mcp-portal

mcp-portal is a stdio Model Context Protocol server that lets a frontier agent (Claude Code, Codex, Cursor, or any MCP host) delegate two jobs to the Cursor CLI on its own quota: bounded bulk_read (read explicitly selected files, answer with verified quotes) and code_write (generate boilerplate from a reference file + spec; the server writes the target file). Python stdlib only—no Node runtime and no MCP SDK dependency.

On 2026-09-08, composer-2.5-fast generated roughly 5× faster than a frontier model on the same brief (line-rate measurement). Cursor quota is separate from the host model's.

Quick start

Until the first PyPI release lands, install straight from GitHub:

uvx --from git+https://github.com/apollion69/mcp-portal mcp-portal

After the PyPI release the short forms work:

uvx mcp-portal
pipx install mcp-portal
pip install mcp-portal

Requirements: Python 3.10+, the Cursor CLI (cursor-agent) installed and logged in.

Doctor (CLI inventory, no model call):

mcp-portal-doctor

Configure per host

Claude Code

claude mcp add --scope user mcp-portal -- uvx mcp-portal

Codex (~/.codex/config.toml)

[mcp_servers.mcp-portal]
command = "uvx"
args = ["mcp-portal"]
tool_timeout_sec = 150

Cursor (~/.cursor/mcp.json)

{
  "mcpServers": {
    "mcp-portal": {
      "command": "uvx",
      "args": ["mcp-portal"]
    }
  }
}

VS Code (.vscode/mcp.json, servers key)

{
  "servers": {
    "mcp-portal": {
      "type": "stdio",
      "command": "uvx",
      "args": ["mcp-portal"]
    }
  }
}

Generic mcpServers JSON

{
  "mcpServers": {
    "mcp-portal": {
      "command": "uvx",
      "args": ["mcp-portal"]
    }
  }
}

Environment (optional):

Tools

bulk_read

Returns status, run_id, answer.findings[] (file, start, end, quote, fact), gaps[], metrics, model_decision.

code_write

Returns generated code, optional bytes_written, run_id, metrics.

status

No arguments. Returns CLI path, auth hint, default model, policy summary, cache location, receipt counters.

Model policy

Shipped in model-policy.json (package data). Defaults:

• Prefer Cursor-native models (composer-2.5, then cursor-grok-*)
• Strip -fast suffixes (never auto-select fast variants)
• Other vendors only when explicitly requested and listed by cursor-agent --list-models

Override by editing model-policy.json in the installed package or setting policy fields via a custom file at MCP_PORTAL_HOME (future) — today, replace the package file or patch preferred in your fork. Each tool result includes model_decision.reason (default_preferred, fast_suffix_stripped, cursor_native_explicit, explicit_other_vendor, requested_unavailable_fallback).

How it works

1. Authorize — Server reads only listed paths; blocks credential-like paths and secret patterns.
2. Manifest — Request JSON includes per-file SHA-256 hashes.
3. Isolate — Cursor CLI runs with fresh CURSOR_CONFIG_DIR, deny-all permissions, --mode ask, sandbox enabled.
4. Verify — Every quote in bulk_read answers must appear verbatim in the cited line range; bad citations are dropped or fail closed.
5. Evidence — Per-run directory under MCP_PORTAL_HOME/runs/<run_id>/ with manifest (hashes, metrics; not full source).
6. Budgets — 16 files, 128 KiB combined input, 90s timeout, bounded stdio frames.

Windows

On Windows, the delegate uses a local Cursor CLI run when either:

• MCP_PORTAL_CLI points at an executable (including test stubs), or
• cursor-agent / agent is found on PATH and is a real file.

Otherwise it falls back to the wsl.exe bridge into Ubuntu/WSL (python3 -m mcp_portal.delegate --worker). Force either mode with MCP_PORTAL_BACKEND=local or MCP_PORTAL_BACKEND=wsl.

• MCP config can use native uvx mcp-portal when the CLI is on PATH, or wsl.exe + uvx mcp-portal when it is not
• Helpers in clients/windows/ (delegate.ps1, parse_read.ps1)
• MCP_PORTAL_WORKER overrides the default WSL worker command
• MCP_PORTAL_WSL_CD sets the WSL working directory (default ~)

Optional Claude Code routing hook

Install read gate + skill (generic, transactional):

python3 -m mcp_portal.install_router prepare --client claude --python python3 \
  --state-root ~/.cache/mcp-portal/router-tx --shell bash --command-shell bash
# then apply with the printed transaction id

See docs/skills/cursor-bulk-reader/SKILL.md for agent-facing guidance. The router blocks or warns on large full-file reads (>350 lines or >128 KiB) and points agents at bulk_read.

Repo-level MCP registration helper:

python3 -m mcp_portal.install plan
python3 -m mcp_portal.install apply --target claude-mcp

Security

See SECURITY.md. Summary: you choose which files leave the machine; the CLI runs read-only with tools denied; quotes are verified server-side. Not a substitute for secret hygiene.

Related projects

Several Node-based bridges expose Cursor via MCP (different tradeoffs: SDK/Node stack, varying isolation and verification):

• lipey1/cursor-agent-mcp
• andreilungeanu/cursor-delegate-mcp
• sailay1996/cursor-agent-mcp
• ai-nuke/cursor-agent-mcp
• JaimeJunr/cursor-mcp-bridge

mcp-portal focuses on stdlib Python, hash-pinned manifests, quote verification, server-side writes for code_write, model policy, and WSL-first Windows support.

License

MIT — see LICENSE.