Connect
Overdrive, your efficiency agent
Clock in here. Copy-paste this into your agent, or Install here. Opening this page does not mean live ingest is on.
Copy-paste this into your agent
EasiestWorks in Claude, ChatGPT, Cursor, Claude Code, Codex, or any agent.
Set up Overdrive for me. Overdrive shows how my team uses its AI plans. 1. If you don't have Overdrive's tools yet, add the Overdrive connector (https://mcp.useoverdrive.ai/mcp) and I'll sign in when asked. 2. If you can run terminal commands and `roy connect` is installed, run it, show me the link and code it prints, and I'll click Approve in my browser. Don't open, print, or copy anything in ~/.roy. If it isn't installed, skip this step. 3. Ask me which AI plans I pay for (tool, plan, how many, price) and save them with Overdrive's declare_plans tool. Read back what you saved. 4. Offer to review my prompting in this chat with Overdrive's review_prompting tool. It sends counts, never my words. 5. Then ask Overdrive: which of my plans are underused? Tell me the answer in plain words. Never ask me to paste a password or secret key into chat.
Nothing secret is in this prompt, so it's safe to share with your team.
Or
Install here
Claude
- Click Add custom connector, or add Overdrive.
- Name it Overdrive and paste https://mcp.useoverdrive.ai
- Click Connect, sign in with your email, then click Allow.
Codex, ChatGPT, Cursor, OpenRouter
A desktop app for Mac and Windows is coming later.
Other ways to connect
Settings files, admin keys, and inviting a teammate
Command-line helper
roy connectFor people comfortable in a terminal.
The roy connect command reads Claude Code and Codex sessions on a computer, including past ones, and asks you to approve it in this browser. Your agent runs it for you when you use the prompt above. It isn't published yet.
Invite a teammate
ShareSend this to the person who manages your tools. It has no secrets. They sign in with their own email, and you'll see each connection here once data flows.
Can you connect our AI tools to Overdrive? Paste this into Claude Code, Cursor, or Codex: Set up Overdrive for me. Overdrive shows how my team uses its AI plans. 1. If you don't have Overdrive's tools yet, add the Overdrive connector (https://mcp.useoverdrive.ai/mcp) and I'll sign in when asked. 2. If you can run terminal commands and `roy connect` is installed, run it, show me the link and code it prints, and I'll click Approve in my browser. Don't open, print, or copy anything in ~/.roy. If it isn't installed, skip this step. 3. Ask me which AI plans I pay for (tool, plan, how many, price) and save them with Overdrive's declare_plans tool. Read back what you saved. 4. Offer to review my prompting in this chat with Overdrive's review_prompting tool. It sends counts, never my words. 5. Then ask Overdrive: which of my plans are underused? Tell me the answer in plain words. Never ask me to paste a password or secret key into chat.
Paste a read-only key from your admin settings
AdminsFor company plans. Overdrive reads seats, usage, and spend, and can't change anything.
Claude Enterprise
Overdrive does not store an Anthropic admin key. Connect Claude Code with telemetry above. Workload identity for Console reports is a later step.
Cursor Teams or Enterprise
- Open the Cursor dashboard as a team admin.
- Go to Settings, then API keys.
- Create a key with read-only access, and copy it.
- Paste it below.
OpenRouter
- Open openrouter.ai, then Settings, then Keys.
- Create a management key. It reads activity and can't run models.
- Paste it below.
Sign in to paste a key.
Claude Code settings file
TelemetryWorks on every Claude plan, including Pro and Max. Add it on each machine you use Claude Code on.
Sign in to create this connection.
Prompting signals (optional)
A prompt hook adds per-turn counts so Overdrive can coach how you prompt.
What the hook sends to Overdrive:
- conversation id and turn id
- model name
- time of each prompt and turn end
- whether a prompt exactly repeats the previous one
- whether a prompt closely rephrases the previous one (only with the rephrase opt-in)
- turn status (completed, error, aborted) and duration
Never sent:
- response text
- file paths and workspace folders
- attachments
- tool input and output
- commands
- transcripts
- Cursor version, loop counts, and effort settings
- prompt text, unless you opt in and turn on prompt capture in Overdrive
Kept on your computer:
- ~/.roy/local.key: a random key that never leaves this machine
- ~/.roy/state-claude-code.json: a keyed hash of the last prompt per conversation, and with the rephrase opt-in a keyed similarity sketch; no prompt text, dropped after a day
- ~/.roy/spool-claude-code.jsonl: events waiting to send; with the prompt-text opt-in, prompt text waits here unredacted until it is sent, and deleting the opt-in file strips it
Retry and rephrase detection: exact repeats are spotted with a keyed hash on your machine. Rephrase detection is opt-in and off until you turn it on. Token counts aren't available from hooks, so Overdrive shows them as not reported, not zero.
Script SHA-256: 2ff5585951f5a8ce908871b5850e8a6c8061b2da09f8bf640e661343f5a9a936. The setup page shows a one-line check.
Sign in to create this connection.
Codex settings file
TelemetryCodex CLI on a ChatGPT or API sign-in. Codex reports tokens but no cost.
Sign in to create this connection.
Cursor hooks
HooksCursor Pro and Ultra have no usage export. Hooks in ~/.cursor/hooks.json send a few fields for each agent run. Cursor's hooks don't report token counts, so Overdrive shows them as not reported.
Cursor hooks (Pro, Ultra)
What the hook sends to Overdrive:
- conversation id and turn id
- model name
- time of each prompt and turn end
- whether a prompt exactly repeats the previous one
- whether a prompt closely rephrases the previous one (only with the rephrase opt-in)
- turn status (completed, error, aborted) and duration
- your Cursor account email, which Overdrive stores only as a keyed hash
Never sent:
- response text
- file paths and workspace folders
- attachments
- tool input and output
- commands
- transcripts
- Cursor version, loop counts, and effort settings
- prompt text, unless you opt in and turn on prompt capture in Overdrive
Kept on your computer:
- ~/.roy/local.key: a random key that never leaves this machine
- ~/.roy/state-cursor.json: a keyed hash of the last prompt per conversation, and with the rephrase opt-in a keyed similarity sketch; no prompt text, dropped after a day
- ~/.roy/spool-cursor.jsonl: events waiting to send; with the prompt-text opt-in, prompt text waits here unredacted until it is sent, and deleting the opt-in file strips it
Retry and rephrase detection: exact repeats are spotted with a keyed hash on your machine. Rephrase detection is opt-in and off until you turn it on. Token counts aren't available from hooks, so Overdrive shows them as not reported, not zero.
Script SHA-256: 2ff5585951f5a8ce908871b5850e8a6c8061b2da09f8bf640e661343f5a9a936. The setup page shows a one-line check.
Sign in to create this connection.
OpenRouter Broadcast
BroadcastOpenRouter sends per-request tokens and cost to Overdrive. Turn Privacy Mode on so prompts are not sent.
Sign in to create this connection.
Grok Bot
Other hostsGrok Bot uses the same Streamable HTTP URL as Codex, ChatGPT, Cursor, and OpenRouter.
Add a custom MCP server named Overdrive at https://mcp.useoverdrive.ai/mcp using Streamable HTTP, then sign in when asked. Never paste a token or API key into chat.
Claude and ChatGPT chat
In chatclaude.ai and ChatGPT have no usage export, so there's nothing to connect here. Instead, in any chat where Overdrive is installed, ask "Overdrive, review my prompting in this chat". Your assistant sends counts about that one conversation (rephrases, retries, repeated pasted context, prompt length, model), not your text. Overdrive can't see your other chats or your history, and it doesn't keep the counts.
Sign in to allow in-chat reviews.
Overdrive reads usage counts, model names, timings and the provider's estimated cost. It doesn't store prompts or replies unless you turn that on for a connection. Overdrive keeps usage for 90 days, then deletes it. Sign in to turn off a connection or delete the usage it stored. Ending the session does not delete usage. Privacy · Support