Introducing: Loki! An all-in-one, batteries-included LLM CLI tool

Loki started out as a fork of the fantastic AIChat CLI, where I just wanted to give it first-class MCP server support. It has since evolved into a massive passion project that's a fully-featured tool with its own identity and extensive capabilities! My goal is to make Loki a true "all-in-one" and "batteries-included" LLM tool.

Check out the release notes for a quick overview of everything Loki can do!

What Makes Loki Different From AIChat?

• First-class MCP support, with support for both local and remote servers
  • Agents, roles, and sessions can all use different MCP servers and switching between them will shutdown any unnecessary ones and start the applicable ones
  • MCP sampling is coming next
• Comes with a number of useful agents, functions, roles, and macros that are included out-of-the-box
• Agents, MCP servers, and tools are all managed by Loki now; no need to pull another repository to create and use tools!
  • No need for any more *.txt files
• Improved DevX when creating bash-based tools (agents or functions)
  • No need to have argc installed: Loki handles all the compilation for you!
  • Loki has a --build-tools flag that will build your bash tools so you can run them exactly the same way Loki would
  • Built-in Bash prompting utils to make your bash tools even more user-friendly and flexible
• Built-in vault to securely store secrets so you don't have to store your client API keys in environment variables or plaintext anymore
  • Loki also will inject additional secrets into your agent's tools as environment variables so your agents can also use secrets securely
• Multi-agent support out-of-the-box: You can now create agents that route requests to other agents and use multiple agents together without them trampling all over each other's binaries
• Improved documentation for all the things!
• Simplified directory structure so users can share full Loki directories and configurations without massive amounts of data, or secrets being exposed accidentally

What's Next?

• MCP sampling support, so that MCP servers can send back queries for the LLM to respond to LLM requests. Essentially, think of it like letting the MCP server and LLM talk to each other to answer your query
• Give Loki a TUI mode to allow it to operate like claude-code, gemini-cli, codex, and continue. The objective being that Loki can function exactly like all those other CLIs or even delegate to them when the problem demands it. No more needing to install a bunch of different CLIs to switch between!
• Integrate with LSP-AI so you can use Loki from inside your IDEs! Let Loki perform function calls, utilize agents, roles, RAGs, and all other features of Loki to help you write code.