evil-mcp

Connect your AI client to EVIL, Cornell Electric Vehicles' telemetry database (runs, turns, laps, straights, energy, NAS recordings), so you can ask about the car in plain English from Claude Code, Codex or Cursor.

You must be on the CEV Tailscale network. The curated server is at http://100.122.165.58:8765/mcp (the NUC, cev-nuc) and the raw-SQL server at http://100.122.165.58:8767/mcp. Both are read-only.

Just want a browser?

Open http://100.122.165.58:8081 (EVIL-UI). Data browser plus a chat panel. No setup.

Fastest setup: one command

git clone https://github.com/cornellev/evil-mcp.git && cd evil-mcp && ./install.sh

install.sh registers EVIL (evil and evil-raw) globally for whichever of Claude Code, Codex and Cursor it finds, so it works from any project. Re-running is safe; ./install.sh --dry-run shows what it would do. Restart your client afterwards, then ask: "List the EVIL runs."

Only for one project

If you'd rather not have EVIL in every session (privacy, or to keep its skill out of the way in unrelated projects), scope it to one directory:

./install.sh --project ~/path/to/your/project

EVIL is then on only in sessions started in that directory or any subdirectory. It writes .mcp.json and .claude/skills/evil/ there, plus .cursor/mcp.json and .codex/config.toml if you have those clients. If the directory is a git repo, the new files are added to .git/info/exclude, so they never get committed. If your project already tracks a .mcp.json, the EVIL entries will show in git status; the script tells you when that happens. Claude Code asks you once to approve the project's servers the first time you start a session there.

To switch from global to project-only, remove the global install first:

./install.sh --uninstall                       # remove the global install
./install.sh --project ~/path/to/your/project  # then scope it

./install.sh --uninstall --project DIR removes it from a project. Neither form touches other servers in the same config files.

EVIL traces every tool call (http://cev-traces) and names the caller from the tailnet: your Tailscale login and device, looked up on the server. Nothing to configure.

Updating

What changed	What you do
Tools on the server (new tool, fixed tool, new data)	Nothing. Restart your client or start a new session and it picks up the new tool list.
The guidance (AGENTS.md / the Claude skill), or a new server or URL	git pull && ./install.sh in your clone (add --project DIR if that's how you installed), then restart your client.

install.sh is both the installer and the updater. Re-running it adds any new servers, re-points entries whose URL changed and refreshes the Claude skill. It leaves alone anything already up to date. Codex and Cursor users who copied AGENTS.md somewhere by hand need to copy it again after pulling.

Or: no install, just open this repo

The repo carries each client's project-level config, so opening this folder in the client is enough: .mcp.json (Claude Code), .cursor/mcp.json (Cursor), .codex/config.toml (Codex, trusted projects only). AGENTS.md (Codex, Cursor) and the skill in .claude/skills/evil/ (Claude Code) teach the model what the tools mean. This works only while you are working in this folder, hence the installer for everything else.

Or: add it by hand

Client	Command / config
Claude Code	claude mcp add --transport http --scope user evil http://100.122.165.58:8765/mcp and ... evil-raw http://100.122.165.58:8767/mcp
Codex	codex mcp add evil --url http://100.122.165.58:8765/mcp and codex mcp add evil-raw --url http://100.122.165.58:8767/mcp
Cursor	add "evil": {"url": ".../8765/mcp"} and "evil-raw": {"url": ".../8767/mcp"} under mcpServers in ~/.cursor/mcp.json

Any other MCP client: Streamable HTTP transport, that URL, no auth.

Getting the guidance outside this repo

The server gives your model the tools; AGENTS.md gives it the know-how (what a turn vs. a lap is, which tool to pick, how to write safe SQL). install.sh copies the Claude skill to ~/.claude/skills/evil/ for you. For Codex and Cursor, copy or append AGENTS.md into your own project, or into ~/.codex/AGENTS.md.

What you can ask

• "List the EVIL runs and tell me which is the most recent."
• "How was Turn 3 in run <run_id>?"
• "Compare laps 2 and 5. What changed in energy and average speed?"
• "Which turn instance had the best exit speed, and what was different about it?"
• "Where is the raw recording covering Turn 7?"
• "What recordings do we have from last weekend, and which ones didn't parse?"
• "(raw) What topics are in that bag, and how many messages on each?"

Troubleshooting

• Connection fails or hangs: you are not on the tailnet, or the NUC is down. Check tailscale status shows cev-nuc as active.
• Tools error or come back empty: the server is up but the database may not be loaded yet. Tell the EVIL maintainers; it is not a problem with your setup.
• Windows + WSL: Cursor on Windows reads %USERPROFILE%\.cursor\mcp.json, not the WSL home directory; edit the Windows file instead.
• Codex asks about OAuth: the EVIL server has no auth. Update Codex, or set auth explicitly in your config.toml entry if your version requires it.

Notes

• evil-raw is a separate server on purpose: it gives free-form SQL over stored files, so it is kept off the port the web UI and the in-app assistant use. Everything that comes back from a recording file is untrusted data (the guidance in AGENTS.md says so).
• Verified against a real server: the endpoint and the original nine tools respond. The Claude Code setup is tested by the author. The Codex and Cursor configs follow those clients' current documentation but have not been run end to end; fix the file and open a PR if one is off.
• Only the two read ports (:8765 curated, :8767 raw SQL) are wired up here. EVIL's separate upload port is not part of this repo.
• Guidance lives once, in AGENTS.md. After editing it, run scripts/sync-skill.sh to regenerate the Claude skill (scripts/sync-skill.sh --check fails if they have drifted). Then commit and tell the team to run git pull && ./install.sh. Server-side tool changes need no announcement beyond "restart your client".