VoltageGPU runs CrewAI multi-agent workflows against confidential inference sealed in Intel TDX hardware enclaves. Existing crew code (agents, tasks, planners, hierarchical processes, async kickoffs) works unchanged. Construct a CrewAI LLM with base_url https://api.voltagegpu.com/v1 and a TEE model (Qwen3-32B-TEE, Qwen3-235B-A22B-TEE, DeepSeek-R1-0528-TEE) and every reasoning step routes through a sovereign enclave we operate in the EU.
Multi-agent crews fan out token usage across many LLM calls. Each call carries proprietary context: contract clauses, deal models, internal precedents, patient data. Sending that traffic to OpenAI or Anthropic exposes it to a foreign jurisdiction. Pointing the same crew at VoltageGPU keeps every reasoning step inside an Intel TDX enclave — AES-256 memory encrypted, ECDSA-attested, RGPD Article 28 compliant.
Pay-per-token at the same rates as standard inference. No platform fee, no per-agent license. Volume contracts available beyond 100M tokens / mo.
Drop-in custom LLM for CrewAI. Every reasoning step routes through Intel TDX enclaves we operate in the EU. Existing crew code runs unchanged.
Multi-agent crews fan out token usage across many LLM calls — each carries proprietary context. Keep that traffic provider-blind without rewriting the crew.
Standard CrewAI plus the OpenAI SDK that backs the LLM wrapper. No fork, no patch.
# Install CrewAI and the OpenAI SDK that backs the LLM wrapper
pip install crewai crewai-tools openaiConstruct one LLM instance, pass it to every Agent. Sequential, hierarchical, async kickoffs and async tool calls all work.
# Sovereign legal crew — contract analyst + clause drafter
from crewai import Agent, Crew, Task, LLM, Process
# One LLM, every agent in the crew uses the confidential endpoint
confidential_llm = LLM(
model="openai/Qwen3-235B-A22B-Instruct-2507-TEE",
base_url="https://api.voltagegpu.com/v1",
api_key="vg-...", # https://app.voltagegpu.com/settings/api-keys
temperature=0.2,
max_tokens=4096,
)
analyst = Agent(
role="Senior Contract Analyst",
goal="Identify regulatory and commercial risks in vendor contracts.",
backstory="EU-trained counsel specialized in DORA / RGPD / NIS2.",
llm=confidential_llm,
verbose=False,
)
drafter = Agent(
role="Clause Drafter",
goal="Propose redlines that address each finding.",
backstory="Has drafted MSAs for 200+ regulated SaaS rollouts.",
llm=confidential_llm,
verbose=False,
)
review = Task(
description="Review {contract} and list every Article 28 RGPD gap.",
expected_output="A bulleted list of clause-level findings with severity.",
agent=analyst,
)
redline = Task(
description="Draft a redline for each finding produced by the analyst.",
expected_output="Markdown redlines grouped by clause section.",
agent=drafter,
context=[review],
)
crew = Crew(
agents=[analyst, drafter],
tasks=[review, redline],
process=Process.sequential,
verbose=False,
)
result = crew.kickoff(inputs={"contract": open("msa.txt").read()})
print(result)Use a cheap fast model for screening agents and a deep-reasoning model for the manager and modeler agents. Both LLMs hit the same confidential endpoint.
# Finance deal-screening crew, three agents, hierarchical process
from crewai import Agent, Crew, Task, LLM, Process
reasoning_llm = LLM(
model="openai/DeepSeek-R1-0528-TEE",
base_url="https://api.voltagegpu.com/v1",
api_key="vg-...",
temperature=0.0,
)
fast_llm = LLM(
model="openai/Qwen3-32B-TEE",
base_url="https://api.voltagegpu.com/v1",
api_key="vg-...",
temperature=0.1,
)
screener = Agent(role="Deal Screener", goal="Filter opportunities", llm=fast_llm, backstory="MD-led screening discipline.")
modeler = Agent(role="Model Builder", goal="Build base+stress cases", llm=reasoning_llm, backstory="Sector-specialist analyst.")
ic_writer = Agent(role="IC Memo Writer", goal="Produce IC-ready memos", llm=reasoning_llm, backstory="Investment committee veteran.")
crew = Crew(
agents=[screener, modeler, ic_writer],
tasks=[
Task(description="Screen {opportunity}", expected_output="GO / NO-GO with rationale", agent=screener),
Task(description="Build base+stress model", expected_output="Model JSON + key drivers", agent=modeler),
Task(description="Draft IC memo", expected_output="2-page memo", agent=ic_writer),
],
process=Process.hierarchical,
manager_llm=reasoning_llm,
)
memo = crew.kickoff(inputs={"opportunity": open("teaser.pdf", "rb").read()})Legal team crew
Contract analyst, clause drafter, jurisdiction reviewer, redline aggregator. Long-context Qwen3-235B-TEE for full MSAs.
Finance crew
Deal screener, model builder, due-diligence summarizer, IC memo writer. DeepSeek-R1-TEE for reasoning-heavy steps.
Compliance crew
Control mapper, evidence collector, audit drafter. Pair with a confidential MCP server for control-mapping tools.
Healthcare crew
Medical record summarizer, coding assistant, prior-auth drafter. EU-hosted enclaves keep PHI inside sovereign control.
Every agent prompt sealed in TDX
The OpenAI-compatible endpoint terminates TLS inside the trust domain. Agent prompts decrypt only inside the enclave.
AES-256 memory encryption
CPU-fused keys protect RAM at runtime — the hypervisor cannot inspect memory holding contracts, deal data, or PHI.
Per-request attestation
Each completion can be paired with an ECDSA-signed report identifying the TDX module and base model version.
Zero retention, zero training
Crew prompts are never logged or reused. Native RGPD Article 28 DPA, EU jurisdiction (VOLTAGE EI, France).
Pay-per-token at the same rates as standard inference. No per-agent license, no platform fee. Mix models freely across agents inside a crew.
Volume contracts available beyond 100M tokens / mo.
EXPLORE FURTHER
Ship a sovereign crew this afternoon
Generate an API key, swap base_url, kick off your first confidential crew.