Getting Started

Each flux7 project works independently. Add layers as you need them.

flux7-mesh alone

Policy enforcement in 5 minutes. No other component required.

Install

go install github.com/KTCrisis/flux7-mesh/cmd/mesh7@latest

Write a policy

config.yaml

mcp_servers:
  - name: filesystem
    transport: stdio
    command: npx
    args: ["-y", "@anthropic/mcp-filesystem"]

policies:
  - tools: ["filesystem.read_file"]
    action: allow
  - tools: ["filesystem.write_file"]
    action: human_approval
  - tools: ["*"]
    action: deny

Run

mesh7 --config config.yaml

Agents connect via MCP. Reads are allowed, writes require approval, everything else is denied. Traces are logged to traces.jsonl.

---

Add flux7-memory

Persistent memory + auto-approve from decision history.

Install

go install github.com/KTCrisis/flux7-memory/cmd/mem7@latest

Start the daemon

MEM7_TOKEN=mem7_secret123 mem7 serve --listen :9070

Add memory to mesh config

config.yaml

mcp_servers:
  - name: filesystem
    transport: stdio
    command: npx
    args: ["-y", "@anthropic/mcp-filesystem"]
  - name: memory
    transport: stdio
    command: mem7
    env:
      MEM7_DIR: /home/user/.mem7

policies:
  - tools: ["filesystem.read_file"]
    action: allow
  - tools: ["filesystem.write_file"]
    action: human_approval
  - tools: ["memory.*"]
    action: allow
  - tools: ["*"]
    action: deny

The mesh automatically checks memory : if a tool call has been approved 3+ times before, it's auto-approved. Decisions are stored as facts for next time.

Use memory from Python

pip install flux7-memory

from mem7 import Mem7

m = Mem7("http://localhost:9070", token="mem7_secret123")

# Store a decision
m.store("deploy.approval", "approved by ops lead",
        tags=["decision"], agent="supervisor")

# Search later
for mem in m.context("deployment approval", limit=5):
    print(f"{mem.key}: {mem.value}")

---

Add flux7-supervisor

Automated evaluation for pending approvals. Reduces approval fatigue.

Install

pip install flux7-supervisor

Run

sup7 --mesh http://localhost:8080 --provider ollama

The supervisor polls the mesh for pending approvals, applies rules, and evaluates ambiguous cases via LLM. Three providers supported : Ollama, Anthropic API, Claude Code MCP.

# With Anthropic API
sup7 --mesh http://localhost:8080 --provider anthropic --api-key $ANTHROPIC_API_KEY

# With local Ollama
sup7 --mesh http://localhost:8080 --provider ollama --model gemma4:e4b

---

Add flux7-console

Human oversight dashboard. Approval UI, trace viewer, memory browser.

docker run -p 3000:3000 ktcrisis/flux7-console

Open http://localhost:3000 — the console connects to mesh and memory to show the full governance picture.

---

The full chain

Agent → mesh7 (policy) → mem7 (history) → sup7 (LLM eval) → console (human)
  L0                       L1                L1+                L2

Each layer reduces the load on the next. Most tool calls resolve at L0 (policy) or L1 (memory). The supervisor handles novel cases. Humans only see what truly requires judgment.

Layer	Component	Latency	What it does
L0	flux7-mesh	<1ms	Policy match — allow, deny, or escalate
L1	flux7-memory	~100ms	Check decision history — 3+ past approvals → allow
L1+	flux7-supervisor	2-20s	Rules + LLM evaluation — approve, deny, or escalate
L2	flux7-console	minutes	Human reviews in web UI