// governance infrastructure for AI agents

flux7

Four projects, one system. Policy enforcement, governed memory, automated evaluation, human oversight.

Get started Read the docs
01

The problem

AI agents call tools. Tools have consequences. An agent with access to your filesystem, email, database, and cloud APIs is one hallucination away from an irreversible action.

The current answer is either trust everything (dangerous) or approve everything manually (doesn't scale). Between "allow all" and "human in the loop for every call," there's a gap — and no standard infrastructure to fill it.

flux7 fills that gap with a layered governance system that handles the routine automatically and escalates the ambiguous to humans.

02

Architecture

L0  flux7-mesh          Policy engine             allow / deny / escalate    <1ms
     │
     │  pending approval
     ▼
L1  flux7-mesh          Memory lookup             3+ past approvals → allow  ~100ms
     │
     │  novel case
     ▼
L1+ flux7-supervisor    Rules + LLM evaluation    approve / deny / escalate  ~2-20s
     │
     │  uncertain
     ▼
L2  Human               Console UI / Claude Code  final decision             minutes
     │
     ▼
    flux7-memory         Decisions as facts        remembered for next time

Each layer reduces the load on the next. Most tool calls never reach a human.

03

The system

flux7-mesh Go
L0 — Policy engine

Enforces what agents can do. YAML policies, first-match-wins, glob patterns, approval queue, rate limiting, JSONL tracing. One binary, one config file. Fail-closed.

MCP proxyhot-reloadper-agent policies
flux7-memory Go
Memory substrate

Remembers what happened. Persistent, searchable, governed memory. Hybrid search (BM25 + dense cosine + LLM reranking). Decisions stored as queryable facts.

decisions as factstemporalmulti-agent
flux7-supervisor Python
L1+ — Automated evaluation

Evaluates what's ambiguous. Polls pending approvals, applies rules, evaluates via pluggable LLM (Ollama, Anthropic, Claude Code MCP), auto-resolves the routine.

rule engine3 LLM providersconfidence threshold
flux7-console Next.js
L2 — Human oversight

Shows it all and lets humans intervene. Dashboard, approval UI, trace viewer, memory browser, OTEL waterfall. The management plane for the entire system.

dashboardapproval UIOTEL traces
04

Progressive adoption

Day 1
flux7-mesh
Policies, tracing, approval queue. Five minutes from install to first governed tool call.
Day 30
+ flux7-memory
Persistent memory, decision history, auto-approve patterns seen 3+ times.
Day 45
+ flux7-supervisor
Automated L1 evaluation, reduced approval fatigue, LLM-driven judgment for novel cases.
Day 60
+ flux7-console
Dashboard, team approval UI, governance scoring, full audit trail.

Each step is independently valuable. No project requires the others to function.

05

How it compares

Anthropic Console LangSmith Agent Gateways flux7
Scope Anthropic agents LangChain ecosystem HTTP/API routing Any agent, any SDK
Governance Permission policies None Rate limiting Policy + rules + LLM + human
Memory None Trace replay None Decisions as facts
Approvals Inline in SDK None None Web UI + API + supervisor
Self-hosted No Partial Yes Yes, fully
06

Source

flux7-meshApache 2.0 flux7-memoryApache 2.0 flux7-supervisorApache 2.0 flux7-consoleApache 2.0