An LLM Agent to help you from and within the terminal.
Recommended way of using the terminal-agent is by compiling the code into an agent binary, and placing it within your $PATH. This can be achieved through executing the Taskfile command task install which builds the binary and copies it to /usr/local/bin/agent.
For majority of cases, do check out the Taskfile.dist.yaml as it has most relevant tasks / receipies.
Intended for general questions.
agent ask what is a file descriptor?If you like to see characters appearing in terminal, add --stream flag.
Life hack: If you set alias alias aa="agent ask" you'll have a quick shortcut to ask questions from terminal. It's quicker than opening browser to search! Execute task install:alias for auto-setup.
Generates code for execution, and asks for execution permissions.
agent task list files 3 dirs upNote: Since more testing is required, there are only a few methods allowed.
Given that a few providers and their models are supported, there's a rudamentry configuration support. Whole config
To change the provider (to $PROVIDER) and model (to $MODEL) do
$ agent config set provider $PROVIDER
$ agent config set model $MODELAnthropic is best known for their Sonnet and Haiku models. From subjective experience, they're one of the best commercially available models. One can use them either directly, or via Amazon Bedrock.
To use directly, check out their Getting Started page. Primarely what you need is to create an account, then create an API key, and set it as your env variable, e.g.
export ANTHROPIC_API_KEY=sk-ant-api00-something-somethingOnce you have the key set, then configure the terminal-agent to use your preferred provider and model. To see all available models, check User Guide -> Models.
Setting the provider can be done using task run:set:anthropic or directly
$ agent config set provider anthropic
$ agent config set anthropic.claude-3-haiku-20240307Terminal Agent uses the Bedrock Converse API and thus supports the same models. These include models from Anthropic (newer than the Instance v1) and Llama 3+ version.
To use these models you need to make sure that you have valid aws credentials set. We support having auth via local credentials or via SSO.
Setting bedrock provider with Anthropic Haiku model can be done with task run:set:bedrock or directly
$ agent config set provider bedrock
$ agent config set anthropic.claude-3-haiku-20240307-v1:0 Perplexity AI has an "in development" functionality to provide access to models via their API. To read more on this visit their (documentation)[https://docs.perplexity.ai/guides/getting-started].
To use agent with Perplexity AI, you need to set PERPLEXITY_KEY env variable with your individual key, and then set the provider to perplexity.
Setting perplexity provider with llama-3.1-8b-instruct model can be done with task run:set:perplexity or directly
$ agent config set provider bedrock
$ agent config set llama-3.1-8b-instructTo use agent with OpenAI, you need to set the OPENAI_API_KEY environment variable with your individual key, and then set the provider to openai.
Setting openai provider with gpt-4o-mini model can be done with task run:set:openai or directly
$ agent config set provider openai
$ agent config set model gpt-4o-mini- ask: Ask general questions via terminal (all models)
- task: Ability to propose and execute a command (expect Perplexity)
- stream: Ask only (except Perplexity)
- unix: Designs and evaluates a unix command
- python: Designs and evaluates a python command
- git: Designs and evaluates a git command
The agent is not you.
You interact with the agent.
The agent knows about the tools.
The agent can use tools.
Tools are independent from the agent.
Each request can have its own configuration, and there are defaults provided.
To see default execute agent config get all.
To override defaults, use config set command, e.g.
$ agent config set model ai21.jamba-1-5-mini-v1:0The config is stored in json file at $HOME/.config/terminal-agent/config.json.
- Taskfile, instead of Makefile. See Taskfile installation
- Docker, needed for running integ tests