Skip to content

Air-Gapped Deployment

AgentCost is fully self-contained and runs without internet access. This guide covers deploying AgentCost in air-gapped, restricted, or on-premises environments where outbound connectivity is not available.

How It Works

AgentCost ships with everything it needs inside the Docker image:

  • Pricing database — A vendored JSON file containing pricing for 2,610+ models from 83+ providers (OpenAI, Anthropic, Azure OpenAI, GCP Vertex AI, AWS Bedrock, Google Gemini, Mistral, Cohere, DeepSeek, and more). No external API calls are made to calculate costs.
  • Dashboard — The React dashboard is bundled as static files. No CDN dependencies.
  • Database — Community edition uses SQLite (single file, zero config). Enterprise supports PostgreSQL.
  • No telemetry — AgentCost does not phone home, send analytics, or make any outbound requests.

Deployment

Step 1: Transfer the Docker image

On a machine with internet access, pull and save the image:

docker pull agentcost/agentcost:latest
docker save agentcost/agentcost:latest -o agentcost-image.tar

Transfer agentcost-image.tar to your air-gapped environment via USB, secure file transfer, or your approved media process.

Step 2: Load and run

On the air-gapped machine:

docker load -i agentcost-image.tar
docker run -d \
  --name agentcost \
  --restart unless-stopped \
  -p 8100:8100 \
  -v agentcost_data:/data \
  agentcost/agentcost:latest

The dashboard is available at http://<host>:8100.

Step 3: Instrument your code

Your application code connects to the local AgentCost instance. No external endpoints involved:

from agentcost.sdk import trace
from openai import OpenAI

# Point your LLM client at your internal endpoint (Azure OpenAI, etc.)
client = trace(
    OpenAI(base_url="https://your-internal-openai.company.com/v1"),
    project="my-app",
    agentcost_url="http://agentcost-host:8100"
)

All cost tracking happens locally — traces are stored in the local database, costs are calculated from the local pricing file.

Pricing Database

Where costs come from

AgentCost vendors the pricing database from the LiteLLM open-source project. This is a community-maintained JSON file that covers:

Provider Coverage
OpenAI (direct API) All GPT-4.1, GPT-5.x, o3, o4-mini models
Anthropic (direct API) Claude Opus 4.6, Sonnet 4.6, Haiku 4.5
Azure OpenAI All deployed models with Azure-specific pricing
GCP Vertex AI Gemini 3.x, PaLM, Claude on Vertex
AWS Bedrock Claude, Titan, Llama, Mistral on Bedrock
Google Gemini Gemini 3 Pro, Flash, and variants
Mistral All Mistral and Mixtral models
Cohere Command, Embed models
DeepSeek DeepSeek-V3, DeepSeek-R1
Open-source (Ollama, vLLM) Tracked as $0 cost (infrastructure cost only)

The file is located at agentcost/cost/model_prices.json inside the Docker image and the Python package.

Offline pricing updates

When model pricing changes, you can update the pricing database without internet access:

Option A: New Docker image

Pull the latest image on an internet-connected machine, transfer it as described in Step 1. The updated pricing database is included in every new release.

Option B: Manual file update

  1. On an internet-connected machine, download the latest pricing file:
curl -o model_prices.json \
  https://raw.githubusercontent.com/agentcostin/agentcost/main/agentcost/cost/model_prices.json

  1. Transfer the file to your air-gapped environment.

  2. Copy it into the running container:

docker cp model_prices.json agentcost:/app/agentcost/cost/model_prices.json
docker restart agentcost

Pricing is effective immediately after restart.

Option C: Custom pricing overrides

For models with internal or negotiated pricing (e.g., your Azure OpenAI enterprise agreement), you can set custom costs via the AgentCost API:

curl -X POST http://agentcost-host:8100/api/models/pricing \
  -H "Content-Type: application/json" \
  -d '{
    "model": "azure/gpt-4.1",
    "input_cost_per_token": 0.0000018,
    "output_cost_per_token": 0.0000072
  }'

Custom pricing overrides take precedence over the vendored database.

On-Premises LLM Cost Tracking

For self-hosted models (Ollama, vLLM, TGI, etc.), AgentCost tracks token usage, latency, and call volume — costs are recorded as $0.00 since there's no per-token charge. This is useful for:

  • Usage analytics — who's using which models, how much, and how often
  • Capacity planning — token volume trends help plan GPU scaling
  • Comparison — see the cost delta between your self-hosted models and equivalent cloud APIs

To track infrastructure costs for on-premises LLMs (GPU hours, electricity, etc.), you can set a custom per-token cost that reflects your amortized infrastructure cost:

curl -X POST http://agentcost-host:8100/api/models/pricing \
  -H "Content-Type: application/json" \
  -d '{
    "model": "ollama/llama3.2:70b",
    "input_cost_per_token": 0.0000005,
    "output_cost_per_token": 0.0000010
  }'

Enterprise Edition in Air-Gapped Environments

The enterprise edition adds governance features that are especially relevant for restricted environments:

Feature Why It Matters in Air-Gapped
SSO/SAML Integrate with your internal identity provider (ADFS, Okta on-prem, Keycloak)
Policy Engine Enforce which models can be used, cap costs per project
Approval Workflows Require human approval for expensive model usage
Audit Logs Hash-chained compliance trail for regulatory requirements
Budget Enforcement Auto-downgrade models when budgets are hit — critical when chargebacks are involved
Anomaly Detection Catch runaway agents without relying on external monitoring

Enterprise licensing works offline — the license key is validated locally using HMAC-SHA256 signatures. No license server or internet access required.

Architecture Summary

Air-Gapped Network Architecture

FAQ

Does AgentCost make any outbound network calls? No. All cost calculations use the local pricing database. No telemetry, no analytics, no license server calls.

How often does the pricing database need updating? Major pricing changes happen a few times per year. For most deployments, updating quarterly is sufficient. If you need exact-to-the-penny accuracy for billing, update monthly.

Can I run AgentCost on Kubernetes? Yes. The Docker image works in any container runtime — Docker, Kubernetes, OpenShift, ECS. Use a PersistentVolume for the /data directory to retain the SQLite database across pod restarts.

What about PostgreSQL in air-gapped? Enterprise edition supports PostgreSQL. Deploy PostgreSQL within your air-gapped network and set the DATABASE_URL environment variable. The Docker Compose file includes a PostgreSQL configuration.

Questions about air-gapped deployment? Contact us at care@agentcost.in or open an issue on GitHub.