OpenManus

OpenManus: Advanced

This page is the seventh of eight, and it covers six power features that OpenManus documents in its README, its example files, or its code. Try them only after the jobs in the Playbook feel routine, because each one gives the agent more reach or more of your money.

Checked 2026-11-14. Several of these come from the code rather than the README, and the page says so each time.

One job straight from the command line

What it is for: Starting a job without the Enter your prompt: question. You can then keep a request you trust in a text file and paste the same words each time.

How to turn it on: The start file accepts a --prompt option, which the code calls the "Input prompt for the agent". The README does not mention it. Inside agent-tryout/OpenManus, with the Python space on, type the line below.

python main.py --prompt "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."

A token cap for each run

What it is for: Stopping a run once it has sent a set number of tokens to the model, so a stuck job cannot keep spending.

How to turn it on: The settings loader reads a max_input_tokens line in the [llm] block. The code describes it as the "Maximum input tokens to use across all requests (None for unlimited)", but neither the README nor the example file shows it. Open config.toml and add one line under the other [llm] lines.

max_input_tokens = 200000

The number is only an example, and this guide did not test which value suits which job. The check runs before each request, so the request that would go over the cap is never sent. In a Windows 11 test with a tiny cap, the code retried for about 20 seconds and printed a line with Request may exceed input token limit. Then it ended the job normally. The [llm.vision] block takes the same value unless you give it its own.

Add-on tools through MCP

What it is for: Giving the agent extra tools from MCP servers. These are add-ons that other projects publish, such as ones that read a calendar or a database.

How to turn it on: The main agent reads a file named config/mcp.json and connects to every server listed there each time it starts. The project ships an example to copy.

cp config/mcp.example.json config/mcp.json

On Windows, the copy line is copy config\mcp.example.json config\mcp.json. The example lists one server reached by web address.

{
    "mcpServers": {
      "server1": {
        "type": "sse",
        "url": "http://localhost:8000/sse"
      }
    }
}

In the code, each server is either "sse" with a url, or "stdio" with a command and args that start a program on your computer. Replace the example with the server you want, following that server's own instructions.

The README also has a separate "MCP tool version", started with python run_mcp.py. That version connects to OpenManus's own server. In the code, that server gives the agent a bash tool that runs shell commands, a file editor, and a finish tool. The bash tool starts /bin/bash, so this guide could not confirm that it works on Windows. An open issue from 3 November 2026 also reports that a connection fails when a server's tool has no description.

To compare MCP servers with skills, and to see how other agents add them, read Skills and MCP.

Several agents on one job

What it is for: Splitting a bigger job into a plan with steps, and handing data and chart steps to a second agent built for them.

How to turn it on: The README calls this the "unstable multi-agent version" and starts it with python run_flow.py. To add the data analysis agent, change the line in the [runflow] block of config.toml from false to true.

[runflow]
use_data_analysis_agent = true

That agent needs Node.js 18 or newer and its own parts. Its README gives these lines for after Node is installed, which it does with nvm on a Mac or nvm-windows on Windows.

cd app/tool/chart_visualization
npm install

Its charts are saved as picture or web-page files under workspace\visualization. The whole run has a limit of 60 minutes, after which the code says the request "timed out after 1 hour".

A local model through Ollama

What it is for: Running the model on your own computer, so prompts and file contents do not go to a provider and each use costs nothing.

How to turn it on: The project ships an Ollama example file. Copy it over your settings file, then add the two [daytona] lines from Connect a model to its end again.

cp config/config.example-model-ollama.toml config/config.toml

On Windows, the copy line is copy config\config.example-model-ollama.toml config\config.toml, and that copy replaces the saved config.toml.

The example file still names llama3.2 and llama3.2-vision. Do not pull those.

Pull qwen3.5:4b when the computer has about 8 GB, or pull qwen3.8:27b when it has about 24 GB. Then write that same name on both the main model line and the vision model line. Leave the address at http://localhost:11434/v1, and leave the key set to ollama, because Ollama does not check it.

Ollama's documentation says its OpenAI-compatible address supports tools, and that agents need a context length of at least 64000 tokens. It also says Ollama starts with only 4k on computers with less than 24 GiB of graphics memory. You can raise that with the slider in the Ollama app's settings, or by starting Ollama with the line below. On a Mac or Linux, quit the Ollama app first, because ollama serve fails if the app is already running.

OLLAMA_CONTEXT_LENGTH=64000 ollama serve

On Windows, use the slider in the Ollama app instead, or quit the Ollama app and run $env:OLLAMA_CONTEXT_LENGTH=64000; ollama serve in PowerShell. Give it enough context for agent work covers both in more detail.

This guide did not test any local model with OpenManus. Issue reports show small local models failing to save files, and llama3.2-vision is not on the code's list of models that get screenshots.

A real box: a Daytona cloud sandbox

What it is for: Running the agent's commands and browser on a separate computer in the cloud, so your own files and your own Chrome are out of its reach.

How to turn it on: The project's Daytona page gives these steps. Make a key at app.daytona.io/dashboard/keys, and put it between the quote marks on the daytona_api_key line you added to config.toml. Then install the two parts the page names, using uv as the rest of this guide does, and start the sandbox version.

uv pip install daytona==0.21.8 structlog==25.4.0
python sandbox_main.py

In the code, this version switches off the local Python runner and file editor. Its browser, file, shell, and screenshot tools run inside the Daytona sandbox instead. It prints a VNC link where you can watch its browser, and the Daytona page says the VNC password is 123456 unless you set VNC_password in the [daytona] block.

Do not confuse this with the local [sandbox] section on page 4. A reading of the code on 14 November 2026 found that use_sandbox moves only the file editor into Docker, while the Python runner stays on your computer.

This guide did not try Daytona or check its prices. It also could not confirm that those pinned versions still install alongside the current parts list, and an issue from May 2026 reports the Daytona parts crashing at startup.

Shared advanced topics

These pages cover setups that work with any agent. Only the ones that apply to OpenManus are listed.

Sources

Opened 2026-11-14.

Next

Continue to OpenManus: Persona and rules.

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