Report Abuse

Basic Information

Mentals AI is a developer-focused framework for creating and running LLM-driven agents defined as plain text markdown files with a .gen extension. The repo treats each .gen file as an executable agent that embeds instructions, control flow, and calls to built‚Äëin tools, enabling developers to express loops, branching, and state entirely in natural language instead of writing scaffolding code in Python or other languages. Its Agent Executor runs agents in a recursive loop where the model chooses the next instruction and manages context. The project includes examples of game generators, multi-agent interactions and content pipelines, and targets integration with OpenAI-compatible providers including optional Llama3 support. The repository also documents build and runtime requirements, a config.toml for API configuration, and work-in-progress components like a local vector memory branch and a web UI.

Links

App Details

Features
The README highlights several core features: executable .gen agent files with named instructions and directives (for example ## use: and ## input:), a recursive Agent Executor that implements loops and semantic control flow, and native tools for file handling, user input, running Python, executing bash, and an experimental short-term memory tool. Instruction-level working memory is preserved by default and can be limited or cleared using directives. The system supports multiple reasoning frameworks including ReAct, Tree of Thoughts, and Auto-CoT and allows linking frameworks together. Included examples live in the agents folder and demonstrate tasks like game generation, web scraping and content assembly. Build and run instructions show dependencies (libcurl, libfmt, poppler, optional pgvector) and a simple make, config.toml and run workflow.
Use Cases
Mentals AI helps developers prototype and operate complex AI agents without writing traditional orchestration code by encoding logic, memory and tool usage in readable markdown. Users can rapidly iterate agent behavior using plain language directives, experiment with different reasoning strategies, compose multi-step loops, and reuse instruction blocks across agents. Built-in tools and file handling simplify tasks like saving CSVs, fetching transcripts, or running shell commands while short-term memory enables cross-instruction context preservation. The framework makes it straightforward to build content pipelines, multi-agent workflows, or game generators demonstrated in example files, and lowers integration overhead by relying on an OpenAI-compatible API key and minimal build steps. Work-in-progress features include vector-backed memory and a web UI for managing agents.

Please fill the required fields*