TinyPyMCP
A sandboxed MCP server for agents that provides file read/edit, code intelligence, search, allowlisted command execution, package lookups, persistent memory, and VPS channel clients, all confined to C:\Work.
README
TinyPyMCP
A Python MCP server: a sandboxed workshop for an agent — file read/edit, code
intelligence, search index, local command execution, package lookups, persistent
memory, and a client for bounded VPS channels. All filesystem access is confined
to C:\Work. 50 tools.
For development/maintenance notes and the roadmap, see DEVNOTES.md.
Run
Local clients (Claude Desktop, Cursor) — stdio, no auth:
python -m src.server
Remote / hosted connectors — HTTP (Streamable HTTP), auth required:
# Bearer mode (token from file): bearer_token.json = {"token": "<secret>"}
python -m src.server --auth bearer --token-file C:\Users\mczyz\.romion\bearer_token.json --transport http --port 8765
# OAuth 2.1 mode (recommended for hosted connectors): see below.
Auth is selected by --auth bearer|oauth|none (default bearer). Secrets come
from --token-file / --secret-file (JSON, kept outside C:\Work) or env. none
runs open — never on a public tunnel.
Endpoint: http://127.0.0.1:8765/mcp/v5 (public via Cloudflare Tunnel at
https://tiny-py-mcp.romionologic.dev/mcp/v5).
The HTTP transport refuses to start without MCP_AUTH_TOKEN (fail-closed),
because the tunnel exposes tools like run_command. For trusted localhost-only
dev you may set MCP_AUTH_DISABLE=1 — never on port 8765 while the tunnel runs.
How clients authenticate
- Header clients: send
Authorization: Bearer <secret>. - ChatGPT connector (can't send custom headers): set the URL to
https://tiny-py-mcp.romionologic.dev/mcp/v5?token=<secret>and Authentication to "No authorization". The token in the URL is redacted from logs.
Environment variables
| Var | Purpose |
|---|---|
MCP_AUTH_TOKEN |
Bearer/query token for HTTP. Required unless MCP_AUTH_DISABLE=1. |
MCP_AUTH_DISABLE=1 |
Run HTTP without auth (trusted local only). |
MCP_PUBLIC_HOST |
Public host for the DNS-rebinding allowlist (default tiny-py-mcp.romionologic.dev). |
MCP_MEMORY_DB |
SQLite memory path (default data/memory.db). |
MCP_WORKSPACE_ROOT |
Where clone_repo/indexes live (default workspaces/). |
MCP_AUDIT_LOG |
Audit JSONL path (default logs/.tinypymcp-audit.jsonl). |
MCP_VPS_CONFIG |
VPS-channel creds JSON (default ~/.romion/vps-channel.json, kept OUTSIDE C:\Work). |
MCP_AUTH_MODE |
bearer (default) or oauth. Selects HTTP auth. |
MCP_OAUTH_OPERATOR_SECRET |
OAuth mode: the secret you type on the authorization page (the human gate). Required for oauth mode. |
MCP_OAUTH_ISSUER |
OAuth mode: public issuer URL (default https://<MCP_PUBLIC_HOST>). |
MCP_OAUTH_DB |
OAuth store path (default data/oauth.db). |
MCP_EXEC_ALLOWLIST |
Comma/space-separated exec allowlist override (e.g. "git" drops interpreters). Default: git/node/npm/python/pytest/... |
MCP_CF_PROTECTED |
Override the protected-CF-host denylist (delete/remove of these needs force=true). |
MCP_RATE_LIMIT_PER_MIN |
Per-IP request cap (0 = off, default). Anti-runaway; set generous to not break connector bursts. |
OAuth 2.1 mode (standard, for hosted connectors)
python -m src.server --auth oauth --secret-file C:\Users\mczyz\.romion\oauth_secret.json --transport http --port 8765
oauth_secret.json (keep OUTSIDE C:\Work):
{ "operator_secret": "<the secret you type at login>", "issuer": "https://tiny-py-mcp.romionologic.dev" }
The server becomes its own OAuth authorization server (DCR + PKCE; no external IdP). Connectors discover it automatically; during each connector's login you enter the operator secret on the TinyPyMCP page. See CONNECTORS.md.
(Env-var form also works: MCP_AUTH_MODE=oauth, MCP_OAUTH_OPERATOR_SECRET,
MCP_OAUTH_ISSUER. Flags take precedence.)
Tools (37)
- Read/inspect: read_file, read_file_chunk, get_info, list_files, get_project_structure, find_phrase_occurrences, search_codebase
- Edit/write: write_file, create_file, append_file, safe_replace_in_line, edit_code_block, edit_file_patch
- Manage: copy_path, move_path, delete_path (soft → .trash), restore_path, list_trash
- Code intelligence: code_dependencies, code_impact, code_symbols, build_index, search_index, index_status
- Exec/acquire: run_command (allowlisted), clone_repo
- Network: http_probe, check_npm_package, check_pypi_package
- Memory (SQLite): memory_get_state, memory_set_state, memory_save, memory_search, memory_create_task, memory_get_tasks
- VPS channel: vps_status, vps_request
- Cloudflare admin (token from ~/.romion/cloudflare.json; writes need confirm=true): cf_verify_token, cf_list_dns, cf_create_dns_record, cf_delete_dns_record, cf_list_tunnels, cf_get_tunnel_config, cf_add_tunnel_route, cf_remove_tunnel_route, cf_create_access_app, cf_delete_access_app, cf_add_access_service_policy, cf_create_service_token, cf_delete_service_token
Security model
- Sandbox: every file tool resolves paths through
path_guard.ensure_within— confined toC:\Work(../symlink escapes blocked). Network is unrestricted. - Exec:
run_commandruns only allowlisted binaries (git/node/npm/python/ pytest/...), no shell, child env stripped to a safe allowlist (secrets don't leak), timeout + output caps, cwd confined toC:\Work. - Auth: bearer header or
?token=query, fail-closed; redacted from logs. - Deletes are soft (move to
.trash, reversible viarestore_path). - Secrets (VPS channel tokens) live OUTSIDE C:\Work so the agent's own file tools can't read them.
- All mutations and command runs are recorded in the audit log.
Tests
python tests/smoke.py # 34 offline checks over the whole tool surface
VPS deploy channel
Heavy deploy (docker on the VPS) is handled by the separate romion-deploy
service (C:\Work\romion-deploy), reached via vps_request. See its README.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.