MCP

TARX uses MCP to expose approved tools and context while preserving local-private boundaries.

Current MCP landscape

MCP has moved from a local-desktop experiment into a multi-client integration surface. Current platform docs now commonly distinguish:

local stdio servers for same-machine private tools
Streamable HTTP servers for local or remote MCP
hosted remote MCP servers with OAuth/account scope
desktop extension packages for one-click local install
client-specific tool visibility, approval, and admin controls

TARX docs should follow that shape without implying every client, tool, or transport is fully production-ready.

What MCP connects

MCP lets compatible clients discover TARX tools and call them through a clear boundary. In TARX, MCP is for context, files, memory, tool access, and evidence-aware workflows.

Client paths

Path	Best for	Status	TARX rule
Local stdio MCP	Private memory, files, Vault, and device context on the user’s Computer.	Pilot-ready	Default for local-private work.
Local Streamable HTTP MCP	Local clients that support HTTP transport and need explicit local endpoint routing.	Pilot-ready / route-specific	Verify local runtime and Bridge state first.
Remote Streamable HTTP MCP	Public-safe tools, discovery, and approved relay flows.	Pilot-ready	No private localhost context by default.
Hosted/OAuth MCP	Account-scoped hosted clients and enterprise agents.	Evaluation lane	Requires auth, scopes, audit, and route policy.
Desktop extension package	One-click local install in clients that support extension bundles.	Future packaging lane	Do not claim until packaged and tested.

Local MCP

Use local MCP when the client runs on the same Computer as TARX and needs private context.


{
  "mcpServers": {
    "tarx": {
      "command": "npx",
      "args": ["@tarx/mcp"]
    }
  }
}

Local MCP is the right default for:

Space context on the user’s Computer.
Local files and private project material.
Memory operations that should not leave the device by default.
Tool calls that require local approval.

Remote MCP

Use remote MCP only for public-safe tools, approved relay flows, or account-scoped workflows where the route is explicit.


{
  "mcpServers": {
    "tarx": {
      "url": "https://mcp.tarx.com/mcp"
    }
  }
}

Remote MCP should not imply private localhost access. If a hosted client needs private context, use a user-approved relay, an export bundle, or another policy-approved route.

Remote MCP uses a streaming MCP transport. A plain browser or generic GET may not return a useful response; use an MCP-compatible client or headers.

Latest verified remote MCP signal from the legacy-docs salvage audit: initialize succeeded against https://mcp.tarx.com/mcp and reported TARX server version 3.16.0.

Compatibility checklist

Question	Required answer
Which transport is used?	stdio, local Streamable HTTP, remote Streamable HTTP, or hosted/OAuth MCP.
Which client is connecting?	Claude Desktop, Claude Code, Cursor, VS Code, OpenAI Agents, enterprise agent, or other compatible host.
Which tools are exposed?	Live `tools/list` or registry proof, not stale marketing counts.
Are resources/prompts exposed?	State explicitly. Do not imply resources or prompts from tools alone.
What auth applies?	Local user context, account auth, OAuth, API key, or no auth for public-safe read-only tools.
What needs approval?	Sensitive actions, external sends, file mutations, credential use, deploys, purchases, and privileged tools.
What evidence is recorded?	Route, tool, context, approval, result, blocker, and timestamp.

Tool contract

Tool area	Purpose	Evidence
Status	Check runtime availability and route health.	Runtime status.
Spaces	List Spaces and load approved Space context.	Space and context record.
Memory	Store and search approved work memory.	Memory event.
Files	List or search files available to the Space.	File/source record.
Actions	Propose work that may require review.	Tool and approval record.

Exact hosted or local tool counts must come from live MCP registry proof. Do not publish stale numeric claims from legacy docs.

Platform notes

Platform	What to know
MCP specification	Tools are model-invoked and should keep humans able to see, approve, or deny risky calls.
OpenAI Agents / API	Current docs distinguish hosted MCP tools, Streamable HTTP MCP, and stdio MCP. Use the route that matches privacy and execution policy.
Claude Desktop	Local MCP can be distributed as desktop extensions; enterprise admins can control extension availability.
Cursor / VS Code	Treat them as MCP-compatible clients whose exact setup and auth UX can change by version.
Enterprise agent platforms	Remote MCP requires connector/admin governance; do not route private local context there by default.
Vercel MCP pattern	Remote OAuth MCP is a useful model for account-scoped operational tools and docs search.

Quick client examples

Use the remote endpoint for public-safe tools:


{
  "mcpServers": {
    "tarx": {
      "url": "https://mcp.tarx.com/mcp"
    }
  }
}

Use local MCP when private files, local memory, device context, or Computer-only tools are required.

Troubleshooting

Symptom	Check
Client cannot see TARX tools	Confirm the MCP server process starts and the client config points to the right command or URL.
Tool call cannot reach local runtime	Confirm TARX Desktop or local runtime is running on the Computer.
Private context missing	Confirm the request is scoped to a Space and the client is using local MCP or an approved relay.
Hosted route appears unexpectedly	Check route policy, account settings, and Evidence before continuing.

Next steps

Review Connections.
Connect Claude Desktop.
Connect OpenAI Agents.
Review Skills.

Ready to test MCP with TARX?

Request Mac access.

Use local MCP for private files, device context, Memory, and Vault-bound workflows. Use remote MCP only for public-safe or approved relay paths.

Request Mac Access

Compare connection paths