Onit

OnIt

OnIt — the AI is working on the given task and will deliver the results shortly.

OnIt is an intelligent agent for task automation and assistance. It connects to private vLLM servers, OpenRouter.ai, and Ollama cloud for hosted models — and uses MCP tools for web search, file operations, and more. It also supports the A2A protocol for multi-agent communication.

Getting Started

1. Install

pip install onit

Or from source:

git clone https://github.com/sibyl-oracles/onit.git
cd onit
pip install -e ".[all]"

2. Setup

onit setup

The setup wizard walks you through configuring your LLM endpoint, API keys, and preferences. Secrets are stored securely in your OS keychain. Settings are saved to ~/.onit/config.yaml.

To review your configuration at any time:

onit setup --show

3. Run

onit

That's it. MCP tools start automatically, and you get an interactive chat with tool access.

CLI at a Glance

onit                                          # interactive terminal chat
onit setup                                    # configure LLM endpoint, API keys
onit resume [TAG_OR_ID]                       # continue a previous session
onit sessions                                 # list saved sessions

onit serve a2a                                # A2A protocol server (port 9001)
onit serve web                                # Gradio web UI (port 9000)
onit serve gateway [telegram|viber|auto]      # Telegram or Viber bot
onit serve loop "task" --period 60            # repeat a task on a timer

onit ask "what is the weather in Manila"      # send a task to a running A2A server

onit --rules                                  # load coding rules from ONIT.md
onit --rules path/to/rules.md                # load coding rules from a custom file
onit --container                              # run in a hardened Docker container
onit --sandbox                                # delegate code execution to a sandbox
onit --unrestricted                           # unrestricted host filesystem access

Configuration

onit setup is the recommended way to configure OnIt. It stores:

• Settings in ~/.onit/config.yaml (LLM endpoint, theme, ports, timeout)
• Secrets in your OS keychain (API keys, bot tokens)

You can also use environment variables or a project-level YAML config:

# Environment variables
export ONIT_HOST=https://openrouter.ai/api/v1
export OPENROUTER_API_KEY=sk-or-v1-...

# Or a custom config file
onit --config configs/default.yaml

Priority order: CLI flags > environment variables > ~/.onit/config.yaml > project config file.

Example config (configs/default.yaml)

serving:
  host: https://openrouter.ai/api/v1
  host_key: sk-or-v1-your-key-here   # or set OPENROUTER_API_KEY env var
  # model: auto-detected from endpoint. Set explicitly for OpenRouter:
  # model: google/gemini-2.5-pro
  think: true
  max_tokens: 32768   # max output tokens per response (fits any single answer)
  # Sampling parameters (all optional — sensible defaults apply):
  # temperature: 1.0
  # top_p: 0.95
  # top_k: 20
  # presence_penalty: 1.5
  # repetition_penalty: 1.0

verbose: false
timeout: 600
sandbox: false

web_port: 9000
a2a_port: 9001

theme: white         # or "dark"
topic: ~             # default topic context, e.g. "machine learning"
template_path: ~     # custom prompt template YAML
documents_path: ~    # local documents directory
data_path: ~         # working directory for file operations (default: system temp)

mcp:
  servers:
    - name: PromptsMCPServer
      url: http://127.0.0.1:18200/sse
      enabled: true
    - name: ToolsMCPServer
      url: http://127.0.0.1:18201/sse
      enabled: true

Sampling parameters

Sampling parameters (temperature, top_p, top_k, min_p, presence_penalty, repetition_penalty) are set in configs/default.yaml under serving:. They are not exposed as CLI flags to keep the command line clean.

Recommended parameters for Qwen3.5:

| Mode | Use case | temperature | top_p | top_k | presence_penalty | |------|----------|:---:|:---:|:---:|:---:| | Thinking (think: true) | General | 1.0 | 0.95 | 20 | 1.5 | | Thinking (think: true) | Precise coding | 0.6 | 0.95 | 20 | 0.0 | | Instruct (no think) | General | 0.7 | 0.8 | 20 | 1.5 | | Instruct (no think) | Reasoning | 1.0 | 1.0 | 40 | 2.0 |

Set repetition_penalty: 1.0 in all cases.

CLI Reference

Interactive chat (default)

onit [OPTIONS]

Starts an interactive terminal chat with tool access. MCP servers start automatically.

| Flag | Description | Default | |------|-------------|---------| | --config FILE | Path to YAML configuration file | configs/default.yaml | | --host URL | LLM serving host URL. Overrides config and ONIT_HOST | — | | --model NAME | Model name. Skips auto-detection from endpoint | — | | --verbose | Enable verbose logging | false | | --think | Enable thinking/reasoning mode (CoT) | false | | --no-stream | Disable token streaming | false | | --show-logs | Show tool execution logs | false | | --rules [FILE] | Load a coding-rules .md file to guide the agent's behaviour. Defaults to ONIT.md in the current directory when no path is given | ONIT.md | | --resume TAG_OR_ID | Resume a previous session by tag, UUID, or last | — | | --sandbox | Delegate code execution to an external MCP sandbox provider | false | | --unrestricted | Unrestricted host filesystem access (trusted environments only) | false | | --container | Run the entire OnIt process inside a hardened Docker container | false | | --mcp-sse URL | Add an external MCP server (SSE transport, repeatable) | — | | --mcp-server URL | Add an external MCP server (Streamable HTTP transport, repeatable) | — |

onit setup

Interactive setup wizard. Configures the LLM endpoint, API keys, and preferences. Stores settings in ~/.onit/config.yaml and secrets in the OS keychain.

onit setup           # run the wizard
onit setup --show    # print current configuration

onit sessions

List and manage saved sessions.

onit sessions                          # list recent sessions (default: 20)
onit sessions --limit 50               # list up to 50 sessions
onit sessions --tag abc123 "my-chat"   # tag a session for easy recall
onit sessions --rebuild                # rebuild session index from JSONL files
onit sessions --clear                  # delete all session history

onit resume

Resume a previous session by tag or UUID.

onit resume              # resume the most recent session
onit resume my-chat      # resume by tag
onit resume abc123       # resume by session UUID prefix

Equivalent to onit --resume TAG_OR_ID.

onit ask

Send a single task to a running OnIt A2A server and print the response. Useful for scripting, pipelines, or one-shot queries without starting a local agent.

onit ask "what is the weather in Manila"
onit ask "summarize this document" --file report.pdf
onit ask "describe this image" --image photo.jpg
onit ask "write a script" --server http://192.168.1.10:9001

| Argument / Flag | Description | Default | |-----------------|-------------|---------| | task (positional) | Task to send to the server | required | | --file PATH | File to upload along with the task | — | | --image PATH | Image file for vision processing (model must be a VLM) | — | | --server URL | A2A server URL | http://localhost:9001 |

onit serve

Run OnIt in a persistent server or daemon mode. All serve modes run indefinitely until interrupted (Ctrl+C).

onit serve a2a

Run OnIt as an A2A protocol server so other agents or clients can send tasks.

onit serve a2a                 # listen on port 9001 (default)
onit serve a2a --port 9100     # custom port

| Flag | Description | Default | |------|-------------|---------| | --port PORT | A2A server port | 9001 (or a2a_port in config) |

The agent card is available at http://localhost:9001/.well-known/agent.json.

Send a task from another agent (Python A2A SDK):

from a2a.client import ClientFactory, create_text_message_object
from a2a.types import Role
import asyncio

async def main():
    client = await ClientFactory.connect("http://localhost:9001")
    message = create_text_message_object(role=Role.user, content="What is the weather?")
    async for event in client.send_message(message):
        print(event)

asyncio.run(main())

onit serve web

Launch the Gradio web chat UI.

onit serve web                 # open on port 9000 (default)
onit serve web --port 9500     # custom port

| Flag | Description | Default | |------|-------------|---------| | --port PORT | Web UI port | 9000 (or web_port in config) |

Supports optional Google OAuth2 authentication — see [docs/WEBAUTHENTICATION.md](docs/WEBAUTHENTICATION.md).

onit serve gateway

Run OnIt as a Telegram or Viber bot. Configure bot tokens via onit setup or environment variables.

onit serve gateway                                      # auto-detect from env vars
onit serve gateway telegram                             # Telegram bot
onit serve gateway viber --webhook-url https://...      # Viber bot

| Argument / Flag | Description | Default | |-----------------|-------------|---------| | gateway_type (positional) | telegram, viber, or auto | auto | | --webhook-url URL | Public HTTPS URL for Viber webhook (or set VIBER_WEBHOOK_URL) | — | | --port PORT | Local port for Viber webhook server | 8443 (or viber_port in config) |

Required environment variables (set via onit setup or export):

• Telegram: TELEGRAM_BOT_TOKEN
• Viber: VIBER_BOT_TOKEN, VIBER_WEBHOOK_URL

Install gateway dependencies if not using [all]:

pip install "onit[gateway]"

onit serve loop

Repeat a task on a configurable timer. Useful for monitoring, polling, or autonomous scheduled work.

onit serve loop "check the weather in Manila" --period 60
onit serve loop "summarize today's news" --period 3600

| Argument / Flag | Description | Default | |-----------------|-------------|---------| | task (positional) | Task to execute repeatedly | required | | --period SECONDS | Seconds between iterations | 10 (or period in config) |

Coding Rules (--rules)

--rules injects a Markdown rules file into the agent's system prompt, giving it explicit coding guidelines to follow throughout the session. This is the primary way to improve and customise the agent's coding behaviour.

onit --rules                       # load ONIT.md from the current directory (default)
onit --rules path/to/RULES.md      # load a custom rules file
onit --rules --host http://localhost:8000/v1   # combine with any other flags

When no file is specified, --rules reads ONIT.md in the current working directory. The file contents are wrapped in a `` block and prepended to the system prompt:

You are an expert coding agent.
Follow these rules precisely when writing, reviewing, or modifying code.

... contents of ONIT.md ...

Creating your own rules file:

Place an ONIT.md at the root of your project and run onit --rules. The agent will follow your rules automatically for every task in that session.

Example ONIT.md:

## Style
- Use snake_case for all identifiers.
- Maximum line length: 88 characters.

## Tests
- Every public function must have a test.
- Tests must assert the business intent, not just the return value.

## Safety
- Never silence exceptions. Log and re-raise.
- Validate all external inputs at the boundary.

With a vLLM or OpenRouter backend:

onit --rules --host http://localhost:8000/v1
onit --rules --host https://openrouter.ai/api/v1 --model google/gemini-2.5-pro

With Ollama cloud:

onit --rules --host https://api.ollama.com --model glm-5.1:cloud

Isolation Modes

OnIt offers three isolation levels. They can be combined (e.g. --container --sandbox).

--sandbox

Delegates individual code-execution tool calls to an external MCP sandbox provider. Complementary to --container.

onit --sandbox
onit --container --sandbox   # defense in depth

Requires an MCP server that provides sandbox tools (sandbox_run_code, sandbox_install_packages, sandbox_stop). Set sandbox: true in config.yaml to enable by default.

--container

Runs the entire OnIt process inside a hardened Docker container so a breach cannot reach the host OS.

onit --container                                          # interactive terminal in container
onit --container serve web                                # web UI, port 9000 published
onit --container serve a2a --port 9100                    # A2A server on custom port
onit --container --container-gpus all                     # NVIDIA GPU pass-through
onit --container --container-mount "$HOME/docs:/home/onit/documents:ro" \
  serve web                                               # expose host path read-only
onit --container --sandbox                                # combine with per-tool sandboxing

The first run auto-builds the onit:local image from the repo Dockerfile. Subsequent runs reuse the image.

Container sub-flags:

| Flag | Description | |------|-------------| | --container-gpus SPEC | NVIDIA GPU pass-through (e.g. all, "device=0,1"). Requires NVIDIA Container Toolkit. | | --container-mount HOST:CONTAINER[:ro] | Extra bind mount. Repeatable. Prefer :ro. | | --container-memory SIZE | Hard memory cap (e.g. 16g). Default: unlimited. | | --container-shm-size SIZE | /dev/shm size (default: 4g). Raise for PyTorch DataLoader. | | --container-tmp-size SIZE | /tmp tmpfs size (default: 16g). Backed by host RAM. |

Isolation posture: non-root user, read-only rootfs, --cap-drop=ALL, no host mounts by default, outbound network allowed.

What crosses the boundary:

| Resource | Default behavior | |---|---| | ~/.onit/config.yaml | Bind-mounted read-only | | Host keychain secrets | Passed as ephemeral env vars | | Session data | Named volume onit-data (writable, persistent) | | Ports | Published only for the active mode | | Host filesystem | Nothing beyond config/secrets unless --container-mount is set |

Published ports by mode:

| Mode | Default port | Override | |---|---|---| | (terminal) | — (no ports) | — | | serve web | 9000:9000 | --port | | serve a2a | 9001:9001 | --port | | serve gateway viber | 8443:8443 | --port |

See [docs/DOCKER.md](docs/DOCKER.md) for full details.

--unrestricted

Runs OnIt with lifted filesystem restrictions on the host — the agent can read/write any path, use any working directory, and install packages freely (pip, apt, brew, etc.). Use only in trusted, isolated environments.

onit --unrestricted

Catastrophic commands (disk wipe, reboot, kernel module loading) are always blocked regardless of this flag.

MCP Tool Integration

MCP servers start automatically. Tools are auto-discovered and available to the agent.

| Server | Description | |--------|-------------| | PromptsMCPServer | Prompt templates for instruction generation | | ToolsMCPServer | Web search, bash commands, file operations, and document tools |

Connect to additional external MCP servers:

onit --mcp-sse http://localhost:8080/sse
onit --mcp-server http://localhost:8080/mcp

Model Serving

Private vLLM

Serve models locally with vLLM:

CUDA_VISIBLE_DEVICES=0,1,2,3 v

…

## Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [sibyl-oracles](https://github.com/sibyl-oracles)
- **Source:** [sibyl-oracles/onit](https://github.com/sibyl-oracles/onit)
- **License:** Apache-2.0

Install and usage instructions live in the source repository linked above.