@tasklite/mcp

@tasklite/mcp

A backend and an admin for apps built by AI agents — tasklite.net · docs

Describe a system and get the thing behind it: a project with typed tables andrelations, rows, REST endpoints, an API key, and an admin interface thebusiness itself operates afterwards. From Claude Code, Claude Desktop, ChatGPT,Gemini, Cursor or VS Code. Deploy a frontend onto it and hand the whole thingover.

Not the CLI task manager. There is an older, unrelated project also calledTaskLite — ad-si/TaskLite attasklite.org, a command-line task manager. This is adifferent product from a different author: a hosted backend attasklite.net, published as @tasklite/mcp and as net.tasklite/mcp inthe MCP registry.

The 48 tools

Every tool declares readOnlyHint, destructiveHint and openWorldHint, so aclient can tell what is safe to run unattended.

build_backend is the one to reach for first: it takes a description of asystem and creates the project, the boards, their typed columns including therelations between them, sample rows, and a published REST API with a key — inone call, instead of a dozen.

Setup

Install the package (and Claude Code if you don't have it), then add the serverby its binary. That avoids a known Windows issue where claude mcp addmis-parses npx -y.

npm install -g @anthropic-ai/claude-code @tasklite/mcp
claude mcp add tasklite -- tasklite-mcp

Add -s user to claude mcp add to make it available in every project(claude mcp add -s user tasklite -- tasklite-mcp); the default scope is thecurrent project only.

Then tell Claude what you want to build. The sign_up tool creates your account, organization, and connection from the conversation (a strong random password is generated locally and never shown; use "forgot password" with your email for web access).

Already have an account? Create a key at TaskLite → Integrations → "Connect Claude Code" and use:

claude mcp add tasklite -e TASKLITE_API_KEY=tl_xxx -- tasklite-mcp

Hosted (no install)

Point any MCP client at the hosted server; there is nothing to install locally:

claude mcp add --transport http tasklite https://mcp.tasklite.net/mcp \
  --header "Authorization: Bearer tl_xxx"

The hosted server is stateless: every request carries its own credential, soone endpoint serves every account safely.

Simplest of all: the connector. In Claude (claude.ai or the desktop app),Settings → Connectors → add TaskLite, or add it by address using the URL above.You sign in once over OAuth; there is no key to create or store. Note that thetools appear in a new chat, not in the conversation you were already in.

Optional env: TASKLITE_API_URL (default https://api.tasklite.net), TASKLITE_APP_URL (default https://app.tasklite.net).

Other clients

One hosted server, every MCP client. Full setup notes: https://tasklite.net/docs/guides/connector-from-cursor-codex-desktop

• Cursor — Add to Cursor or put {"mcpServers":{"tasklite":{"url":"https://mcp.tasklite.net/mcp"}}} in .cursor/mcp.json.
• VS Code — Install in VS Code or .vscode/mcp.json with {"servers":{"tasklite":{"type":"http","url":"https://mcp.tasklite.net/mcp"}}}.
• ChatGPT — Settings → Connectors (developer mode) → add https://mcp.tasklite.net/mcp. The server implements search and fetch.
• Gemini CLI — gemini extensions install https://github.com/shimon-ks/tasklite-mcp (this repo ships gemini-extension.json), or add httpUrl + oauth to ~/.gemini/settings.json.
• OpenAI Responses API / Agents SDK, Gemini API — pass the hosted URL with Authorization: Bearer tl_….

Typical flow (what Claude Code does)

1. create_project → create_board → create_column × N builds the schema.
2. create_item / query_items seed and inspect data.
3. create_app → create_app_endpoint (with exposedColumns + RLS) → create_app_api_key expose the REST surface for your frontend.
4. get_app_spec / get_frontend_prompt generate the frontend against it.
5. Every tool returns an adminUrl, the ready-made admin for the end client.

Hosting your frontend

deploy_frontend puts a static frontend on https://{slug}.tasklite.dev.No server, no hosting account, no CI to configure. Hand the site over in oneof three ways:

deploy_frontend(appId: "app-xxxxxx", files: [{ path: "index.html", content: "<!doctype html>…" }, { path: "app.js", content: "…" }])
deploy_frontend(appId: "app-xxxxxx", zipUrl: "https://github.com/you/site/releases/download/v1/dist.zip")
deploy_frontend(appId: "app-xxxxxx", dir: "./dist")

files is the path from ChatGPT or any hosted client: the assistant writes thepage and deploys it in the same turn. zipUrl takes an export from Lovable,Bolt or a GitHub release. dir is for an MCP running on the machine with thebuild output.

Hosted pages call the app API through the relative path /api/{endpoint}. Thehosting proxy attaches the app identity server-side, so the browser nevercarries an API key. list_deployments shows the versions androllback_deployment points the live URL back at an earlier one.

Security model

• The tl_ key is exchanged for a short-lived JWT (POST /public/v1/auth/session); all calls run with the key owner's own permissions, never super-admin.
• App API keys belong in server-side env vars (Next.js API routes), never in browser code. Frontends hosted on tasklite.dev need no key at all.

Development

npm install
npm run build
TASKLITE_API_KEY=tl_xxx TASKLITE_API_URL=http://localhost:3333 node dist/index.js

License

MIT