# Overdrive

> Overdrive, your efficiency agent

Teams already run a software factory across Claude, Codex, Cursor, and OpenRouter. Overdrive is how you see which stations and habits pay off, with a receipt. Overdrive is not the factory and does not run the agents. Works with Claude, ChatGPT, and Grok Bot.

Overdrive finds where agents waste effort across engineering, sales, finance, and customer support. Every figure can be traced to its source.

## Server

- Claude: https://mcp.useoverdrive.ai
- Codex, ChatGPT, Cursor, OpenRouter: https://mcp.useoverdrive.ai/mcp
- Grok Bot and other Streamable HTTP clients: that same URL, https://mcp.useoverdrive.ai/mcp

## Connect usage

Clock in at the public page: https://mcp.useoverdrive.ai/connect
Two alternatives, not numbered steps:
- Copy-paste this into your agent
- Install here
Claude is its own connector; Codex, ChatGPT, Cursor, and OpenRouter share https://mcp.useoverdrive.ai/mcp.
If you added Overdrive in your agent, ask Overdrive in that chat. It should offer those two choices and that URL. Do not ask for a CSV. Opening that page does not mean live ingest is on.
The collector and hook POST to a fixed host and do not follow redirects. Reinstall after the move so ingest is https://mcp.useoverdrive.ai/collector/v1/requests and hook events are https://mcp.useoverdrive.ai/hooks/v1/events.

## Docs

- [Docs](https://useoverdrive.ai/docs): Hub
- [Connect](https://useoverdrive.ai/docs/connect): Clock in. Claude as its own connector; Codex, ChatGPT, Cursor, and OpenRouter share one MCP URL
- [Sign-in](https://useoverdrive.ai/docs/auth): OAuth, the email magic link, tokens, and how to revoke
- [Tools](https://useoverdrive.ai/docs/tools): What each tool is for
- [Your data](https://useoverdrive.ai/docs/data): Bring operating data with agent fields or a CSV
- [Troubleshooting](https://useoverdrive.ai/docs/troubleshooting): Connect, magic link, revoke, packs, and ask
- [Privacy](https://useoverdrive.ai/privacy)
- [Support](https://useoverdrive.ai/support): support@useoverdrive.ai
- [Terms](https://useoverdrive.ai/terms)

## Auth

OAuth with PKCE. Claude Connect is Connect, then an email magic link, then Allow. There is no Allow step that skips the email link. Overdrive does not have a separate password account. Do not paste a Bearer token or an API key.

- Access token: 1 hour
- Refresh token: 14 days, rotating
- Disconnecting Overdrive in Claude does not end the Overdrive session. There is no /oauth/revoke endpoint.
- Overdrive logout: POST https://mcp.useoverdrive.ai/logout or POST https://mcp.useoverdrive.ai/logout/all. After that, the next MCP call with the old access token is 401.

## Tools

- ask (Ask ROI): Answers one ROI question from the loaded pack, or from the pack you name, using before, after, or period as YYYY-MM-DD:YYYY-MM-DD, and a pack name that does not load is an error.
- connect_usage (Connect usage): Returns the two ways to connect usage, the steps for this agent, and the connect page, after checking that the caller is signed in.
- declare_plans (Declare plans): Saves a plan the caller stated, with provider, plan name, holders, an optional price and currency, and a date window that defaults to this calendar month.
- delete_requests (Delete request content): Deletes stored request content on one pack for one tenant or one person_key, plus the derived copies Overdrive can name for that pack (excerpts, findings, export hash fields, stored pack copies, and server-side export files), and fails if a named copy can't be removed. It needs an admin or self grant on the pack, and a self grant can delete only the caller's own person_key. It does not recall copies already delivered, delete provider backups, or delete usage from live connections, which is deleted on the connect page.
- disconnect_usage (Disconnect usage): Stops accepting new data from one connection, and usage already stored stays for up to 90 days.
- export_otlp (Export OTLP findings): Returns an OTLP document of this caller's visible request findings, does not include prompt text, and does not write a server file.
- export_pack (Export pack): Returns the pack's raw uploaded tables and needs an admin grant on the pack.
- lineage (Lineage): Opens one published number from the last ask, run_roi, or roadmap scan and returns its source row location.
- load_pack (Load pack): Opens one company pack by directory name under the pinned pack root and clears any previous report.
- review_prompting (Review prompting in this chat): Reviews this conversation from counts only, with no prompt text, and is consent-gated.
- revoke_requests (Revoke request content): Turns off request-content collection on one pack for one tenant or one person_key and strips stored content, excerpts, and content hashes for that scope. It needs an admin or self grant on the pack, and a self grant can name only the caller's own person_key.
- roadmap (Roadmap): Sequences next moves from published findings on the loaded pack, from the last ask or run_roi or from lenses you pass.
- run_roi (Run ROI): Runs one named lens on the loaded pack and returns that report.
- upsert_pack (Upsert pack): Stores one normalized JSON pack for the verified principal and refuses read-only tour packs.
- usage_connections (Usage connections): Lists the caller's connections and declared plans, with label, tool, plan, status, and when data last arrived.

## Data

Bring operating data through agent fields or a CSV. northline and halyard are sample packs for a tour.
A temporary allowlist may still limit which pack names the live host accepts.
Overdrive answers from the pack loaded for this sign-in. This file does not claim that a shared host isolates every customer.

## Writing

- [Blog](https://useoverdrive.ai/blog): Where AI spend pays off
- [Why AI ROI Beats Another Dashboard](https://useoverdrive.ai/blog/why-roi-beats-dashboards): Illustrative essay. No customer results. Samples are a tour, not a limit.
- [The true cost of reasoning models vs cheap models](https://useoverdrive.ai/blog/true-cost-reasoning-vs-cheap-models): Expensive reasoning is often overkill, and cheap or open-weight models carry costs too: retries, review time, failure handling. How to weigh both.
- [Agent loops vs long-running work worth paying for](https://useoverdrive.ai/blog/agent-loops-vs-justified-long-work): Long agent runs aren't always waste, and short ones aren't always fine. How to separate loops from justified long work with session evidence.
- [Tool errors, handoffs, and when an AI agent should stop](https://useoverdrive.ai/blog/agent-tool-errors-handoffs-stopping): Failed tool calls and missed handoffs quietly add spend. How to find where agents break, hand off, or should stop, with cited rows.
- [Repeated context and caching: paying for the same tokens](https://useoverdrive.ai/blog/repeated-context-and-caching): Resending the same context on every call adds up. How to spot repeated context in your usage and check whether caching applies.
- [AI prompting best practices: structure one prompt well](https://useoverdrive.ai/blog/ai-prompting-best-practices): How to word and structure a single prompt: goal, context, constraints, output format, examples. Practical habits, checked against your own results.
- [Requirements and task decomposition for AI agents](https://useoverdrive.ai/blog/requirements-task-decomposition-ai-agents): Vague asks turn into long, costly agent runs. How clearer requirements and smaller tasks show up in your usage evidence.
- [Unused AI plans, overlap, and limit-driven routing](https://useoverdrive.ai/blog/unused-overlapping-ai-plans-routing): Seats nobody uses, plans that overlap, and work routed by limits not fit. How to see it in attributable spend with cited evidence.
- [When an existing tool beats an AI agent](https://useoverdrive.ai/blog/existing-tool-instead-of-ai-agent): Some agent work is a job a script, query, or tool you already own does fine. How to spot those tasks with evidence, not instinct.
- [Before and after: measuring AI quality, cost, and time](https://useoverdrive.ai/blog/measure-ai-before-after-quality-cost-time): A change only counts if you can show it. How to compare quality, cost, and time before and after, with receipts tied to source rows.

## Directory listing (draft, not submitted)

Name: Overdrive
OpenAI short (26 characters): Overdrive, your efficiency agent
Claude one-liner: Overdrive, your efficiency agent. Teams already run a software factory across Claude, Codex, Cursor, and OpenRouter.
Long: Overdrive, your efficiency agent. Teams already run a software factory across Claude, Codex, Cursor, and OpenRouter. Clock in at https://mcp.useoverdrive.ai/connect. Claude is its own connector; Codex, ChatGPT, Cursor, and OpenRouter share https://mcp.useoverdrive.ai/mcp. Overdrive does not rank people, invent savings, run the agents, or read live account usage while flags are off.

## Hosts

- Claude: today. URL https://mcp.useoverdrive.ai
- Codex, ChatGPT, Cursor, OpenRouter: today. One shared URL https://mcp.useoverdrive.ai/mcp
- Grok Bot and other MCP clients: that same URL, https://mcp.useoverdrive.ai/mcp, and OAuth