OpenManus for Local LLMs

OpenManus Installation and Usage Guide

OpenManus repository on GitHub.

Install uvicorn for running the FastAPI server.

1$ curl -LsSf https://astral.sh/uv/install.sh | sh
2uv -h
3uv -V

Essential uv Commands

Category	Command	Description
Environments	uv venv	Creates a new virtual environment in the .venv folder.
	source .venv/bin/activate	Activates the environment (macOS/Linux).
	.venv\Scripts\activate	Activates the environment (Windows).
	deactivate	Exits the currently active environment.
Project Management	uv init	Sets up a new project with a pyproject.toml file.
	uv add `<package>`	Adds a dependency and automatically updates the environment.
	uv sync	Synchronises the environment with the lockfile (installs/removes).
	uv run `<script>`	Runs a script without needing to manually activate the env.
Pip Interface	uv pip install -r reqs.txt	Fast alternative to pip install for existing requirement files.
	uv pip tree	Displays a visual tree of all installed dependencies.
Python Versions	uv python list	Shows available Python versions you can install.
	uv python install 3.12	Downloads and installs a specific Python version automatically.

Clone the OpenManus repository and navigate to the project directory.

1git clone https://github.com/FoundationAgents/OpenManus.git
2cd OpenManus

Install the required dependencies using pip.

1uv pip install -r requirements.txt

Configure toml file

OpenManus uses a config.toml file to specify the LLMs and their configurations. You can create this file in the root directory of the project.

 1# This configuration is for local LLMs Ollama and qwen2.5
 2
 3# OpenManus configuration - config/config.toml
 4
 5[llm]
 6# Provider type for OpenManus: use "ollama" for local Ollama
 7api_type = "ollama"
 8# Base URL for the Ollama OpenAI-compatible API (must include /v1)
 9base_url = "http://localhost:11434/v1"
10# Local Ollama does not require an API key
11api_key = "ollama"
12# Default model for code/text tasks
13model = "qwen2.5-coder"
14# Generation defaults
15max_tokens = 4096
16temperature = 0.0
17
18[llm.vision]
19# Vision model for screenshot/image analysis tasks
20api_type = "ollama"
21base_url = "http://localhost:11434/v1"
22api_key = "ollama"
23model = "llama3.2-vision"
24max_tokens = 4096
25temperature = 0.0

Once installed and configured correctly, it can be run with python main.py.
There is no web UI, like Vscode Copilot Chat — you interact with it purely through the terminal. You can provide prompts directly as command-line arguments.

Issues Installing OpenManus

requirements.txt does not work, needs corrections to include missing packages and fix names. Following fixes were applied on 2026-05-10:

 1aiofiles~=24.1.0
 2docker~=7.1.0
 3mcp~=1.5.0
 4boto3~=1.37.18
 5huggingface-hub~=0.29.2
 6# --- OpenManus requirements (2026-05-10) ---
 7# The following block is the correct set of requirements for OpenManus, matching setup.py and avoiding dependency conflicts.
 8pydantic~=2.10.4
 9openai>=1.58.1,<1.67.0
10tenacity~=9.0.0
11pyyaml~=6.0.2
12loguru~=0.7.3
13numpy
14datasets>=3.2,<3.5
15html2text~=2024.2.26
16gymnasium>=1.0,<1.2
17pillow>=10.4,<11.2
18browsergym~=0.13.3
19uvicorn~=0.34.0
20unidiff~=0.7.5
21browser-use~=0.1.40
22googlesearch-python~=1.3.0
23aiofiles~=24.1.0
24pydantic_core>=2.27.2,<2.28.0
25colorama~=0.4.6
26structlog~=25.5.0
27tiktoken~=0.9.0
28baidusearch~=1.0.3
29duckduckgo_search~=7.5.3
30daytona-sdk
31
32# --- Optional/extra packages (uncomment if needed) ---
33# fastapi~=0.115.11
34# tiktoken~=0.9.0
35# playwright~=1.51.0
36# docker~=7.1.0
37# pytest~=8.3.5
38# pytest-asyncio~=0.25.3
39# mcp~=1.5.0
40# httpx>=0.27.0
41# tomli>=2.0.0
42# boto3~=1.37.18
43# requests~=2.32.3
44# beautifulsoup4~=4.13.3
45# crawl4ai~=0.6.3
46# baidusearch~=1.0.3
47# duckduckgo_search~=7.5.3
48# huggingface-hub~=0.29.2
49# setuptools~=75.8.0
50
51# --- Previous requirements (commented out for reference) ---
52# pydantic~=2.10.6
53# openai~=1.66.3
54# tenacity~=9.0.0
55# pyyaml~=6.0.2
56# loguru~=0.7.3
57# numpy
58# datasets~=3.4.1
59# fastapi~=0.115.11
60# tiktoken~=0.9.0
61# html2text~=2024.2.26
62# gymnasium~=1.1.1
63# pillow~=11.1.0
64# browsergym~=0.13.3
65# uvicorn~=0.34.0
66# unidiff~=0.7.5
67# browser-use~=0.1.40
68# googlesearch-python~=1.3.0
69# baidusearch~=1.0.3
70# duckduckgo_search~=7.5.3
71# aiofiles~=24.1.0
72# pydantic_core~=2.27.2
73# colorama~=0.4.6
74# playwright~=1.51.0
75# docker~=7.1.0
76# pytest~=8.3.5
77# pytest-asyncio~=0.25.3
78# mcp~=1.5.0
79# httpx>=0.27.0
80# tomli>=2.0.0
81# boto3~=1.37.18
82# requests~=2.32.3
83# beautifulsoup4~=4.13.3
84# crawl4ai~=0.6.3
85# huggingface-hub~=0.29.2
86# setuptools~=75.8.0

Issue with installed packages

ModuleNotFoundError: No module named 'structlog' — fixed by adding structlog~=25.5.0 to requirements and installing it.
ModuleNotFoundError: No module named 'tiktoken' — fixed by adding tiktoken~=0.9.0 to requirements and installing it.
ModuleNotFoundError: No module named 'baidusearch' — fixed by adding baidusearch~=1.0.3 to requirements and installing it.
ModuleNotFoundError: No module named 'duckduckgo_search' — fixed by adding duckduckgo_search~=7.5.3 to requirements and installing it.
ModuleNotFoundError: No module named 'daytona' — fixed by adding daytona-sdk to requirements and installing it.

OpenManus Setup — Full Summary

Environment

OS: Ubuntu Linux
Python: 3.12.3 (via uv venv --python 3.12)
Package manager: uv 0.11.12
LLM: Ollama running qwen2.5-coder locally

Issues Found & Fixes Applied

1. `daytona_api_key` required field crash

Error:

1pydantic_core.ValidationError: 1 validation error for DaytonaSettings
2daytona_api_key — Field required

Cause: DaytonaSettings.daytona_api_key was a required str. Even when no [daytona] section existed in config, a bare DaytonaSettings() was instantiated.

Fix in app/config.py:

Changed daytona_api_key: str → daytona_api_key: Optional[str] = None
Changed daytona_settings = DaytonaSettings() → daytona_settings = None when the config section is absent

2. Wrong Ollama `base_url` in config

Cause: http://localhost:11434 is Ollama’s native API. OpenManus uses the OpenAI-compatible endpoint which requires /v1.

Fix in config/config.toml:

1base_url = "http://localhost:11434/v1"   # was: http://localhost:11434
2model = "qwen2.5-coder"                  # was: qwen2.5

3. Missing packages not in `requirements.txt`

Three packages were imported by the codebase but absent from requirements:

Package	Where used	Error
`structlog`	`app/logger.py`	`ModuleNotFoundError: No module named 'structlog'`
`tiktoken`	`app/llm.py`	`ModuleNotFoundError: No module named 'tiktoken'`
`baidusearch`	`app/tool/search/baidu_search.py`	`ModuleNotFoundError: No module named 'baidusearch'`
`duckduckgo_search`	`app/tool/search/`	Would have failed next
`daytona-sdk`	`app/daytona/tool_base.py`	`ModuleNotFoundError: No module named 'daytona'`

Fix: Added all five to requirements.txt and installed them.

4. Wrong import name for Daytona SDK

Error:

1ModuleNotFoundError: No module named 'daytona'

Cause: The pip package is daytona-sdk but imports as daytona_sdk, not daytona.

Fix in app/daytona/tool_base.py and app/daytona/sandbox.py:

1# Before
2from daytona import Daytona, DaytonaConfig, Sandbox, SandboxState
3# After
4from daytona_sdk import Daytona, DaytonaConfig, Sandbox, SandboxState

5. Daytona module-level crash when not configured

Error:

1AttributeError: 'NoneType' object has no attribute 'daytona_api_key'
2# then after partial fix:
3DaytonaAuthenticationError: Authentication credentials not found

Cause: Both sandbox.py and tool_base.py construct DaytonaConfig(...) and Daytona(...) at module import time — not inside a function — so they crash immediately even if you never use Daytona.

Fix: Wrapped the initialization in both files with a None guard:

1# sandbox.py and tool_base.py
2daytona_settings = config.daytona
3if daytona_settings is not None:
4    daytona_config = DaytonaConfig(...)
5    daytona = Daytona(daytona_config)
6else:
7    daytona_config = None
8    daytona = None

Full Rebuild Steps (in order)

 1# 1. Delete old broken venv
 2rm -rf .venv
 3
 4# 2. Create fresh venv with Python 3.12 (uv downloads it if needed)
 5uv venv --python 3.12 .venv
 6
 7# 3. Install all dependencies
 8uv pip install -r requirements.txt
 9
10# 4. Install Playwright Chromium + system deps
11.venv/bin/playwright install chromium --with-deps

VSCode Configuration

Added to .vscode/settings.json:

1"python.defaultInterpreterPath": "${workspaceFolder}/.venv/bin/python",
2"python.terminal.activateEnvironment": true

This ensures VSCode always uses the .venv Python 3.12 interpreter for IntelliSense, linting, and integrated terminal.

Running OpenManus

1# Interactive mode
2.venv/bin/python main.py
3
4# With a prompt directly
5.venv/bin/python main.py --prompt "your task here"

Harmless Warnings (can ignore)

UserWarning: Valid config keys have changed in V2: 'underscore_attrs_are_private' has been removed — comes from a third-party library (browser-use) using an old Pydantic V1 config style; doesn’t affect functionality.
Daytona not configured, skipping Daytona initialization, expected, you don’t use Daytona.