SAFLA hardens the self aware feedback loop before anything else

What SAFLA is

A feedback loop architecture that keeps a model of its own state, exposed three ways: an MCP server for Claude Code, a CLI, or a Python SDK.

What changed

The May security pass replaced MD5 with SHA-256, closed an argument injection path, removed a tracked copy of an environment file and fixed a JSON serialisation bug. Earlier commits in the same run fixed runtime crashes and undeclared runtime dependencies.

A tracked environment file in a repository is the kind of thing worth calling out rather than quietly deleting.

Get started

Python.

pip install safla

Register it with Claude Code:

git clone https://github.com/ruvnet/SAFLA.git
claude mcp add safla python3 /absolute/path/SAFLA/safla_mcp_enhanced.py

Expected result: SAFLA's tools available in your Claude Code session.

Rotate any credential that ever lived in a tracked environment file, in this or any other repository.

Use it today

Practical case: giving an agent a persistent model of its own recent behaviour. Input is session activity; the workflow is feed the loop and query it; the output is state the agent can reason about instead of re-deriving.

Acceptance test: run the MCP discovery test in the repository and confirm the tools enumerate before wiring anything real.

Push it further

Experimental commentary. Self awareness in this sense is bookkeeping, not consciousness, and that is why it works. The limitation is that a loop with no external check drifts.

Falsifiable test: run the loop for a long session and compare its self model against ground truth logs. Divergence is the number that matters.