Mana-Agent

Think beyond prompts. Build beyond boundaries.

Your AI-powered execution workspace.
One goal. Multiple agents. Real tools. Verified results.

English · العربية · বাংলা · Brezhoneg · Bosanski · Dansk · Deutsch · Español · فارسی · Français · Ελληνικά · Italiano · 日本語 · 한국어 · Norsk · Polski · Русский · ไทย · Türkçe · Українська · Tiếng Việt · 简体中文 · 繁體中文

Get Started · Capabilities

What if your AI could do more than answer?

Most AI tools stop at the answer.

You ask a question. They generate a response. Then you're left to figure out what to do next.

Mana-Agent is built for what happens after the answer.

It can understand your goals, explore your environment, plan actions, coordinate specialized agents, use real tools, and verify outcomes.

Not just for writing code.

For researching, building, debugging, managing, automating, and solving complex tasks across connected systems.

For unresolved coding bugs, the Main Model can select bounded, version-aware web/GitHub research after local investigation and return sourced findings to the same coding task. Both coding engines use the existing Coding Activity stream. See coding research configuration and recovery.

You define the destination. Mana-Agent helps navigate the journey.

Meet Mana-Agent

Mana-Agent is an open-source, Python-based AI agent orchestration platform combining an autonomous execution loop, specialized agents, persistent task management, supervised tool execution, and extensible integrations.

It brings multiple capabilities into one coordinated environment:

Understand: Analyze repositories, documents, APIs, and available evidence before taking action.
Decide: Select suitable models, agents, tools, and execution strategies.
Execute: Use coding tools, browser automation, external services, and connected systems.
Coordinate: Distribute work among specialized agents while tracking dependencies and progress.
Remember: Reuse authorized context, decisions, and lessons across relevant tasks.
Verify: Evaluate actual outcomes rather than treating tool completion as success.
Continue: Replan, recover, or request human intervention when the goal remains unresolved.

The objective is simple:

Make AI useful not only for thinking about work, but for carrying work through to a trustworthy result.

What can Mana-Agent do?

Your challenge	What Mana-Agent can help with
"I don't understand this codebase."	Explore architecture, dependencies, execution flows, and relevant files.
"Something is broken."	Investigate symptoms, identify likely causes, propose fixes, and run verification.
"I want to build something complex."	Decompose goals, coordinate coding agents, manage dependencies, and review implementation.
"This task has too many steps."	Track progress through the autonomous loop, TaskBoard, and execution supervisor.
"I need information from an API."	Inspect API documentation, validate requests, and execute supported operations.
"I keep repeating this workflow."	Create automations or teach repeatable actions using Teach Mode.
"I want AI to work across tools."	Integrate browser, Git, SSH, email, Telegram, protocols, and external services.
"What happened during execution?"	Inspect activity, tools, decisions, errors, checkpoints, and artifacts.

One interface. Many kinds of work.

> Explain how authentication works in this repository.

> Find why the latest deployment failed.

> Plan and implement a new feature, then run its tests.

> Read this API documentation and help me call the right endpoint.

> Analyze these files and produce an actionable report.

> Break this project into independent tasks and track progress.

> Review previous mistakes before attempting this operation again.

Mana-Agent is designed to support both focused questions and multi-step workflows.

Quick Start

1. Install

Requires Python 3.10–3.14.

Recommended installation with pipx:

pipx install git+https://github.com/manaplatform/mana-agent.git

Or install directly from the repository:

git clone https://github.com/manaplatform/mana-agent.git
cd mana-agent

python3 -m venv .venv
source .venv/bin/activate

python -m pip install -e .

2. Configure

mana-agent --configure

Configure your model provider, credentials, model preferences, and supported capabilities.

3. Start

cd /path/to/your/project
mana-agent

You're ready.

Execution & Approval Modes

Mana-Agent supports three execution approval modes:

default: Interactive approval behavior (prompts for mutating/consequential operations).
auto_approve_safe: Automatically approves explicitly classified safe operations (e.g., non-destructive shell, test runs, linters, local git queries); prompts for destructive actions.
full_access: Automatically approves all supported operations without interactive confirmation prompts while preserving operating-system boundaries and security controls.

# Enable Full Access Mode via CLI
mana-agent --full-access
mana-agent --approval-mode full_access

# Or set via environment variable
export MANA_AGENT_APPROVAL_MODE=full_access

# Or configure in ~/.mana/config.toml
# [execution]
# approval_mode = "full_access"

Start with a question about your repository, ask for an investigation, or describe something you want to accomplish.

Your first task doesn't need to be complicated.

Try:

Analyze this project and explain its architecture,
main components, dependencies, and potential risks.

4. Explore the dashboard

mana-agent dashboard --root-dir .

The optional dashboard provides another way to interact with Mana-Agent, inspect activity, and manage workflows.

Live agent activity

Textual chat, Rich terminal chat, and the dashboard show observable operations,
agent explanations when provided, and supported public provider summaries while
work is running. Activity and Coding Activity stay inside their existing turn
containers. Concurrent agents retain their own task, attempt, and operation
identity; backend completion and verified goal completion are separate states.

Use /activity minimal, /activity normal (default), or /activity verbose to
persist MANA_ACTIVITY_VISIBILITY through the existing user configuration.
Activity appears inside Main Activity and Coding Activity, including internal
coding operations. /activity shows recent activity in terminal chat. In the
TUI, Ctrl+End returns to latest activity. Visibility never
reveals private reasoning, credentials, prompts, or raw invocation arguments.

During an active TUI or dashboard turn, submit text to steer that execution.
The user row appears pending immediately and follows recorded delivery and
acknowledgement. Dashboard steering is text-only and requires the same runtime
to own the live session. Approval requirements remain unchanged. Plan Mode
activity describes planning and artifact saving, not completed implementation.

See activity architecture, limitations, regression rules, and user verification.

License

Mana-Agent is open source under the MIT License.

Mana-Agent
From intention to execution. From execution to evidence.

Get Started · GitHub · Report an Issue · Explore the Docs

Have a goal in mind? Start with Mana.