> ## Documentation Index
> Fetch the complete documentation index at: https://docs.corsair.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Corsair MCP

> One address, https://mcp.corsair.dev/mcp, that lets Claude, ChatGPT, Cursor and other agents use the apps you've connected to Corsair.

Corsair MCP lets an AI agent use the apps you've connected to Corsair. It can read your inbox, post in Slack, open a GitHub issue, search data Corsair has synced, hand a file to one of your apps, or set up an automation that runs when something happens in one of your apps.

Every client uses the same address:

```
https://mcp.corsair.dev/mcp
```

You add it once, sign in to Corsair, and choose which of your links the client uses.

## Before you start

* A Corsair account. Sign up at [hub.corsair.dev](https://hub.corsair.dev) with Google, GitHub or an emailed sign-in link.
* Apps to use. You can connect them in Corsair first, or ask the agent to connect one later. It hands you a sign-in link for that app.

## How sign-in works

**1. Add the address to your client**

See [Add it to your client](#add-it-to-your-client) below.

**2. Sign in to Corsair**

The client opens Corsair in your browser.

**3. Pick a link**

A link is a set of your connected apps plus a [mode](#modes-and-approvals) that decides what the agent may do without asking. Pick an existing link or make a new one for this client. If you have no link yet, Corsair asks you to make one first.

**4. Allow**

The client gets the apps on that link and your email address. Nothing else on your account is shared.

One connection uses one link. To switch links, remove Corsair from the client, add it again and pick the other link. You can sign a client out of a link at any time from [hub.corsair.dev/me](https://hub.corsair.dev/me).

## Add it to your client

> [tab] Claude
>
> 1. In Claude, open **Customize → Connectors → Add custom connector**.
> 2. Name it Corsair, paste `https://mcp.corsair.dev/mcp` and add it.
> 3. Select **Connect** and sign in to Corsair.
>
> On a Team or Enterprise plan, an owner adds connectors under Organization settings.

> [tab] ChatGPT
>
> 1. In ChatGPT, add a custom MCP server and paste `https://mcp.corsair.dev/mcp`.
> 2. Choose OAuth, then sign in to Corsair.
>
> On personal plans, ChatGPT currently uses only the read tools from a custom server, so Corsair can look things up but not make changes. Write tools such as `run_script` need a Business or Enterprise/Edu workspace, where an admin adds the server or gives you permission to. OpenAI's [guide to custom MCP servers](https://help.openai.com/en/articles/12584461-connecting-custom-mcp-servers-in-chatgpt) has the current rules.

> [tab] Cursor
>
> [Add to Cursor](cursor://anysphere.cursor-deeplink/mcp/install?name=corsair\&config=eyJ1cmwiOiJodHRwczovL21jcC5jb3JzYWlyLmRldi9tY3AifQ==), or add this to `~/.cursor/mcp.json`:
>
> ```json theme={null}
> {
>   "mcpServers": {
>     "corsair": { "url": "https://mcp.corsair.dev/mcp" }
>   }
> }
> ```
>
> Cursor opens Corsair to sign in the first time it connects.

> [tab] VS Code
>
> Add this to `.vscode/mcp.json`, or run **MCP: Add Server → HTTP** and paste the address:
>
> ```json theme={null}
> {
>   "servers": {
>     "corsair": { "type": "http", "url": "https://mcp.corsair.dev/mcp" }
>   }
> }
> ```
>
> VS Code signs you in when it first connects.

> [tab] Claude Code
>
> ```bash theme={null}
> claude mcp add --transport http corsair https://mcp.corsair.dev/mcp
> ```
>
> Then run `/mcp` in Claude Code to sign in. Add `--scope user` to have it in every project.

> [tab] Codex
>
> ```bash theme={null}
> codex mcp add corsair --url https://mcp.corsair.dev/mcp
> ```
>
> Codex usually opens Corsair to sign in as soon as you add it. If it says the server may or may not require login instead, or you need to sign in again later, run `codex mcp login corsair`.

> [tab] Gemini CLI
>
> ```bash theme={null}
> gemini mcp add --transport http corsair https://mcp.corsair.dev/mcp
> ```
>
> Then run `/mcp auth corsair` to sign in.

> [tab] Zed
>
> Add this to your Zed settings. Zed signs you in when it first connects.
>
> ```json theme={null}
> {
>   "context_servers": {
>     "corsair": { "url": "https://mcp.corsair.dev/mcp" }
>   }
> }
> ```

Any other client that supports remote MCP servers with OAuth sign-in works the same way.

> [!NOTE]
>
> Some clients can't sign in to an MCP server, such as GitHub Copilot's coding agent. For those, use a link's own address with an API key. Open the link at [hub.corsair.dev/me](https://hub.corsair.dev/me) and follow the steps for your client there.

## What the agent can do

| Tool | What it does | Type |
| - | - | - |
| `list_integrations` | Lists the apps on this link, whether each is ready and which account it's signed in as, with the Corsair credits left and spent this month. Also checks whether Corsair can connect an app that isn't here yet | Read |
| `get_connect_link` | Returns the sign-in link for an app, adding the app first if it isn't here. Apps that run on Corsair's own account come back ready, with no link. Shows a [connect card](#cards) where the chat supports it | Write |
| `list_operations` | Shows what connected apps can do, with each action's risk and price. Searched by what you want to do, it returns the matching actions with the arguments they require | Read |
| `get_schema` | Shows the full arguments and return shape of one or more actions | Read |
| `search` | Answers a question about your data, usually in one call: Corsair's own PM, CRM, routines and files, and the email, chat, issues, docs and calendars of your connected apps. When an app's synced copy is behind, it reads that app live. When nothing matches, it says so instead of offering the closest records as an answer. It also opens results in full | Read |
| `read_data` | Reads your apps live with a short script, for what `search` doesn't have or what must be current. It stops before any action that would change something | Read |
| `run_script` | Makes changes in your apps with a short script: send, create, update or delete | Sensitive write |
| `get_job` | Waits for a [long job](#long-jobs) a script started, such as a crawl or a document parse, and returns its result. Shows a [job card](#cards) where the chat supports it | Read |
| `upload_file` | Adds a file to your [Files](#files) so your apps can use it: a file you attached in the chat, a public URL or a file on your computer. With none of these, it gives you a link to pick the file. It only adds files and never replaces one. Shows an [upload card](#cards) where the chat supports it | Write |
| `view_routines` | Lists the routines on this link and their recent runs, or shows one run step by step with its error | Read |
| `update_routine` | Creates, changes, pauses or removes a routine, an automation that runs on an app event or a schedule. Describe it in plain words and Corsair writes and tests it | Sensitive write |
| `use_skill` | Loads one of your saved [skills](#skills) by name: its `SKILL.md`, then any file that comes with it | Read |
| `find_skill` | Searches the public [skills.sh](https://skills.sh) directory for skills you haven't saved yet. Shows a [skills card](#cards) where the chat supports it | Read |
| `save_skill` | Saves a skill for every agent on your link, from text, a GitHub URL or a skills.sh id. For a folder on your computer, such as `~/.claude/skills`, it gives you one command that uploads it | Write |
| `submit_feedback` | Sends the Corsair team a short note about Corsair itself. The agent sends one on its own when Corsair lacks an app or action you asked for, or when a reply says the fault is on Corsair's side, and tells you it did. For a wrong or confusing answer it asks you first. A note never includes your app data | Write |

Read tools never change anything in your apps. A read can still use [credits](#corsair-credits): `read_data` on an app that runs on Corsair's account is billed like any other call there. Write tools change only your Corsair setup and files, which you can undo, or send Corsair a note. Sensitive writes act in your apps, or set up a routine that will act later without asking, so a client should ask you before every call.

Scripts run in a sandbox that can only call your connected apps' actions. They have no network, filesystem or other system access, and the agent never sees your app credentials.

Not every link shows every tool. A link that isn't allowed to add apps has no `get_connect_link`, and a link that can't change routines has no `update_routine`. The skill tools come on the links you make at [hub.corsair.dev/me](https://hub.corsair.dev/me), and a link in Read mode has no `save_skill`.

`list_integrations`, `list_operations`, `get_schema`, `search`, `read_data`, `get_job`, `view_routines`, `use_skill` and `find_skill` are marked read-only, so most clients run them without asking. Clients usually ask you before running the others.

## Long jobs

Some actions start work that takes longer than one call: a site crawl, an async document parse, an agent run, an image or video prediction. Corsair notices when a script starts one and waits a few seconds. If the job finishes in that time, the result comes back with the script's. If not, the reply names the job, and the agent calls `get_job` to wait for it and get the result. The reply also gives a line to stop the job, where the app has one.

This works whether the app is connected with your own account or runs on Corsair's.

## Files

Every link has Files, where Corsair keeps files for your apps to use, up to 25 MB each. A script can't carry a file itself, so the agent first adds the file with `upload_file`, then passes its id (`fil_…`) to the app. Corsair gives the app a short-lived link to the file in its place.

How a file gets in depends on where you're chatting:

* **ChatGPT.** A file you attach goes straight to Files.
* **Claude.** The agent sends a file you attached from its code sandbox with one command. That needs the sandbox to reach Corsair: in claude.ai, add `auth.corsair.dev` under **Settings > Capabilities > Allowed websites**. Until you do, you pick the file in the upload card instead.
* **An agent with a shell**, such as Claude Code or Codex, sends a file from your computer with one command. The file goes straight to Corsair without passing through the chat.
* **Anywhere else**, you pick the file in the upload card, or on the upload page its link opens. The link works for 10 minutes.
* **A public URL**: Corsair fetches it.

`upload_file` only adds files, and sending the same file twice keeps one copy. A file is private to the link that added it until you share it with your workspace. In a script, the agent can read text files and the text of PDFs, change part of a text file, and save an app's output, such as a filled-in PDF, as a new file or a new version of one. Search finds files by name, and by what they say when they have text.

See, share and delete your files on your link's **Files** page: open the link at [hub.corsair.dev/me](https://hub.corsair.dev/me).

## Corsair's own apps

Besides the apps you connect, the agent can use apps that Corsair runs itself. They need no other account.

* **PM**: issues and projects, with a status, priority, assignee, labels and a due date.
* **CRM**: people, companies and deals, with notes and tasks.
* **pg**: your own Postgres databases, queried by name. See [Postgres databases](#postgres-databases).
* **Files**: files your apps can use. See [Files](#files).
* **Mate**: the chat assistant in Corsair. The agent can find and read your Mate chats, including ones in Shared.

Each one has a page on your link at [hub.corsair.dev/me](https://hub.corsair.dev/me).

### Postgres databases

pg lets the agent run SQL on your own Postgres databases without ever seeing their connection URLs.

* **Adding a database.** Ask the agent to add one, and it gives you a setup link. The workspace owner opens it within 30 minutes and pastes the connection URL there. Corsair stores the URL encrypted; it never passes through the chat.
* **Reading and writing.** The agent runs one statement at a time. A read runs in a transaction that is always rolled back, so nothing it does persists, and returns up to 1,000 rows. A write commits, and needs write access. A database marked read-only takes reads only, for everyone.
* **Access.** The owner chooses which teammates can read or write each database on the link's **pg** page, and its **Audit** tab lists every query.
* **Limits.** Per database: 10 queries per 10 seconds and 60 per minute for each person, and 120 per minute for everyone together.

## Skills

A skill is a set of instructions an agent loads when a task calls for it: a `SKILL.md`, sometimes with files it refers to. Save a skill to Corsair once and every agent on the same link can use it, in whichever client it runs. Corsair keeps your skills in your own Corsair instance.

* **Use one.** Each agent sees the list of your skills when it connects, and loads one with `use_skill` when a task matches it. Clients that show MCP prompts also list each skill as a prompt.
* **Find one.** Ask the agent for a skill, such as "find me a skill for writing PR descriptions". It searches skills.sh with `find_skill`, and you pick the one to add.
* **Save one.** Ask the agent to save a skill you wrote together in the chat, or one from GitHub or skills.sh. Saving your own skill again under the same name adds a new version.
* **Upload your own.** To add the skills you keep on your computer, such as `~/.claude/skills`, ask the agent to upload the folder. It gives you one command to run in your terminal, and an agent with a shell, such as Claude Code, can run it for you. The files go straight to Corsair without passing through the chat. The command works for 10 minutes.

Corsair keeps only text files and removes any keys or tokens it finds when it saves a skill. A library holds up to 100 skills, and one skill up to 200 files and 8 MB.

See, switch off or roll back your skills at [hub.corsair.dev/me/skills](https://hub.corsair.dev/me/skills).

## Cards

In chats that show MCP Apps, such as Claude, four tools add a small card to the conversation:

* **Connect card** (`get_connect_link`): the app's logo, its name and one button. **Connect** opens the sign-in. The card shows when the sign-in has landed and offers **Continue** so the agent carries on. If the link runs out (after 10 minutes) it shows **New link** and gets a fresh one itself.
* **Job card** (`get_job`): a job that is still running, which keeps checking until the job finishes, fails or is cancelled, then offers **Continue**.
* **Skills card** (`find_skill`): the skills it found, each with an **Add** button that saves it to your library. It shows only on links that can save skills.
* **Upload card** (`upload_file`): a place to choose or drop files. It shows each file as it's saved, then offers **Continue**. When the agent is sending the file itself, the card says so and still lets you pick it if that doesn't work.

Clients without cards get the same information as text, so nothing is lost.

## Apps on Corsair's account

Some apps, such as Firecrawl, Exa, Reducto and TinyFish, can run on Corsair's own account with nothing to sign in to. Each call is billed to your organization's Corsair credits. `list_operations` and `get_schema` give each action's price.

## Corsair credits

Each organization gets free credits. To add more, open **Upgrade** in your Corsair account at [hub.corsair.dev/upgrade](https://hub.corsair.dev/upgrade). The agent doesn't link you there itself.

`list_integrations` gives the credits left, when the free ones come back and what this project spent this month. When credits run low, the agent says so in its next reply. When they run out or a spend limit is reached, the call doesn't run, nothing is charged and the agent tells you.

When a call fails at the app, Corsair gives its credits back. That refund is best effort: if it can't be recorded, the charge stays, so email [primary@corsair.dev](mailto:primary@corsair.dev) about a failed call you were charged for.

## Modes and approvals

Each link has a mode. Corsair checks every action against it, whatever the client allows.

| Mode | What it allows |
| - | - |
| Read | Reads only. Writes are refused. |
| Cautious (default) | Writes go through. A destructive action, such as a delete, needs your approval. |
| Strict | Every write needs your approval. Destructive actions are refused. |
| Open | Everything, without asking. |

When an action needs approval, the agent gives you a link. It shows exactly what will run and with which arguments. Approve or deny it there, then tell the agent to carry on. An approval covers that one action once.

Change a link's mode at [hub.corsair.dev/me](https://hub.corsair.dev/me).

## Troubleshooting

> [accordion] The agent says an app isn't connected
>
> Ask it to connect the app. It calls `get_connect_link` and gives you a sign-in link, or in Claude a card with a **Connect** button. Open it, sign in to the app, then tell the agent you're done or press **Continue** on the card. A few apps Corsair provides directly come back ready with no link, and the agent can use them right away.

> [accordion] The sign-in link stopped working
>
> Links last 10 minutes. On the card, press **New link**. Without a card, ask the agent for a new link.

> [accordion] I want to tell Corsair something went wrong
>
> Ask the agent to send Corsair feedback. It calls `submit_feedback` with a short note in its own words, plus the error's reference code when the note is about an error, without anything from your apps. When Corsair can't do what you asked, or an error is on Corsair's side, the agent sends the note on its own and tells you. Or email [primary@corsair.dev](mailto:primary@corsair.dev).

> [accordion] An app needs setting up before it can be connected
>
> The project behind your link has to finish setting the app up first, for example by adding its OAuth credentials. Whoever manages the project does this in Corsair.

> [accordion] An action is blocked by the permission mode
>
> The link's mode doesn't allow that kind of action. Change the mode at [hub.corsair.dev/me](https://hub.corsair.dev/me), or use a link with a different mode.

> [accordion] I want this client to use a different link
>
> Remove Corsair from the client, add it again and pick the other link when you sign in.

> [accordion] I added [https://mcp.corsair.cloud/mcp](https://mcp.corsair.cloud/mcp) earlier
>
> That address still works and reaches the same server, so nothing needs to change. New setups use `https://mcp.corsair.dev/mcp`.

## Support and privacy

Email [primary@corsair.dev](mailto:primary@corsair.dev). See the [privacy policy](https://corsair.dev/privacy-policy) and [terms](https://corsair.dev/terms-of-service).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.