documentation
Caipi docs
Caipi is a hosted, file-based workspace your AI agent writes into over
MCP.
Agents read and write files, run SQL, and build small SQLite-backed web
apps; you browse and use the result in a web UI at
caipi.ai/{username}. Workspaces are private by
default, and you choose what to publish.
Connect your agent
Caipi is a remote MCP server over Streamable HTTP. Point any MCP-capable client at the server URL below. There is no API key to paste: the first connection opens a browser sign-in (Google is the identity provider behind it), and Caipi registers your client automatically over OAuth. Sign in with the same account that owns your workspace.
https://caipi.ai/mcp
Apps
Chat clients with an official connector UI. You add Caipi from inside the app, then complete the browser sign-in.
Claude (web & desktop)
Open Settings → Connectors → Add custom connector,
paste https://caipi.ai/mcp, click
Add, then complete the browser sign-in. Leave any
advanced or client-ID fields blank; Caipi registers automatically.
Available on Free, Pro, Max, Team, and Enterprise plans.
ChatGPT (web & desktop)
Open Settings → Apps & Connectors → Advanced
settings and turn on Developer Mode. Back in
Apps & Connectors, choose Create, enter a name and
the MCP server URL https://caipi.ai/mcp,
select OAuth, then sign in. Developer Mode must be enabled, and
ChatGPT supports remote servers only.
Coding tools
CLI and config-file clients. Run the command or drop the snippet in, then complete the browser sign-in.
Claude Code
claude mcp add --transport http caipi https://caipi.ai/mcp
Add --scope user to use it across all
projects. Then run /mcp inside Claude Code
and complete the browser sign-in.
OpenAI Codex
codex mcp add caipi --url https://caipi.ai/mcp codex mcp login caipi
Requires a recent Codex CLI; its HTTP MCP support is newer and still experimental, so update Codex if the connection hiccups. Do not set a bearer token.
Cursor
Add to ~/.cursor/mcp.json (or a project
.cursor/mcp.json):
{
"mcpServers": {
"caipi": { "url": "https://caipi.ai/mcp" }
}
}
Or via the UI: Settings → Tools & MCP → add a new MCP server, paste the URL, then click Needs login to sign in.
opencode
Add to opencode.json (project root) or
~/.config/opencode/opencode.json:
{
"$schema": "https://opencode.ai/config.json",
"mcp": {
"caipi": { "type": "remote", "url": "https://caipi.ai/mcp", "enabled": true }
}
}
If the browser sign-in doesn't open automatically, run
opencode mcp auth caipi.
Any other MCP client
Any MCP client can connect to the same URL,
https://caipi.ai/mcp, as a remote server
over Streamable HTTP. Clients without native remote-OAuth support can
reach it through an
mcp-remote
bridge.
What your agent can do
Once connected, your agent has ten tools, grouped into reading and
writing. All paths are workspace-relative (like
notes/todo.md), never local filesystem
paths.
Read
read_file— read a text file or image, or fetch metadata only.find_files— list or search the workspace by path and text; returns up to 50 files.query_sql— run one read-only SQLite statement (SELECT or PRAGMA) against a.sqlitefile.get_upload_status— check a browser upload link and see which files arrived.
Write
write_file— create or overwrite a text file (.md,.html,.js,.css,.json,.sql).execute_sql— run SQLite write statements (DDL and DML) against a.sqlitefile.delete_file— delete a file.move_file— move or rename a file.request_upload— create a short-lived browser link so you can hand off files from your device.set_visibility— set a file or folder to private, public, or password-protected.
There is also a build_app prompt: invoke it
for the full, copy-pasteable guide to building a SQLite-backed app.
The workspace model
A workspace is a tree of plain files. Markdown and HTML files render as
pages; a folder serves its index.html or
index.md like a small web server. SQLite
databases get a read-only table browser. Uploads (PDFs, images,
archives) sit alongside the rest.
Every file or folder has one of three visibility tiers: private (owner only, the default), public (reachable by anyone), or password-protected (a browser prompt gates access). A folder rule cascades to everything inside it.
Published content is served from a separate content origin,
share.caipi.ai, so shared pages and apps run
isolated from your authenticated session on the main site.
Building apps
A Caipi app is a folder with an index.html
and a manifest.json. The manifest declares
named SQL operations; the page calls them with a plain
fetch to ./api/{op},
and Caipi runs the SQL server-side. The browser never sends raw SQL, so
a public app exposes only the operations you declare. Apps follow normal
file visibility, private by default. For the complete, copy-pasteable
recipe, invoke the build_app prompt from your
agent.
Limits
The free tier includes 250 MCP tool calls per week (the counter resets Monday at 00:00 UTC) and 50 MB of workspace storage. Pro is coming soon.
FAQ & troubleshooting
My agent can't connect.
Confirm the server URL is exactly
https://caipi.ai/mcp and that your client
supports remote MCP over HTTP. The first connection must complete the
browser sign-in; sign in with the account that owns the workspace. You
can review and revoke connected agents anytime under
Settings in your workspace.
Where did my file go? How do I find content?
Everything lives in your workspace at
caipi.ai/{username}. Browse the folder tree
there, or ask your agent to run find_files to
list or search by name and text.
How do I make something public?
Ask your agent to call set_visibility on the
file or folder, or use the publish control in the workspace UI. Public
content is served from share.caipi.ai.
How do I delete my account or data?
Open Settings in your workspace to export a zip of your files or delete your account. See the privacy policy for what deletion covers and how to make a request by email.
Your data
Your files live in your workspace on Caipi's hosting and stay private unless you publish them. You can export everything as open-format files from Settings, and delete your account and workspace from the same place. For details on what we store, retention, and how to make a request, see the privacy policy.