AI-agent integration

v_ase 0.3.3 gives an external AI agent typed access to the same document that a researcher sees and edits. MCP is the primary integration for tool-capable hosts. Native function tools serve application developers. The CLI, HTTP JSON bridge and JavaScript interface remain supported adapters.

Choose a connection

Agent environment

Interface

Start here

An MCP desktop or coding host

stdio MCP, optionally progressive discovery

MCP setup

A local model application

strict native function schemas and dispatcher

Native functions

An HTTP MCP host on the same machine

loopback Streamable HTTP

HTTP transport

Shell automation, HPC, or a host without MCP

Skill + CLI/HTTP

CLI collaboration

v_ase runs no LLM and requires no model API key. Natural language belongs between the human and the external agent; v_ase executes structured requests.

Share one document

Human instructions, typed AI operations and manual refinement in one shared v_ase GUI.

One live document connects the human, agent and exported result.

Human <-------------- natural language -------------> AI agent
  |                                                       |
  | live GUI                                      typed tools
  |                                                       |
  +---------------- v_ase live document <------------------+
                            |
                 CLI / HTTP / JavaScript adapters

The human and AI work in the same live GUI. The raw API uses expectedRevision and expectedDocumentId for guarded edits.

State profiles expose only the structure, appearance, bonds, camera or analysis needed for the next decision. Editing tools require the inspected document ID and revision. They return exact changed paths; a later human edit becomes a new revision that the agent must review. Rendering is for final visual checks, while scientific state is read semantically.

The bundled workflow Skill explains scientific interpretation, validation and recovery. Tool schemas supply the actual parameters. This separates scientific workflow knowledge from transport syntax.

Find a scientific workflow

Use tools by feature to find the relevant operation. For interpretation, follow the dedicated distribution, constraint, trajectory, RDF, volumetric, interface matching, or render/export guide.

Verify the result

Read the relevant state after each edit. Confirm unchanged structure for visualization-only work, and check ASE positions and constraints after a physical edit. Pause a running movie before a frame-dependent read. Inspect a saved image and reopen exported projects or standalone HTML when applicable.

MCP removes shell quoting from agent calls; it does not prove that an agent’s scientific choices are correct or establish a universal SOTA ranking. The implementation and evaluation notes state the scope of the evidence.