pyro-mcp/docs/integrations.md

Integration Targets

These are the main ways to integrate pyro-mcp into an LLM application.

Use this page after you have already validated the host and guest execution through the CLI path in install.md or first-run.md.

Recommended Default

Use vm_run first for one-shot commands, then move to workspace-core when the agent needs to inhabit one sandbox across multiple calls. Only promote the chat surface to workspace-full when it truly needs shells, services, snapshots, secrets, network policy, or disk tools.

That keeps the model-facing contract small:

• one tool
• one command
• one ephemeral VM
• automatic cleanup

Profile progression:

• vm-run: one-shot only
• workspace-core: persistent workspace create/list/update/status/sync/exec/logs/file ops/diff/export/reset/delete
• workspace-full: the full stable workspace surface, including shells, services, snapshots, secrets, network policy, and disk tools

OpenAI Responses API

Best when:

• your agent already uses OpenAI models directly
• you want a normal tool-calling loop instead of MCP transport
• you want the smallest amount of integration code

Recommended surface:

• vm_run for one-shot loops
• the workspace-core tool set for the normal persistent chat loop
• the workspace-full tool set only when the host explicitly needs advanced workspace capabilities

Canonical example:

• examples/openai_responses_vm_run.py
• examples/openai_responses_workspace_core.py

MCP Clients

Best when:

• your host application already supports MCP
• you want pyro to run as an external stdio server
• you want tool schemas to be discovered directly from the server

Recommended entrypoint:

• pyro mcp serve --profile workspace-core

Profile progression:

• pyro mcp serve --profile vm-run for the smallest one-shot surface
• pyro mcp serve --profile workspace-core for the normal persistent chat loop
• pyro mcp serve --profile workspace-full only when the model truly needs advanced workspace tools

Starter config:

• examples/mcp_client_config.md
• examples/claude_desktop_mcp_config.json
• examples/cursor_mcp_config.json

Direct Python SDK

Best when:

• your application owns orchestration itself
• you do not need MCP transport
• you want direct access to Pyro

Recommended default:

• Pyro.run_in_vm(...)
• Pyro.create_workspace(name=..., labels=...) + Pyro.list_workspaces() + Pyro.update_workspace(...) when repeated workspaces need human-friendly discovery metadata
• Pyro.create_workspace(seed_path=...) + Pyro.push_workspace_sync(...) + Pyro.exec_workspace(...) when repeated workspace commands are required
• Pyro.list_workspace_files(...) / Pyro.read_workspace_file(...) / Pyro.write_workspace_file(...) / Pyro.apply_workspace_patch(...) when the agent needs model-native file inspection and text edits inside one live workspace
• Pyro.create_workspace(..., secrets=...) + Pyro.exec_workspace(..., secret_env=...) when the workspace needs private tokens or authenticated setup
• Pyro.create_workspace(..., network_policy="egress+published-ports") + Pyro.start_service(..., published_ports=[...]) when the host must probe one workspace service
• Pyro.diff_workspace(...) + Pyro.export_workspace(...) when the agent needs baseline comparison or host-out file transfer
• Pyro.start_service(..., secret_env=...) + Pyro.list_services(...) + Pyro.logs_service(...) when the agent needs long-running background processes in one workspace
• Pyro.open_shell(..., secret_env=...) + Pyro.write_shell(...) + Pyro.read_shell(...) when the agent needs an interactive PTY inside the workspace

Lifecycle note:

• Pyro.exec_vm(...) runs one command and auto-cleans the VM afterward
• use create_vm(...) + start_vm(...) only when you need pre-exec inspection or status before that final exec
• use create_workspace(seed_path=...) when the agent needs repeated commands in one persistent /workspace that starts from host content
• use create_workspace(name=..., labels=...), list_workspaces(), and update_workspace(...) when the agent or operator needs to rediscover the right workspace later without external notes
• use push_workspace_sync(...) when later host-side changes need to be imported into that running workspace without recreating it
• use list_workspace_files(...), read_workspace_file(...), write_workspace_file(...), and apply_workspace_patch(...) when the agent should inspect or edit workspace files without shell quoting tricks
• use create_workspace(..., secrets=...) plus secret_env on exec, shell, or service start when the agent needs private tokens or authenticated startup inside that workspace
• use create_workspace(..., network_policy="egress+published-ports") plus start_service(..., published_ports=[...]) when the host must probe one service from that workspace
• use diff_workspace(...) when the agent needs a structured comparison against the immutable create-time baseline
• use export_workspace(...) when the agent needs one file or directory copied back to the host
• use stop_workspace(...) plus list_workspace_disk(...), read_workspace_disk(...), or export_workspace_disk(...) when the agent needs offline inspection or one raw ext4 copy from a stopped guest-backed workspace
• use start_service(...) when the agent needs long-running processes and typed readiness inside one workspace
• use open_shell(...) when the agent needs interactive shell state instead of one-shot execs

Examples:

• examples/python_run.py
• examples/python_lifecycle.py
• examples/python_workspace.py
• examples/python_shell.py

Agent Framework Wrappers

Examples:

• LangChain tools
• PydanticAI tools
• custom in-house orchestration layers

Best when:

• you already have an application framework that expects a Python callable tool
• you want to wrap vm_run behind framework-specific abstractions

Recommended pattern:

• keep the framework wrapper thin
• map one-shot framework tool input directly onto vm_run
• expose workspace_* only when the framework truly needs repeated commands in one workspace

Concrete example:

• examples/langchain_vm_run.py

Selection Rule

Choose the narrowest integration that matches the host environment:

1. OpenAI Responses API if you want a direct provider tool loop.
2. MCP if your host already speaks MCP.
3. Python SDK if you own orchestration and do not need transport.
4. Framework wrappers only as thin adapters over the same vm_run contract.