Integrations: AI model providers

Argent with vLLM

Run models on your own hardware and govern them like any cloud model. Argent applies the same approved list, redaction, budgets and signed record to vLLM, with no key to manage and no data leaving your network.

vLLM logo
01 / INTEGRATIONS: AI MODEL PROVIDERS

What Argent does with vLLM

vLLM serves open models at scale on your own GPUs. Argent governs it like any cloud provider, and nothing has to leave your network.

  • Built in: vLLM's connection settings ship with Argent.
  • No key needed: the endpoint is marked self-hosted and keyless, so Argent fetches nothing from a vault for it.
  • Metered on streams: some self-hosted engines report token use on a stream only when asked. Argent asks for them, so cost metering works.
  • Approved models and rules: the same approved list and access rules as every provider.
  • Local data protection: the sensitive-data detector runs as part of Argent, so prompts are scanned without sending text anywhere else.
02 / INTEGRATIONS: AI MODEL PROVIDERS

How it connects

Register the engine, price its models, point agents at Argent.

  1. Register the engine on the AI Providers page as a self-hosted, keyless endpoint.
  2. Register each agent with an owner. It gets its own Argent key and copy-paste setup.
  3. Point it at Argent. The agent's base URL becomes its gateway route.
  4. Price your models. Add your internal rate per model in the price book, in USD per million tokens, so budgets count self-hosted use. An agent under a budget that calls a model with no price is refused.

Argent itself can run inside your network with no internet connection. See Air-Gapped Install and Air-gapped on-prem.

03 / INTEGRATIONS: AI MODEL PROVIDERS

What you configure

Everything below is set once in the Argent console and applies to every agent that calls vLLM through the gateway.

SettingWhereWhat it does
Approved modelsPolicy > Model and endpoint accessDefault deny: a call is allowed only when the model is on the approved list. The list is platform AND tenant AND agent, and each layer can only narrow the one above.
Access rulesSame pageAllow or deny by provider, route type, model, hosted tools, function tools and output type, for the tenant, a department, a person or one agent. The narrowest rule wins, and a 7-day impact preview shows what a draft would change.
Sensitive dataDLPPer entity type: redact and record, record only, or block the request.
Budgets and capsBudgetsBudgets for the company, org, department, person and agent. Each request is priced and held against the budget before it leaves, and an emergency cap stops a runaway agent now.
ModePer controlOff, watch or enforce. Watch records what would have been refused and changes nothing on the wire.

Start in watch, read what would have been refused, then switch on enforce. See Watch Before You Enforce.

04 / INTEGRATIONS: AI MODEL PROVIDERS

One request, step by step

What happens between your agent and vLLM, in order.

  1. The agent sends its request to its Argent route, carrying its own Argent key and nothing from vLLM.
  2. Who is calling. Argent identifies the agent from its key and checks it may use this endpoint. The endpoint comes from the route the gateway matched, never from a header the caller sets.
  3. Which model. Argent reads the model from the request and checks it against the approved list and access rules before anything leaves.
  4. Refuse or allow. A model nobody approved is refused at the gateway with a named reason, such as llm.deny.model.
  5. Sensitive data. The request is scanned by a reader that knows the vLLM format, and secrets and personal data are redacted, recorded or blocked under your policy. If the scan cannot run, the request is refused, never passed as clean.
  6. Budget. The request is priced and the amount is held against the agent's budget, so parallel calls cannot overshoot it.
  7. No key. The endpoint is keyless, so nothing is fetched or attached.
  8. Reply and record. Token usage is read from the reply, the cost is settled, and the signed record is written.
05 / INTEGRATIONS: AI MODEL PROVIDERS

What the record shows

Every vLLM call leaves one signed record, allowed or refused.

  • One row per call: the agent and the person who owns it, the tenant, the model, the decision and the policy that made it, latency, and the cost, token and cache figures.
  • Tamper-evident: each row is hash-chained to the one before and signed with an ed25519 key held in Vault Transit, so an edited or missing row shows.
  • Real cost: cost comes from the usage block in vLLM's own reply, priced from your price book in USD per million tokens.
  • No unpriced spend: an agent under a budget that calls a model with no price is refused, so a new model cannot slip past a cap.

Conversations can be opened and replayed from the record. See Session Replay and One-Click Audit Evidence.

06 / INTEGRATIONS: AI MODEL PROVIDERS

Related capabilities and goals

More AI model providers

Every provider below is governed through Argent.

  • OpenAI
  • Anthropic
  • Google Gemini
  • Azure OpenAI and Foundry
  • Amazon Bedrock
  • Mistral
  • DeepSeek
  • OpenRouter
  • GitHub Copilot
  • Kiro
  • Ollama
  • NVIDIA NIM
  • LM Studio
  • Hugging Face TGI