οΈ
Agent Pilot is an open source desktop app to create, manage, and chat with AI agents.
Using your own keys, models and local data.
Features multi-agent, branching chats with dozens of providers through LiteLLM.
Combine models from different providers under one context, and
configure their interaction with each other in a low-code environment.
Open Interpreter comes built-in as the applications code interpreter, let it run code to
do whatever you ask it to do.
Note
Right now the master branch is the development branch, for the latest stable release checkout the latest tag
How to migrate your data to 0.3.0
Copy your old database (data.db) to the new application folder before you start the app.
Platform | Downloads |
---|---|
Linux |
Mirror: AgentPilot_0.3.1_Linux_Portable.tar.gz |
Windows |
Mirror: AgentPilot_0.3.1_Windows_Portable.zip MD5: 2ce8ba0fa927bed01fac70dc576dd959 SHA1: 4e0d250d9e6e56e2f267797fcd125cb63711eab3 |
Building from source: How to build from source
$ pip install agentpilot
Python docs
How to use
Create a plugin
Create new agents, edit their configuration and organise them into folders.
Multi-member workflows can be saved as a composite agent and nested infinitely.
View, continue and delete previous workflow chats and organise them into folders.
Chats can be exported and imported as .
Seamlessly add multiple agents (or users), and configure how they interact with each other.
Agent pilot supports group chat natively, but can be modified with a plugin:
CrewAI (All workflow agents must be a CrewAI agent, or else native will be used)
Create a workflow plugin
Messages can be edited and resubmitted, and code can be edited and re-run.
Allowing for a more practical way to chat with your workflow.
Branching works with all plugins and multi-agent chats.
Manage a list of context blocks available to use in any agent system message.
Allowing reusability and consistency across multiple agents.
Block types:
- Text - A simple text block that can nest other blocks.
- Code - A code block that is executed and gets the output.
- Prompt - A prompt block that gets an LLM response.
Create, edit and delete tools, configure their parameters, code, language and environment.
Tools can be added to an Agent or used individually as a workflow component.
Agent Pilot supports the following plugin types:
- Agent - Override the default agent behaviour.
- Workflow - Override the default workflow behaviour.
These agent plugins are built-in and ready to use:
Open Interpreter
OpenAI Assistant
CrewAI Agent
Create an agent plugin
Open Interpreter is integrated into Agent Pilot, and can either be used standalone as a plugin
or utilised by any Agent or context block to execute code.
Code auto-run can be enabled in the settings, but use this with caution, you should always
understand the code that is being run, any code you execute is your own responsibility.
Try something like "Split this image into quarters" and see the power of Open Interpreter
Tasks are being reimplemented, coming soon!
Tasks can be recurring or scheduled to run at a later time with requests like "The last weekend of every month", or "Every day at 9am".
Still in development, coming soon.
Agents can be linked to a text-to-speech service, combine with a personality context block and make your agent come to life!
Supported TTS services:
- Amazon Polly
- Elevenlabs (expensive)
- FakeYou (celebrities and characters but too slow for realtime)
- Uberduck (celebrities and characters are discontinued)
Supported LLM providers using LiteLLM:
- Anthropic
- Mistral
- Perplexity AI
- Groq
- OpenAI
- Replicate
- Azure OpenAI
- Huggingface
- Ollama
- VertexAI Google
- PaLM API Google
- Voyage
- AWS Sagemaker
- AWS Bedrock
- Anyscale
- VLLM
- DeepInfra
- AI21
- NLP Cloud
- Cohere
- Together AI
- Cloudflare
- Aleph Alpha
- Baseten
- OpenRouter
- Custom API Server
- Petals
(Anthropic, Mistral, Perplexity, OpenRouter & OpenAI have been tested)
Contributions to Agent Pilot are welcome and appreciated. Please feel free to submit a pull request.
- Changing the config of an OpenAI Assistant won't reload the assistant, for now close and reopen the chat.
- Some others
- Be careful using auto run code and open interpreter, any chat you open, if code is the last message it will start auto running, I'll add a flag to remember if the countdown has been stopped.
- Logs are broken and need reimplementing.
- Flickering when response is generating and scrolled up the page.
- Sometimes the scroll position of the chat page jumps after response has finished.
- Windows exe must have console visible or it affects the streams
If you find this project useful please consider showing support by giving a star or leaving a tip :)
BTC:
ETH: