OpenManus: Install and first job
This page installs OpenManus on Windows 11 or a Mac, sets up the safety rules, connects a free model, and runs one small job. It comes after Pick it and plan, so have your model choice ready.
Checked 2026-11-14.
Install on Windows 11
The OpenManus README shows a Windows line for one step only, the line that turns on the Python space, and its uv install command is the Mac and Linux one. The Windows command below comes from uv's own install page instead. This guide could not confirm that OpenManus is tested on native Windows with uv. It was not tried end to end for this guide.
On 14 November 2026, steps 4 to 8 were run on one Windows 11 PC that already had git and uv, and the install finished without errors. That check did not cover steps 1 to 3, and it did not run a job with a real key.
1. Open PowerShell. Click Start, type
PowerShell, and click it. Do not choose "Run as
administrator."
What you should see a window with a line that ends in
>, waiting for you to type.
2. Install git. Go to git-scm.com/install/windows, download the x64 file (or ARM64 if your PC has an ARM chip), and run it. This guide did not walk through each installer screen. Then close PowerShell, open it again, and type:
git --version
The same page also lists
winget install --id Git.Git -e --source winget, and notes
that package is not maintained by the Git for Windows project.
What you should see a line that starts with git version
and a number.
3. Install uv. This line is from uv's install page:
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"
uv's page says changing the execution policy is what lets a script
from the internet run. Close PowerShell, open it again, and type
uv --version.
What you should see a line that starts with uv and a
version number.
4. Make the box folder. Make one new, empty folder
and move into it. This guide calls it agent-tryout, and Keep it in a box explains why.
mkdir $HOME\agent-tryout
cd $HOME\agent-tryout
What you should see no error. The line where you type now shows a
path ending in agent-tryout.
5. Download OpenManus.
git clone https://github.com/FoundationAgents/OpenManus.git
cd OpenManus
What you should see lines about "Cloning into 'OpenManus'," then a
path ending in agent-tryout\OpenManus.
6. Make its own Python space.
uv venv --python 3.12
.venv\Scripts\activate
The second line is the Windows line from the OpenManus README. PowerShell may refuse to run it and talk about scripts being disabled, and this guide could not confirm a fix from the OpenManus README. If that happens, stop and use Agent Zero instead, or read the Python documentation's fix in Fixes and decide for yourself.
What you should see a new .venv folder, and the line
where you type usually now starts with a name in parentheses.
7. Install the parts.
uv pip install -r requirements.txt
What you should see a long list of package names. The last lines should not say "error."
8. Make your settings file.
copy config\config.example.toml config\config.toml
What you should see no error. dir config lists both
config.example.toml and config.toml.
Now go to Keep it in a box, then Connect a model.
Install on a Mac
Every line in a gray box is one command. Type or paste it, then press Return.
1. Open Terminal. Press Command and Space, type
Terminal, and press Return.
What you should see a window with a line that ends in %
or $, waiting for you to type.
2. Check for git. Type:
git --version
If it says git is missing, the git website says Apple ships git with the Xcode Command Line Tools:
xcode-select --install
Follow the window that opens, then run git --version
again.
What you should see a line that starts with git version
and a number.
3. Install uv. This line is from the OpenManus README:
curl -LsSf https://astral.sh/uv/install.sh | sh
Then type uv --version. If it says "command not found,"
close Terminal, open it again, and retry. This guide did not test
whether reopening is always needed.
What you should see a line that starts with uv and a
version number.
4. Make the box folder. Make one new, empty folder
and move into it. This guide calls it agent-tryout, and Keep it in a box explains why.
mkdir ~/agent-tryout
cd ~/agent-tryout
What you should see no error. Typing pwd shows a path
that ends in agent-tryout.
5. Download OpenManus.
git clone https://github.com/FoundationAgents/OpenManus.git
cd OpenManus
What you should see lines about "Cloning into 'OpenManus'," then no
error. pwd now ends in
agent-tryout/OpenManus.
6. Make its own Python space. uv's own docs say it downloads Python 3.12 by itself if you do not have it.
uv venv --python 3.12
source .venv/bin/activate
What you should see a new .venv folder
(ls -a lists it). The line where you type usually now
starts with a name in parentheses.
7. Install the parts.
uv pip install -r requirements.txt
What you should see a long list of package names. The last lines should not say "error."
8. Make your settings file. This copies the example file to the name OpenManus reads:
cp config/config.example.toml config/config.toml
What you should see no error. ls config lists both
config.example.toml and config.toml.
Now go to Keep it in a box, then Connect a model.
Keep it in a box
OpenManus has no box, because the README describes no sandbox and no approval step. A quick look at the code on 14 November 2026 showed that the main agent's tools are a Python runner, a file editor, an "ask the human" tool, and the web browser. "Ask the human" is something the AI may choose to use, not a yes-or-no gate before each action. Any code it writes runs on your real computer, as you.
Your folder is where it starts, not a wall. By default it saves into
a workspace folder inside OpenManus, but code
it writes can still reach any file your account can reach.
So the box is you. Follow all of these:
- New empty folder. Run it only from
agent-tryout, and put nothing personal in there. - Never your home folder. Never start it from your home folder itself (the one with your name), or from Documents, Desktop, or Downloads.
- Your normal account, not admin. On Windows, never
"Run as administrator," and on a Mac, never type
sudofor it. A separate, ordinary user account on the computer is stronger still, though OpenManus's README does not cover that. - No logged-in email or bank. The README says its browser part "attaches to Chrome or Chromium automatically," so treat Chrome as the agent's browser. Sign out of email, banking, shopping, and calendar in Chrome before you run it, because a web page can hide instructions that the agent may obey.
- A spend cap at the provider. OpenManus has no money limit of its own, so stay on Groq's free plan and do not add a card. If you use a paid key, set a monthly cap first. At OpenAI it is Monthly spend limit with Enforce a hard limit turned on, and at Anthropic it is Monthly spend limit on the Spend limits page.
For the first job, turn the browser part off. The README names the switch, which lasts only for the window you type it in.
On a Mac:
export OPENMANUS_DISABLE_BROWSER_USE=1
On Windows:
$env:OPENMANUS_DISABLE_BROWSER_USE = "1"
To stop OpenManus at any time, click in its window and press Control and C.
The example settings file has a [sandbox] section that
is switched off (use_sandbox = false, with the lines
commented out), and it names a Docker setup. The README does not explain
it, and this guide could not confirm what it covers or that it works, so
it is not part of this guide.
For the reasons behind these rules, see the safety checklist.
Connect a model
This uses a Groq key, though this guide did not test OpenManus with
Groq. The code sends questions through a standard OpenAI-style
connection to whatever base_url you give it, and Groq
publishes that kind of address.
- Make a key at console.groq.com/keys. This guide could not confirm a free Groq model, so read the price before a long job.
- Open the settings file. On a Mac, inside the
OpenManusfolder, typeopen -e config/config.toml. On Windows, typenotepad config\config.toml. - Find the
[llm]block and copy the name from Groq's own list, changing only the three quoted lines:
[llm]
model = "qwen/qwen3.8-27b"
base_url = "https://api.groq.com/openai/v1"
api_key = "paste-your-groq-key-here"
- Find the block that starts with
[llm.vision]and make the same three changes there, so no example address is left behind. This guide could not confirm that your pasted model receives screenshots, so those steps may fail. - Go to the very end of the file and add these two lines on their own:
[daytona]
daytona_api_key = ""
On the code checked on 14 November 2026, OpenManus stops as soon as
it starts, with an error that names daytona_api_key, unless
the settings file has this section. That happens even though this guide
never uses Daytona, so leave the quote marks empty and do not sign up
for anything.
- Save the file and close it.
Never paste your key into the chat with the agent. Never upload or
share the OpenManus folder either, because the key now sits
inside it.
If you skip making config.toml, the code falls back to
the example file, which has no real key, so nothing will work.
What you should see config.toml saved with your model,
address, and key in both blocks, and the two [daytona]
lines at the end.
Or use a local model
OpenManus's example file has an Ollama block, and the project also
ships a whole Ollama example,
config/config.example-model-ollama.toml. This guide did not
test it, and one issue report shows a small local model failing to save
its output file, so read "On a computer you already own" in Pick your model and
budget first. Install Ollama and download a model with Run a local
model with Ollama.
In steps 3 and 4, use these values instead of Groq's, with
qwen3.8:27b, or qwen3.5:4b at about 8 GB:
The example file still names llama3.2 and
llama3.2-vision. Do not pull those.
[llm]
model = "qwen3.8:27b"
base_url = "http://localhost:11434/v1"
api_key = "ollama"
Ollama does not check the key, so ollama is only a
placeholder, as in the project's example. That example names
llama3.2-vision in the [llm.vision] block.
Keep the two [daytona] lines at the end.
What you should see after python main.py and a short
job, ollama ps in a second terminal window lists your
model.
Every agent in this guide reaches Ollama at port 11434, but the exact
address differs: agents in Docker use host.docker.internal,
and some need /v1 on the end. Connect
it to your agent gives each one's address.
Your first safe job
Use the same window where the Python space is on and the browser part
is off. You should be inside agent-tryout/OpenManus.
1. Start it.
python main.py
What you should see Enter your prompt:
If it stops with a block of error text instead, look up the last line of that text in Fixes.
2. Give it one small job. Type:
Write a short packing list for a weekend trip. Save it as packing-list.txt in your workspace folder. Do not open, change, or delete any other file. Do not use the internet.
Then press Enter (Return on a Mac).
Watch the lines as they scroll. If any line talks about deleting, moving, your home folder, Documents, Desktop, email, or a password, press Control and C.
What you should see "Processing your request...", some steps, then "Request processing completed."
3. Check the folder. On a Mac type
open workspace. On Windows type
explorer workspace.
What you should see packing-list.txt in the
workspace folder, and nothing else new.
Each python main.py takes one job and then ends, and the
code stops a job after 20 steps. That limits loops, but it does not
limit money.
Next time, open the terminal, go to
agent-tryout/OpenManus, turn the Python space on again
(step 6 of the install), turn the browser part off, and run
python main.py.
Sources
Opened 2026-11-14.
- OpenManus README (install, uv commands, Browser Use,
OPENMANUS_DISABLE_BROWSER_USE,config/config.toml,python main.py): https://github.com/FoundationAgents/OpenManus/blob/main/README.md - OpenManus example settings (Ollama block,
[sandbox]block, browser options): https://github.com/FoundationAgents/OpenManus/blob/main/config/config.example.toml - OpenManus Ollama example settings (
model = "llama3.2",llama3.2-vision,base_url = "http://localhost:11434/v1",api_key = "ollama"; the same file at tag v0.3.0 and at main commit3309bf4): https://github.com/FoundationAgents/OpenManus/blob/v0.3.0/config/config.example-model-ollama.toml - OpenManus issue #539, "no output file saved" with a local DeepSeek 8B model: https://github.com/FoundationAgents/OpenManus/issues/539
- Ollama OpenAI compatibility (
http://localhost:11434/v1, the key is required but ignored): https://docs.ollama.com/api/openai-compatibility - OpenManus settings loader (fallback to the example file,
workspacefolder,daytona_api_keyrequired): https://github.com/FoundationAgents/OpenManus/blob/main/app/config.py - OpenManus Daytona example settings (the
[daytona]block with an emptydaytona_api_key): https://github.com/FoundationAgents/OpenManus/blob/main/config/config.example-daytona.toml - OpenManus issue #1232, "Error occurred when running main.py"
(
daytona_api_keyfield missing): https://github.com/FoundationAgents/OpenManus/issues/1232 - OpenManus model connection (OpenAI-style client with your
base_url): https://github.com/FoundationAgents/OpenManus/blob/main/app/llm.py - OpenManus main agent (tools, 20-step limit): https://github.com/FoundationAgents/OpenManus/blob/main/app/agent/manus.py
- OpenManus start file (
Enter your prompt:): https://github.com/FoundationAgents/OpenManus/blob/main/main.py - Windows 11 check for this guide, 14 November 2026:
git clone,uv venv --python 3.12,.venv\Scripts\activate, anduv pip install -r requirements.txtfinished without errors on main at commit3309bf4;python main.pystopped with thedaytona_api_keyerror until the two[daytona]lines were added. - uv install page (Mac and Windows installers): https://docs.astral.sh/uv/getting-started/installation/
- uv Python page (downloads Python when needed): https://docs.astral.sh/uv/guides/install-python/
- Git for Windows: https://git-scm.com/install/windows
- Git for macOS: https://git-scm.com/install/mac
- Groq OpenAI compatibility: https://console.groq.com/docs/openai
- Groq models: https://console.groq.com/docs/models
- Groq keys: https://console.groq.com/keys
- OpenAI spend limits: https://developers.openai.com/api/docs/guides/spend-limits
- Anthropic spend limits: https://platform.claude.com/docs/en/api/rate-limits
- This project's notes: safety checklist
Next
Continue to Files, the browser, and memory.