OpenManus

OpenManus: Fixes

This page covers the problems people hit most often with OpenManus. It follows the Playbook, and each problem gives the symptom, the likely cause, and the fix.

Checked 2026-11-14. OpenManus has no troubleshooting page or FAQ of its own. So the fixes here come from its code, its README, its GitHub issues, and one Windows 11 check made for this guide. Where this guide could not confirm a fix, it says so.

Common problems

How to stop it right now

Symptom: It is doing something you did not ask for, such as opening a page you did not name or talking about a file outside workspace.

Likely cause: Nothing in OpenManus pauses for your yes or no, so it keeps going until the job ends, it reaches 20 steps, or you stop it.

Fix: Click in its window and press Control and C. The start file is written to print Operation interrupted. when you do this. You may see a block of error text as well, but either way the job has ended. If Control and C does nothing, close the terminal window. After a job that used the browser part, quit Chrome too, because this guide could not confirm that the connection to Chrome ends with OpenManus. Then check workspace and any folder it mentioned, and restore anything it changed from your backup. After it stops, use After it does something wrong if a file, account, or charge already changed.

The install stops with an error

Symptom: uv pip install -r requirements.txt ends with "No solution found when resolving dependencies" or "ResolutionImpossible", or a later python main.py stops with ModuleNotFoundError and a module name such as structlog.

Likely cause: Between January and May 2026, several GitHub reports showed the parts list asking for versions of crawl4ai and pillow that could not go together. Other reports showed structlog missing from the list. The list on main now pins pillow to a matching version and includes structlog. On 14 November 2026 it installed without errors on one Windows 11 PC. A ModuleNotFoundError can also mean the parts went into a different Python, because the Python space was not on.

Fix:

  1. Check that the line where you type starts with a name in parentheses. If it does not, turn the Python space on again (step 6 of the install).
  2. Get the newest code, then install again. The README does not explain updating, so these are git's and uv's own commands.
git pull
uv pip install -r requirements.txt
  1. If it still fails, copy the last few lines of the error and search the project's issues for them. Issue comments suggest editing the parts list by hand. This guide could not confirm that those edits are safe, so try one only if you are comfortable undoing it.

It stops at once with daytona_api_key

Symptom: python main.py prints a long block of text that ends with 1 validation error for DaytonaSettings, daytona_api_key, and Field required.

Likely cause: On the code checked on 14 November 2026, the settings loader requires a Daytona section in config.toml, even when you never use Daytona. The example settings file does not have one.

Fix: Open config.toml again and add these two lines at the very end, as in step 5 of Connect a model.

[daytona]
daytona_api_key = ""

Leave the quote marks empty. On one Windows 11 PC, this stopped the error.

PowerShell will not turn on the Python space

Symptom: On Windows, .venv\Scripts\activate fails with a message about running scripts being disabled.

Likely cause: Windows has a setting called the execution policy. Microsoft says its strictest level is the default for Windows clients when nothing else is set, and that level blocks every script, including the one that turns on the Python space.

Fix: The safer choice skips turning the space on at all. Python's documentation says you do not need to, as long as you run the Python inside it by its full path. uv's documentation adds that uv pip install finds a .venv folder in the current folder by itself. So, inside agent-tryout\OpenManus, run the install as before. Then start each job with the line below instead of python main.py.

.venv\Scripts\python.exe main.py

The other choice is the Python documentation's own fix, which changes the setting for your account only.

Set-ExecutionPolicy -ExecutionPolicy RemoteSigned -Scope CurrentUser

Microsoft says that level lets scripts made on your computer run, while scripts downloaded from the internet still need a signature. It also says the setting is a safety rule rather than a wall. To undo it later, run Set-ExecutionPolicy -ExecutionPolicy Undefined -Scope CurrentUser. If neither choice suits you, use Agent Zero instead.

The model gives errors

Symptom: Red lines that start with OpenAI API error: Error code: and a number, often repeated several times, then a long block that ends with RetryError.

Likely cause: The code tries each request up to six times, waiting longer each time, so one problem shows up as several lines. The number tells you which problem it is.

Fix: For 401, make a new key at console.groq.com/keys and paste it between the quote marks on both api_key lines. For 429, wait a few minutes and give it a smaller job, because Groq does not publish how much free use you get. For a model error, you must copy the name Groq lists, such as qwen/qwen3.8-27b, into both model lines. For a tools error, pick a model whose provider says it supports tool use.

It keeps going, or repeats itself

Symptom: The steps keep counting up (Executing step 7/20, Executing step 8/20, and so on) after the file is already saved, or the same thought appears again and again. You may see Agent detected stuck state or, at the end, Terminated: Reached max steps (20).

Likely cause: Before step 20, a job ends only when the model calls its finish tool, and an open GitHub issue shows models that do not. The code catches only replies that are exactly the same, and then it asks the model to try something new. Issue reports also show weaker models choosing no tool at all, shown as Manus selected 0 tools to use.

Fix: Press Control and C. Then give a smaller job that ends with a clear stopping point, such as telling it to stop when the file is saved. Every step costs tokens, so if this keeps happening, try a stronger model. Do not leave a run going that you are not watching.

It says it finished, but no file appears

Symptom: You see "Request processing completed." but the file is not in workspace.

Likely cause: Issue reports show some models writing the answer into the chat instead of saving it, and others saving it somewhere else. In one open issue, most of the reports involved small local models.

Fix: Name the file and the folder in your request, as every job in the Playbook does. Then search the newest file in the logs folder for the file name, to see where it was saved, or whether it was saved at all. If a small local model keeps failing, users in that issue reported better results with a larger one.

One job used far more tokens than expected

Symptom: Your provider's usage page shows a much bigger number than one short job should cost.

Likely cause: OpenManus sends the whole job so far with every step, up to 20 steps, and the browser part adds page text and screenshots. One user reported a single run that used 7,681,991 tokens. A reply guessed that pictures were the reason, but the cause was not confirmed.

Fix: Stay on a free plan with no card, or set a spend limit at the provider as in Keep it in a box. Keep the browser part off unless the job needs it. The code also has a token cap for each run, described in Advanced. The log lines that start with Token usage: show what each step used.

Screenshots seem to be ignored

Symptom: During a browser job, the agent acts as if it cannot see the page, or keeps trying the same click.

Likely cause: The code sends pictures only to models on a short built-in list, such as gpt-4o, and removes them for every other model without a warning. A GitHub issue about this was still open on 14 November 2026. openai/gpt-oss-120b is not on the list.

Fix: Give jobs that need only the text of a page, such as copying a price or a heading, and do not ask it to judge how a page looks. This guide could not confirm a setting that fixes it without editing the code.

Update it safely

The OpenManus README does not explain updating. These steps use Git's pull command and the same uv pip install line the README uses to install. Read the project's recent commits before you pull. This guide could not confirm that a past update reset a saved setting. config.toml is left out of Git, so a normal pull does not replace your key.

  1. Close OpenManus. Copy your settings, logs, and workspace before you update. Copy config.toml inside the config folder, the workspace folder, and the logs folder. Also copy mcp.json inside config if you made one. All of those sit inside agent-tryout\OpenManus on Windows, and agent-tryout/OpenManus on a Mac. A copy can live in Windows File History or Previous Versions, or in Mac Time Machine.

What you should see a second copy of config.toml, workspace, and logs, dated today.

  1. Open a terminal in the OpenManus folder. Turn the Python space on if you normally do. Read the notes, then run these two commands.
git pull
uv pip install -r requirements.txt

On Windows, if the space will not turn on, keep the install line above and start the test with .venv\Scripts\python.exe main.py, as earlier on this page. If Git stops and reports a conflict, stop there. Do not delete config.toml. This guide could not confirm a safe way to finish a conflicted pull.

What you should see Git reports updated files, or says the folder is already up to date, and the install ends without an error.

  1. Check the commit, move any old packing-list.txt out of workspace (the file editor refuses to create over an existing file), and run the packing-list job from page 2 with the browser part off.
git rev-parse --short HEAD

What you should see a short commit id, which can differ from 3309bf4 after an update, and packing-list.txt in workspace with nothing else new.

Remove it or start over

Symptom: You want a clean start, or you want OpenManus gone.

Likely cause: Not a fault. Everything OpenManus needs sits inside agent-tryout, apart from uv and the parts uv keeps in its own storage, such as the Browser Use download.

Fix:

uv cache clean
rm -r "$(uv python dir)"
rm -r "$(uv tool dir)"

Then remove the program itself. On a Mac, that is rm ~/.local/bin/uv ~/.local/bin/uvx, and on Windows it is these three lines.

rm $HOME\.local\bin\uv.exe
rm $HOME\.local\bin\uvx.exe
rm $HOME\.local\bin\uvw.exe

Removing uv also removes the Python it downloaded, so do this only if nothing else on your computer uses uv. If you installed Chromium with uvx browser-use install, this guide could not confirm where it went or whether these steps remove it.

Sources

Opened 2026-11-14.

Next

Continue to Advanced.

Checked 14 November 2026 · prices, limits, and versions change