Claude code is terribly inefficient with tokens. It can easily use 50 bucks worth of tokens in an hour.
I've written my version of this tool for coding, and generally use 50$ a day, while retaining a lot of productivity gains. Supports Anthropic, Gemini, Openrouter, Deepseek APIs. All of it is with your own API keys. Relies on aggressive prefix caching to keep the costs down. Can work with extremely long context windows through intelligent history truncation. With a correct prompt, it can keep working on your features for hours on end (Implementing, testing, committing changes automatically).
WARNING: this tool has no safety mechanisms. It'll execute whatever shell command claude will provide (By design).
Termineer was written in Rust (By Claude code at the start, then by itself), over the course of a week.
I'm using this as a platform to experiment with multi-agent workflows, Communication mechanisms, multi-modal interactions (think Manus). Supports basic MCP server interaction as well.
Termineer does not use native tool calling APIs, instead it parses conversation state internally. This allows me to do some interesting things, such as the LLM deciding when to abort shell execution, multiple communicating agents, and other fun stuff.
I'm interested in feedback, both positive and negative is welcome!
Suggestion: Try this with Superwhisper. It has been my workflow for last few weeks and it's extremely productive!
Whenever I tried Aider it seemed it wanted to be a coding helper. This is more like self-directed agent. You just give it a task, and come back half an hour later to see what was performed.
Claude code is terribly inefficient with tokens. It can easily use 50 bucks worth of tokens in an hour.
I've written my version of this tool for coding, and generally use 50$ a day, while retaining a lot of productivity gains. Supports Anthropic, Gemini, Openrouter, Deepseek APIs. All of it is with your own API keys. Relies on aggressive prefix caching to keep the costs down. Can work with extremely long context windows through intelligent history truncation. With a correct prompt, it can keep working on your features for hours on end (Implementing, testing, committing changes automatically).
WARNING: this tool has no safety mechanisms. It'll execute whatever shell command claude will provide (By design). Termineer was written in Rust (By Claude code at the start, then by itself), over the course of a week.
I'm using this as a platform to experiment with multi-agent workflows, Communication mechanisms, multi-modal interactions (think Manus). Supports basic MCP server interaction as well.
Termineer does not use native tool calling APIs, instead it parses conversation state internally. This allows me to do some interesting things, such as the LLM deciding when to abort shell execution, multiple communicating agents, and other fun stuff.
I'm interested in feedback, both positive and negative is welcome!
Suggestion: Try this with Superwhisper. It has been my workflow for last few weeks and it's extremely productive!
How does this compare to Aider though?
Whenever I tried Aider it seemed it wanted to be a coding helper. This is more like self-directed agent. You just give it a task, and come back half an hour later to see what was performed.