent's /workspace
curl -o report.md "http://localhost:5173/v1/containers/$AGENT_ID/files/raw?path=report.md" \
-H "Authorization: Bearer tk_live_seedkey"
Take $AGENT_ID from the first response and $TASK_ID from the second. While
the task runs, the console's task viewer shows each search, each page the agent reads and the moment
report.md is written, with a token meter and a cancel button. You can also open the
agent's Files page and download the file, or the whole workspace as a zip.
The agent keeps its workspace afterwards. Next week you can ask it to update the report, and it starts from the file it already wrote. To make that automatic, save the prompt and put it on a weekly schedule.
Choosing the model
In Manus you choose an agent profile: Standard, Lite or Max, with several thinking levels for Max. Its API docs say model versions are not selected independently, and free accounts are moved to Lite. Manus does let you pick from a list of image, video, audio and 3D models, or leave the choice to it.
In Agenhood every agent has a driver and a model. A research agent can run Claude Sonnet in Agenhood's own tool loop, and the agent next to it can run Codex on an OpenAI model. You can change an agent's model later, and usage goes to your own API key. You can also sign in with your own Claude or ChatGPT plan, on your own server.
How the web gets used
This is where Manus is clearly ahead. Its cloud browser visits sites, clicks buttons, fills forms and can work inside accounts you log into, and it asks you to take over for things like CAPTCHAs or one-time codes. Wide Research splits a big list, such as 100 products, across many parallel agents.
An Agenhood agent searches the web, fetches pages as text, and renders pages that need JavaScript with headless Chromium. That covers most research and monitoring jobs. It doesn't log in or fill forms. To spread a long list across agents, you create several agents and split the work yourself, with a workflow or your own script.
What Manus does better
- No setup at all. Sign up and start. Agenhood needs a Linux machine, Docker and someone to keep it updated.
- An interactive browser. The cloud browser works in logged-in accounts. Agenhood agents only read pages.
- Finished outputs. Manus Studio builds slides, websites, games and edited video. Agenhood hands back files.
- Apps for everyone. Web, desktop and mobile apps, Slack, and real-time collaboration in one task. Agenhood is a web console and an API for developers.
Moving over, or using both
You don't have to pick one for everything. Keep Manus for one-off work that needs its browser or its design tools, and move the jobs that touch private data, run on a schedule or are started from your own code to Agenhood. If a job lives in a Manus project, copy its master instruction into a saved prompt, put its knowledge-base files into the agent's workspace, and recreate any Manus automation as an Agenhood schedule.
What it doesn't do
- No interactive browser. Agents can't log into sites or fill in forms. If that is the job, use Manus or a browser automation tool.
- Built for developers. Setup is a terminal, Docker and an API key. There is no mobile app and no consumer sign-up.
- One Docker host, and you run it. No Kubernetes and no hosted fallback, so updates, backups and uptime are on you.
Questions people ask
Is Agenhood an open-source version of Manus?
No. It is a separate project with a different focus. Manus is a hosted product for anyone, with apps, slides, video and website building. Agenhood is MIT-licensed infrastructure for running sandboxed agents on your own server, and it is built for developers who are happy to use an API and a terminal.
Can Agenhood make slides, videos or websites like Manus?
Not as a built-in feature. An Agenhood agent writes whatever files bash and Python can produce, such as Markdown, CSV, HTML or a chart image, and you download them from its workspace. There is no slide designer, video editor or website hosting.
Does Agenhood need a Manus account or credits?
No. It doesn't talk to Manus at all. You add your own Anthropic or OpenAI API key, and usage is billed to that key. You can also sign in with your own Claude or ChatGPT plan, on your own server, for personal use.
Can an Agenhood agent log into websites for me?
No. Agents search the web and read pages, including pages that need JavaScript, through headless Chromium. They don't click through sites or use your logged-in sessions. For that, the Manus cloud browser or its browser extension is the better tool.
Where does my data go with Agenhood?
Task history, files and logs are stored on the server you run. Each model call still sends its prompt and context to the model provider you picked, Anthropic or OpenAI. By default, web search goes through a SearXNG instance that runs inside the stack.