AgenticOS

In the Activity view, AgenticOS displays the model, Token usage, reported cost, and tool events of a single run together, enabling cost figures to be linked to specific tasks.

If fallback is enabled, the rows recorded in AgenticOS documentation may still retain the primary profile even if another model actually answered, so controlled model comparisons should use a test profile with fallback disabled or take credentials directly from the vendor side.

AgenticOS retains failed runs and their recorded usage in history, making failure records equally meaningful when reviewing workflow costs.

In AgenticOS's Spend view, a missing price does not equal a price of zero; when there are unpriced runs within the time window, the displayed total is a lower bound of that window's usage rather than the full picture, and the real amount may be higher.

AgenticOS checks accounted spending before issuing a model request; if the applicable limit has been exhausted, the next request may be rejected, and to check which limit applies to a given run, one must look at the published configuration rather than draft settings.

A budget is not a full reservation of future bills: charges may still be incurred after a request passes the check, and concurrent or already-submitted work may not yet be reflected in the accounted total.

Model usage is only part of the cost of an Agent workflow; external tool services, sandbox compute, and storage may each have their own bills, and estimating the operational cost of an entire task requires making all these components visible.

The AgenticOS governance guide is located at https://vstorm-co.github.io/agenticos/governance/ and explains budgets and their accounting limitations.

To localize the AgenticOS installer, select --provider none and use --no-mcp to skip downloading the public MCP server registry mirror.

AgenticOS is an open-source AI agent platform developed by the Vstorm team, released under the Apache-2.0 license.

The HIPAA-compliant configuration of AgenticOS resolves the hostname of each model endpoint to determine whether it is local, which is merely a configuration check rather than a compliance certification.

AgenticOS is currently still at version 0.0.x, supports only a single host, has no Kubernetes manifests, and has no reranker.

AgenticOS's conversation model can connect to local models via Ollama configuration files, a self-built LiteLLM proxy, or OpenAI-compatible endpoints from vLLM or LM Studio.

AgenticOS's document parsing defaults to PyMuPDF, running in conjunction with LiteParse and Tesseract OCR within the worker.

If LOGFIRE_TOKEN is not set and no tracing token is configured on any agent or environment, run records will remain local.

Keyless local models do not log any spending; the conventional dollar-based budget mechanism cannot control local models, and it is necessary to prevent tools from getting stuck in infinite loops via each agent's step limit (max_steps).

An agent's capabilities are entirely dependent on the model's tool-calling ability; planning, document search, and sandbox operations are all completed through tool calls.

The Vstorm team currently has no benchmarks for local models to share.

Building the AgenticOS system requires Docker Compose 2.24 or higher, using Postgres with pgvector, Redis, and Prefect, with images supporting amd64 and arm64 architectures.