MCP server de 2ools

2ools MCP Server

Remote MCP server for 2ools — build, version, review and export websites, web apps and games from inside the AI chat you already use.

https://2ools.app/mcp

Your assistant writes the site. 2ools is where it lives: every result becomes a restorable version, previews are shareable links that need no account to open, and the source exports as plain files with no proprietary runtime.

One tool is free and needs no account at all — create_public_preview turns a standalone HTML document your model just wrote into a real, viewable URL. The rest connect over OAuth 2.1 with PKCE and least-privilege scopes.

---

Quick start

Paste this into the AI client you want to connect:

Connect this client to 2ools (github.com/2ools/mcp-server, MCP endpoint https://2ools.app/mcp). Retrieve https://2ools.app/start non-visibly with an HTTP or web-fetch tool; never navigate the in-app browser to that instructions page. Use this client's in-app browser for every 2ools sign-in and authorization page; never use Chrome, another external browser, or the system browser when this client has an in-app browser or preview pane. The exact OAuth approval URL must be the first 2ools page you visibly open. Before I must act, your only user-facing setup message may be Connecting 2ools in this client's browser…. Do not output insight blocks or narrate setup mechanics. Ask me exactly once to sign in and click Authorize. Do not end your turn while authorization is pending; after it succeeds, keep the 2ools app visibly open in that pane. Fetch https://2ools.app/start and follow its "For your AI" steps exactly.

The live start page is the single current connection ritual for every client. It detects the current client path, keeps OAuth alive, and leaves only sign-in and the Authorize click to the person. Manual MCP configuration is a fallback only for clients that cannot manage their own remote-MCP setup.

After authorization, the initiating context selects the app destination: an exact project's Project OS when its ID is known, the new-project or existing-project workspace flow when that intent is known, and the authenticated MCP workspace for a generic install. The public homepage is never the fallback.

Registry introspection proxy

Registries that require a local stdio process can run the included credential-free proxy. It forwards MCP JSON-RPC traffic to the same verified remote endpoint; it does not mock or reimplement any 2ools tools. It translates the stdio handshake into the remote endpoint's current per-request MCP envelope and returns the live schemas from 2ools.

npm ci
npm start

The included Dockerfile runs the same dependency-free proxy for reproducible registry introspection. It intentionally does not perform an interactive OAuth flow. Clients using authenticated workspace tools should connect directly to https://2ools.app/mcp so OAuth remains end-to-end with 2ools.

2ools Skill

For assistants that support repository skills, add or copy the 2ools Skill. It keeps the temporary-preview and OAuth workspace boundaries explicit. Client support for remote MCP and Skill folders varies; the live start page routes each client to its supported flow.

Try it with no account

Ask your assistant:

Write this as one standalone HTML document, create a free 2ools preview, and show me the link.

You get back a real URL that anyone can open. It expires after 24 hours.

---

What you can do with it

Ask your assistant…	It uses
"Open this generated site in 2ools and give me a client-safe preview link"	create_public_preview
"Save this conversation as a durable 2ools project"	build_from_conversation
"Make the hours bigger and show me the new version"	refine_project
"Compare the last two versions"	compare_versions
"What changed on this project since Tuesday?"	get_project_activity
"Export the source as a zip"	prepare_version_export

---

Tools

43 tools, each with explicit readOnlyHint and destructiveHint annotations so a client can reason about blast radius before calling anything.

Free — no account required

Tool	Effect	What it does
create_public_preview	write	Turn a complete standalone HTML document already created in this chat into a temporary, read-only 2ools preview.

Build and refine

Tool	Effect	What it does
build_from_conversation	write	Create a new 2ools project from a structured conversation brief and queue one durable build.
refine_project	write	Create one protected child version from a saved base version and concrete review feedback.
get_build	read-only	Get one build's current status, progress, errors, completed version IDs, and protected preview URLs.
cancel_build	destructive	Cancel a queued or running non-Agent 2ools build.
compare_versions	read-only	Compare two saved versions in one project without returning source.
list_build_systems	read-only	List the Curators, Rulebooks, and Skills this paid account may explicitly apply.

Read a project

Tool	Effect	What it does
list_projects	read-only	List recent 2ools projects the authenticated account can access, including owned and shared projects.
get_project	read-only	Get a 2ools project's brief, selected pages, active version, and compact version history.
get_project_context	read-only	Read one source-free operating packet before acting: the active version, exact frozen Curator snapshot, revision-…
render_project_preview	read-only	Render an interactive, read-only card for one saved 2ools version.
get_project_activity	read-only	Read the attributable project timeline across human work, external MCP agents, builds, versions, feedback, and…
wait_for_project_activity	read-only	Wait up to 20 seconds for attributable project events newer than a head_cursor or prior next_cursor.

Review and approval

Tool	Effect	What it does
list_approval_requests	read-only	List exact saved versions waiting for a separately delegated reviewer.
approve_version	destructive	Use separately delegated review authority to approve one saved version and make it the project's active version.
reject_version	destructive	Use separately delegated review authority to reject one saved version.

Feedback and shared notes

Tool	Effect	What it does
list_project_feedback	read-only	Read page-anchored review feedback for versions in a project, with explicit human or external-agent attribution.
add_project_feedback	write	Add an attributable review comment to a saved project version.
resolve_project_feedback	destructive	Owner-only review decision that resolves or reopens one saved feedback item.
list_project_pins	read-only	Read shared workspace notes and http(s) links, with explicit attribution.
add_project_pin	write	Add an attributable note or safe http(s) link to the shared project workspace.
remove_project_pin	destructive	Remove one shared workspace pin.

Export and sharing

Tool	Effect	What it does
prepare_version_export	read-only	Create a five-minute download URL for an exact saved version.
get_version_share_link	read-only	Read the durable sharing state and current URL for one exact saved version.
create_version_share_link	write	Create or recover a durable public or code-protected link for one exact saved version.
revoke_version_share_link	destructive	Disable every existing public or code-protected URL for one exact saved version.

Artifacts

Tool	Effect	What it does
list_artifacts	read-only	List durable project artifacts and each latest immutable version without returning stored bytes.
get_artifact	read-only	Read one artifact's immutable version history, provenance, attribution, digests, and review states without…
prepare_artifact_upload	write	Create a five-minute PUT URL for exact bytes.
prepare_artifact_download	read-only	Create a five-minute download URL for one immutable artifact version.
list_artifact_approval_requests	read-only	List exact artifact versions awaiting or carrying a human review decision.
approve_artifact_version	write	Approve one pending immutable artifact version and make it current.
reject_artifact_version	write	Reject one pending immutable artifact version with useful feedback.

Agents

Tool	Effect	What it does
list_project_agents	read-only	List the authorizing account's built-in Agents assigned to one readable project.
get_agent	read-only	Read one Agent's saved Limits, trigger, immutable revision summaries, and redacted append-only run traces.
run_project_agent	write	Queue real execution for a project-linked Agent the person already test-ran and armed.
cancel_agent_run	destructive	Stop one queued or running Agent execution and close its durable run history.

Work assignments

Tool	Effect	What it does
list_work_assignments	read-only	List durable project work contracts, bounded attempts, leases, budgets, and result references.
create_work_assignment	write	Create a durable assignment with acceptance criteria and hard step, time, and spend ceilings.
claim_work_assignment	write	Atomically claim one open assignment and receive a five-minute renewable lease plus an immutable budget snapshot.
heartbeat_work_session	write	Renew the authenticated worker's active lease without widening its immutable step, time, or spend ceilings.
submit_work_session	write	End the authenticated worker's lease and submit typed result references for a separate review decision.
decide_work_assignment	write	Complete submitted work or reopen it with useful feedback.

---

Authentication

Access	Method	Scopes
create_public_preview	none — genuinely authless	—
Everything else	OAuth 2.1 authorization code + PKCE (S256)	least-privilege, consented per connection

The authorization server publishes RFC 8414 and RFC 9728 discovery documents at /.well-known/oauth-authorization-server and /.well-known/oauth-protected-resource, supports RFC 7591 dynamic client registration, rotates refresh tokens with family revocation on replay, and stores only token digests.

Review authority is deliberately separate from build authority. projects:build lets an agent produce candidate versions; approve_version and reject_version require projects:approve, which is excluded from default grants. An agent that can build cannot sign off on its own work unless a human explicitly delegates that.

---

Design notes

Source never travels through the model transcript. Exports and artifact downloads are five-minute signed URLs bound to the action, the account and the exact version. The MCP response carries metadata and links, not bytes — so a large site does not consume the context window, and generated source is not replayed into a provider log.

Shared previews run in an opaque-origin sandbox. A shared page is served under a CSP sandbox with connect-src 'none', so a site built by a model cannot read the viewer's cookies, storage, or reach the network.

External agent output enters review, not production. An artifact uploaded by an outside agent is stored immutably and marked pending; it cannot become the current version until a human with review authority approves that exact digest.

Interactive previews use MCP Apps. render_project_preview returns a real interactive card via ui://2ools/project-preview-v1.html rather than a wall of text.

---

Links

• Product — https://2ools.app
• Connection guide — https://2ools.app/connect-ai
• Free website audit (28 deterministic checks, no account) — https://2ools.app/website-audit
• Benchmarks: 36 real sites scored — https://2ools.app/website-audit/benchmarks
• What 2ools does — https://2ools.app/product
• Pricing — https://2ools.app/pricing
• LLM guide — https://2ools.app/llms.txt
• Privacy & terms — https://2ools.app/legal

Registry

Published to the official Model Context Protocol registry under the domain-verified namespace app.2ools/mcp.

curl "https://registry.modelcontextprotocol.io/v0/servers?search=2ools"

Status

The endpoint is live in production. Health, including the exact deployed release:

curl https://2ools.app/api/health

License

MIT — see LICENSE. This repository holds the server's public manifest and documentation; the 2ools application itself is a hosted product.