VXI Local-First MCP

A source-free, read-only MCP status bridge for sensitive development workspaces. Node.js 22+, zero npm dependencies. Original demonstration implementation; no code or configuration was copied from any commercial project.

Security claim (narrow): the MCP tools in this repository export only pre-defined numeric file counts, enums, and static policy. They do not expose source, file paths, Git diffs, logs, environment variables, or command execution. This is a projection boundary, not a claim that an unrelated AI coding client keeps code private. A local desktop application connected to a hosted model may still transmit source. For a strictly no-source-outbound workflow, use local inference or deterministic human-approved local operations for the coding step.

Try in under a minute

node --version  # >= 22
npm test
npm run demo

The demo scans an intentionally synthetic directory containing a fake secret-like text fixture. Its file name and contents never appear in MCP responses. The test suite additionally creates local .env files in temporary workspaces; these files are not included in the public repository.

Inspect your own local project without sharing source

node bin/vxi-local.mjs status --workspace /path/to/local/workspace
node bin/vxi-local.mjs record --workspace /path/to/local/workspace --kind test --status PASS
node bin/vxi-local.mjs status --workspace /path/to/local/workspace

record is a local CLI action, not an MCP tool. It stores only status enums under ~/.vxi-local-first/ (not in the repo). Statuses are human-recorded assertions; they are not verified attestations of test execution. A safer future extension can bind them to a local CI signature.

Run an MCP stdio server

VXI_WORKSPACE=/path/to/local/workspace node bin/vxi-local.mjs mcp

Tools: vxi_project_status, vxi_security_contract. Neither accepts arguments; all other method/tool names are rejected. The server uses line-delimited JSON-RPC on stdio and is intended to run in a trusted local process.

Connect from ChatGPT without a public inbound port (optional)

Use the officially documented OpenAI Secure MCP Tunnel and your own Platform tunnel permissions. The tunnel client can launch a local stdio MCP server. Configure its --mcp-command to point to the mcp CLI above, with VXI_WORKSPACE supplied only on the local machine. Follow the actual current tunnel-client help quickstart and onboarding instructions:

https://developers.openai.com/api/docs/guides/secure-mcp-tunnels

This package has been tested offline only; no ChatGPT/Platform tunnel connection is claimed. Tunnel connectivity alone cannot prevent data leakage if a different MCP tool exports source. The safe surface here is enforced by this server's output schema.

Workflow boundary

1. Cloud model sees only abstract task instructions and the fixed output of this bridge. Never paste business code or raw logs into cloud chat.
2. Developer writes an abstract plan with no confidential file names or snippets.
3. A local human/local-only model determines file-level edits. A cloud-backed desktop coding agent is not a strictly local model.
4. Run real tests locally. Record only PASS/FAIL/UNKNOWN via the local CLI; optionally inspect full diff locally.
5. Cloud model may review evidence sufficiency, not unseen code correctness. No DONE claim without local review.

Threat model

• Reject source export by construction: no read-file or diff tools.
• Paths remain internal; only explicit numeric counters are returned.
• Static symlinks are skipped; each queued directory is rechecked against the canonical root. Like any pathname-based traversal, it is not a proof against a hostile concurrent symlink replacement on a shared filesystem. Dot-directories and common build/cache directories are skipped.
• Check state is stored outside the project. File contents are never opened during scanning.
• A local attacker who controls the Node process, runtime environment, or MCP host is out of scope.
• Metadata can also be sensitive. Do not use without approval where file counts/test statuses are classified.
• MCP callers remain untrusted. Do not extend the tool list with exec, file reads or remote shell without a new threat-model review.

Limitations

• No actual code review by the cloud model; strict privacy trades away code-level cloud reasoning.
• No local LLM or Codex desktop automation bundled; install/configure those independently.
• record assertions are not cryptographically bound to an executed commit.
• No remote connection, cost reduction, or production safety was verified by this offline fixture.
• The scanner is bounded by 20,000 directory entries and does not follow symlinks.

License: MIT.