Project case study
Mamba Agents — AI Agent Framework
AI/ML
An extensible AI-agent framework on pydantic-ai adding context compaction, token/cost tracking, tools, MCP, and observability
PythonPydantic AItiktokenJinja2OpenTelemetrytenacityMCPOllamavLLM
A simple, extensible AI-agent framework built as a thin wrapper around pydantic-ai. Mamba Agents handles context-window management, token tracking, cost estimation, and observability so you can focus on your agent's logic. Published to PyPI with docs at sequenzia.github.io/mamba-agents.
It ships built-in filesystem, glob, grep, and bash tools with security sandboxing, connects to MCP servers over stdio/SSE/HTTP, and works with local model backends out of the box.
Key capabilities:
- Automatic context compaction with five strategies (sliding window, summarize-older, selective pruning, importance scoring, hybrid) to keep conversations within token limits
- Cost visibility via tiktoken usage tracking and estimation across all interactions
- Jinja2 prompt templates with versioning and inheritance, retry logic and circuit breakers, and OpenTelemetry tracing hooks