Sovereign Agent Stack (SAS)

"The model becoming free doesn't mean intelligence becomes sovereign. It just relocates the rent." — Daniel Kliewer, The Rented Sovereign

SAS is a local-first, compile-time AI agent framework. It implements an 8-layer sovereignty model that scores how much of your agent infrastructure you actually own vs. rent.

---

What SAS actually does

1. Sovereignty scoring. Parse a sas.yaml → get a score + verdict. Know exactly which layers you own.

2. Compile-time knowledge graph. Markdown vault → graph → queryable. The agent's long-term knowledge is a build artifact, not a runtime dependency.

3. Local auth broker. Encrypted credential vault + audit trail. Tools call through the broker; credentials never leak to the model.

4. Payments abstraction. Virtual card today, MPP tomorrow. Spending limits enforced at the adapter level.

5. Compute substrate. Boot local desktop containers, auto-destroy on idle. The agent's body lives on your hardware.

6. Identity adapters. Email and phone behind swappable adapters. Mock for dev, production stubs for real APIs.

7. Quant research pipeline. Self-contained professional evaluation worlds: data → research → backtest → risk → report → provenance.

8. Community ecosystem. Plugin system + community registry + ARGO skill pack for harness integration.

---

Quick Start

pip install sovereign-agent-stack

# Initialize config
python -m sas init --output sas.yaml

# Run sovereignty audit
python -m sas dashboard --config examples/agency-worker/sas.yaml --verbose

# Or JSON output
python -m sas dashboard --config examples/agency-worker/sas.yaml --cache ~/.sas --json

---

The 8-Layer Sovereignty Model

#	Layer	Rented (typical)	Owned (SAS default)
1	Model	API provider	Ollama local + API fallback
2	Harness	OpenClaw cloud	ARGO-based, self-hosted (always owned)
3	Compute	Orgo cloud VM	Local Docker desktop
4	Identity	AgentMail/AgentPhone	APIs behind local adapter (unavoidably rented)
5	Short-term Memory	Honcho cloud	Local RAG + session memory
6	Long-term Knowledge	Obsidian (accidental)	Compile-time knowledge graph
7	Auth	Composio hosted broker	Local MCP gateway + encrypted vault
8	Payments	Ramp card	Abstracted: VirtualCard → MPP future

Default score: 6/8 owned (87.5% — "Fully Sovereign"). Identity and Payments are unavoidably rented — excluded from the scoring denominator.

Scoring Verdicts

Score	Verdict	Meaning
>= 87.5%	Fully Sovereign	7-8/8 layers owned
>= 62.5%	Sovereign (target)	5-6/8 layers owned
>= 37.5%	Partially sovereign	3-4/8 layers owned
< 37.5%	Rented	0-2/8 layers owned

---

CLI Reference

# Core
python -m sas init [-o OUTPUT]                          # Create template sas.yaml
python -m sas dashboard [-c CONFIG] [--json] [--verbose] # Sovereignty audit

# Knowledge
python -m sas knowledge compile <source> [--store PATH]  # Compile markdown → graph
python -m sas knowledge query <query> [--store PATH]     # Query the graph
python -m sas knowledge audit [--store PATH]             # Audit graph integrity

# Auth
python -m sas auth register <tool> --token X [--auth-type oauth] [--store PATH]
python -m sas auth list [--store PATH]
python -m sas auth get <tool> [--store PATH]
python -m sas auth unregister <tool> [--store PATH]
python -m sas auth audit [--store PATH]

# Payments
python -m sas payments pay <resource> --price X [--methods card] [--adapter virtual_card]
python -m sas payments limit --daily X --per-transaction Y [--adapter virtual_card]

# Substrate
python -m sas substrate boot [--template xfce]
python -m sas substrate list
python -m sas substrate exec <machine_id> <command>
python -m sas substrate destroy <machine_id>

# Identity
python -m sas identity provision-email <username> [--domain agentmail.to] [--mock]
python -m sas identity send-email <inbox_id> --to X --subject Y --body Z [--mock]
python -m sas identity provision-phone [--region US] [--mock]
python -m sas identity call <number_id> --to X [--mock]
python -m sas identity sms <number_id> --message X [--mock]

# Community registry
python -m sas registry publish <name> <layer_id> <version> [--desc X] [--author Y] [--url Z]
python -m sas registry unpublish <name>
python -m sas registry search <query>
python -m sas registry list
python -m sas registry get <name>
python -m sas registry by-layer <layer_id>

# ARGO skill pack
python -m sas argo info                                        # Show skill metadata
python -m sas argo invoke --params '{"action": "sovereignty_check"}'  # Invoke skill
python -m sas argo schema                                     # Show parameter schema

# Quant research
python -m sas quant status
python -m sas research --universe AAPL --horizon 1y
python -m sas quant backtest --strategy-id momentum-001 --seed 42
python -m sas quant risk --weights AAPL:0.5,MSFT:0.5
python -m sas quant provenance [node_id]

---

Python API

Sovereignty

from sas.core.config import parse_sas_yaml
from sas.core.scoring import generate_report
from sas.layers import LayerRegistry

config = parse_sas_yaml(Path("sas.yaml"))

# Via scorer
report = generate_report(config)
print(report.score)       # 0.875
print(report.verdict)     # "Fully Sovereign"
print(report.owned_count) # 7

# Via registry
registry = LayerRegistry(config)
owned, total, score = registry.sovereignty_score()
print(f"{owned}/{total} = {score:.1%}")

Knowledge Graph

from sas.layers.knowledge import CompileTimeKnowledge
from pathlib import Path

ctk = CompileTimeKnowledge(store_path=":memory:")
graph = ctk.compile(Path("~/sas-knowledge"))
results = ctk.query(graph, "sovereignty")

# Audit
report = ctk.audit(graph)
print(f"{report.total_nodes} nodes, {len(report.orphaned_nodes)} orphaned")

Auth Broker

from sas.layers.auth import LocalAuthBroker, Credentials

broker = LocalAuthBroker(store_path="~/.sas/vault.db")
broker.register_tool("github", Credentials(
    tool_name="github", auth_type="oauth", token="ghp_...", scopes=["repo"]
))

# Make a credential-bearing call
from sas.layers.auth import Request, Response
req = Request(tool_name="github", method="GET", path="/repos", headers={})
resp = broker.call(req, my_http_client)

Payments

from sas.layers.payments import VirtualCardAdapter, PaymentRequirement, SpendingLimit

adapter = VirtualCardAdapter(limit=SpendingLimit(daily=100, per_transaction=50, currency="USD"))
receipt = adapter.pay(PaymentRequirement(
    resource="api", price=25.0, currency="USD", methods=["card"], cadence="one_shot", metadata={}
))
print(receipt.status)  # "completed"

Identity

from sas.layers.identity import MockEmailAdapter, MockPhoneAdapter

# Email
email = MockEmailAdapter()
inbox = email.provision("agent", "agentmail.to")
email.send(inbox, Email(from_="agent@agentmail.to", to="x@y.com", subject="Hi", body="Test"))
sent = email.watch(inbox)

# Phone
phone = MockPhoneAdapter()
number = phone.provision("US")
phone.sms(number, "Integration test")

---

Quant Research Pipeline

Self-contained professional evaluation worlds. Each QuantWorld is: data + documents + portfolio + strategies + policies + agents + tools + task + rubric + gold output.

Portfolio Intelligence World (qw-portfolio-intel-001): 10 tech stocks, $500K portfolio, 5 risk policies, 3 strategies, 15 rubric criteria — all passing.

QuantWorld + Task + Rubric
         │
         ▼
  QuantToolbox (22 tools: market data, computation, backtest, risk, report, provenance)
         │
         ▼
  ModelAdapter.run_loop()  ← ReAct: model calls tools → records trajectory + artifacts
         │
         ▼
  RunEvaluator.evaluate()  ← rubric criteria + sovereignty checks + provenance
         │
         ▼
  RunEvaluation: Pass@1, mean_score, sovereignty_passed

Real Data Providers

Provider	Endpoint	Cache
YFinance	yfinance.Ticker.history()	~/.sas/yf_cache/
Stooq	Free CSV endpoint	~/.sas/stooq_cache/
Alpaca	Data API v2 (paper/live)	~/.sas/alpaca_cache/

All providers implement MarketDataProvider: get_prices(symbol, start, end), validate(), source_info().

---

Architecture

┌─────────────────────────────────────────────────────────────────────────┐
│  CLI: dashboard | init | knowledge | auth | payments | substrate | ...  │
├─────────────────────────────────────────────────────────────────────────┤
│  Layer Registry (single source of truth for 8 sovereignty layers)        │
├─────────────────────────────────────────────────────────────────────────┤
│  Layers: model | harness | compute | identity | memory | knowledge |    │
│          auth | payments                                                │
├─────────────────────────────────────────────────────────────────────────┤
│  Quant: engine | strategy | backtest | risk | broker | market |         │
│         provenance | reports | lifecycle | agents | knowledge            │
├─────────────────────────────────────────────────────────────────────────┤
│  Runtime: AgentRuntime | FleetCoordinator | Orchestrator                 │
├─────────────────────────────────────────────────────────────────────────┤
│  Ecosystem: Plugin system | Community registry | ARGO skill pack         │
└─────────────────────────────────────────────────────────────────────────┘

---

Configuration

# sas.yaml
model:
  primary:
    provider: ollama
    name: llama3.1:8b
    location: local
  fallback:
    provider: openai
    name: gpt-4o
    location: api

compute:
  substrate: local_docker
  resources:
    cpu: 4
    memory: 8Gi
  auto_destroy: 300

memory:
  short_term:
    provider: local_rag
  long_term:
    provider: compile_time_graph
    source: ~/sas-knowledge

auth:
  broker: local_mcp_gateway
  vault: ~/.sas/vault.db
  encryption: libsodium

payments:
  adapter: virtual_card
  virtual_card:
    provider: ramp
    limit: 100

identity:
  email:
    provider: agentmail
    domain: agentmail.to
  phone:
    provider: agentphone
    region: US

Example configs: examples/agency-worker/ (6/6 owned), examples/personal-assistant/, examples/industry-analyst/.

---

Plugin System

Three sources, priority-ordered: LOCAL > PIP > BUILTIN.

from sas.plugins import LayerPlugin, PluginSource, register_plugin

register_plugin(LayerPlugin(
    name="my-payments",
    layer_id="layer_8_payments",
    version="1.0.0",
    description="Custom payment adapter",
    source=PluginSource.LOCAL,
))

Discovery scans built-in plugins, pip entry points (sas.layers group), and ~/.sas/plugins/*.py.

---

ARGO Skill Pack

Exposes SAS as an ARGO-compatible skill with 12 actions: sovereignty_check, compile_knowledge, query_knowledge, register_credential, list_credentials, audit_credentials, quant_status, substrate_list, substrate_boot, substrate_destroy, payments_pay, payments_limit.

python -m sas argo invoke --params '{"action": "sovereignty_check"}'

---

Testing

python -m pytest tests/unit/ tests/integration/ -q \
  --ignore=tests/integration/test_mcp_server.py \
  --ignore=tests/integration/test_rust_extension.py

445 tests passing. Coverage includes: all 8 layers, quant pipeline (15/15 criteria), real data providers (YFinance/Stooq/Alpaca), plugin system, community registry, ARGO skill pack, full stack integration.

---

Documentation

Document	Purpose
docs/ROADMAP.md	8-phase development plan
docs/DEPLOYMENT.md	Production deployment guide
docs/RUNBOOK.md	Operations runbook

---

License

MIT