Agent-K, he's the most capable MiB agent there ever was. - gabereiser/agent-k
gabereiser/agent-k
Agent-K, he's the most capable MiB agent there ever was. - gabereiser/agent-k
0 stars · 0 forks · TypeScript
README
Agent K
Agent K - MiB's most capable agent.
He has access to all the tools, any skills, supports the .agent dir, can load multiple inference providers, multiple models, delegates to subagents
for tasks that are beneath him and will help plan your next project. Seriously capable. Full admin privileges and no confirmations so run
him on a leash in a docker container for your own safety.
Features:
Delegation to subbordinates
Planning
Expansion
Tool Calling
MCP
Reverse Engineering
Multiple Providers
Multiple Models per Provider
Multiple Models per Layer
Ralph Loop built in for long agentic tasks
A2A Server: Expose Agent-K as an A2A-compliant agent for multi-agent ecosystems
To install dependencies:
bun install
To build the binary:
make
To run:
./bin/agent -p "What happened to the Light of Zartha?"
Starting the A2A Server:
Start on default port 3000
./bin/agent --serviceSpecify a custom port
./bin/agent --service --port 8080Use environment variables for port
A2A_PORT=8080 ./bin/agent --service
For optimal performance, install Agent K to ~/bin/agent and include ~/bin in your $PATH
Written from scratch but loosely based on OpenCode's inner workings and the @ai-sdk from Vercel.
That is, until Agent K was able to self-program. Then it was all Agent K.
(WARNING)
Depending on the provider and model you choose to run Agent K with, you'll get drastically different results. I highly recommend a high reasoning model for the planner layer, a solid MoE model for the management layer, and a strong coder model for the worker layer. Replace coder model with writer model if that's your thing.
There are no safe-guards against deleting files, data, databases, operating systems, hard drives, or anything else available via bash. !Use at your own risk!source
