Notion MCP
Notion’s first-party hosted MCP connects AI clients to workspace knowledge through semantic search, Markdown-native retrieval, OAuth and write actions. It is strongest as a knowledge interface and less dependable as a replacement for deterministic database querying.
mcp.notion.com/mcpStart with Decision, Score, Tools, Safety, Token Cost and Fit. Setup answers connection intent; Evidence and External Sources hold the deeper verification trail.
Should you use Notion MCP?
Use Notion MCP when an AI agent repeatedly needs current Notion knowledge, page context, research notes or supervised content edits. Verify database completeness and write behavior before using it for unattended, property-filter-heavy automation.
Knowledge retrieval, project documentation, research synthesis, meeting context, page creation and human-reviewed updates.
The task requires complete enumeration of large databases, exact property filters, guest-only access, or unattended broad workspace writes.
Semantic workspace search plus Markdown-native retrieval gives agents relevant Notion context without the legacy server’s giant API-derived payloads.
The connection inherits the authenticated user’s Notion permissions, while semantic relevance does not guarantee exhaustive structured-data retrieval.
Use Notion MCP like a knowledge engine, not automatically like a database API. “Find what we decided” is its natural job. “Return every record matching this exact filter with no omissions” requires separate verification.
Why does Notion MCP score 77/100?
The 77/100 score reflects strong semantic retrieval, permission fidelity, maintainability and context efficiency, offset by weaker database completeness, uneven OAuth/session reliability and a broad permission boundary. The score evaluates the hosted Remote Notion MCP, not the older local package.
| Universal parameter | Score | Finding |
|---|---|---|
| Effectiveness | 87 | Strong at finding and working with human knowledge stored in Notion. |
| Reliability | 68 | Recent OAuth, client, Cloudflare and schema regressions reduce unattended confidence. |
| Safety | 80 | Notion permissions are respected, but write access and broad workspace visibility remain consequential. |
| Efficiency | 84 | Remote Markdown/relevance-focused results are materially leaner than the old local API-derived implementation. |
| Compatibility | 72 | Major clients connect, but cross-provider schema and OAuth issues have appeared in 2026. |
| Maintainability | 94 | First-party hosted service and the path Notion currently prioritizes. |
| Setup friction | 89 | Remote URL + OAuth is simpler than token-based local deployment when the auth flow works normally. |
Notion-specific parameters
| Parameter | Score | Why it matters |
|---|---|---|
| Semantic retrieval relevance | 88 | Hosted semantic search is a major advantage for fuzzy organizational-knowledge questions. |
| Permission fidelity | 93 | MCP follows the authenticated user’s existing Notion permissions. |
| Database completeness | 61 | Semantic search and query gaps can miss rows required for deterministic workflows. |
| Write precision | 75 | Page creation/editing is useful, but mutation and schema regressions reduce confidence. |
| Workspace-scope minimization | 65 | The user permission boundary is accurate but can be much broader than the task requires. |
| Prompt-injection containment | 67 | Workspace text remains model input and must be treated as potentially untrusted. |
| OAuth/session reliability | 62 | Multiple 2026 reports document callback, state and client interoperability failures. |
| Context efficiency | 86 | Relevant Markdown context is substantially lighter than giant API object trees. |
What is Notion MCP?
Notion MCP is Notion’s first-party hosted Model Context Protocol server for connecting compatible AI clients to a Notion workspace. It lets agents search workspace knowledge, fetch pages, work with databases and comments, and create or edit content using the authenticated user’s existing Notion permissions.
Notion acts as the MCP server. Claude, Cursor, ChatGPT, OpenCode or another compatible application acts as the MCP client. The server exposes tools that the client can invoke when the model requires current workspace information or needs to perform a permitted action.
Which Notion MCP server should you use: remote or local?
Use the hosted Remote Notion MCP for the current default path. Notion’s official repository says the older self-hosted @notionhq/notion-mcp-server implementation is no longer actively maintained or supported and directs users to the remote service.
| Attribute | Remote Notion MCP | Legacy local server |
|---|---|---|
| Current Notion recommendation | Yes | No |
| Canonical endpoint | https://mcp.notion.com/mcp | Local package/runtime |
| Hosting | Notion-hosted | Self-hosted |
| Authentication | OAuth | Notion integration/API token |
| Semantic workspace search | Yes | Basic API-derived behavior |
| Connected-app search | Yes | No equivalent |
| Markdown-native read/edit | Yes | More API-object oriented |
| Permission model | User’s existing permissions | Integration permissions / shared pages |
| Token efficiency | Agent-optimized | Reported ~21K schema footprint in one 2026 issue |
| Maintenance status | Active hosted path | No longer actively supported |
What can Notion MCP do?
Notion MCP can semantically search workspace knowledge, fetch pages, create and update content, move pages, work with databases and views, read or create comments, and retrieve users or teams. The exact remote tool list changes, so MCPVerdict does not hardcode a permanent count.
| Capability | Example hosted tool | Authority | Role |
|---|---|---|---|
| Semantic search | notion-search | Read | Find relevant workspace knowledge by meaning/text. |
| Page retrieval | notion-fetch | Read | Fetch page/entity content for analysis. |
| Create pages | notion-create-pages | Write | Create new Notion pages. |
| Update pages | notion-update-page | Write | Edit existing page content/properties. |
| Move pages | notion-move-pages | Write | Move content within the workspace. |
| Create databases | notion-create-database | Write | Create structured databases. |
| Database/view queries | notion-query-database-view | Read | Query supported database/view structures. |
| Comments | notion-get-comments / notion-create-comment | Read / Write | Read and add comments. |
| Users and teams | notion-get-users / notion-get-teams | Read | Retrieve supported workspace identity data. |
Semantic retrieval is optimized to find relevant information. Exact automation requires proof that every matching row was enumerated, filtered and paginated correctly.
Which AI clients work with Notion MCP?
Notion explicitly names Claude, Cursor and ChatGPT as popular MCP clients. Other compatible remote-MCP clients can connect when they support Notion’s endpoint, transport and OAuth requirements.
| AI client | Support level | Connection note |
|---|---|---|
| Claude.ai / Claude Desktop | Officially named | Connect through supported Notion/Claude integration and OAuth. |
| Claude Code | Supported remote MCP | Use the hosted HTTP endpoint, then authorize through /mcp. |
| Cursor | Officially named | Add Notion MCP in MCP settings and complete OAuth. |
| ChatGPT | Officially named by Notion | Notion’s help page currently lists ChatGPT Pro among popular AI apps. |
| OpenCode | Protocol-compatible path | Remote MCP + OAuth support; verify current client behavior. |
| VS Code / other clients | Conditional | Works when the client supports Streamable HTTP and the required OAuth flow. |
What is Notion MCP best used for?
Notion MCP is strongest when the required outcome is relevant organizational knowledge, not guaranteed database exhaustiveness. Search, summarization, documentation and supervised updates align directly with the hosted server’s design.
| Task | Fit | Reason |
|---|---|---|
| Find what the team decided about authentication | Excellent | Semantic retrieval matches fuzzy knowledge questions. |
| Summarize project documents | Excellent | Relevant pages can be fetched as agent-friendly content. |
| Create technical documentation | Good | Write tools support page creation and updates. |
| Update project knowledge with human review | Good | Useful write surface with a clear review checkpoint. |
| Search connected organizational knowledge | Strong | Hosted MCP can use Notion’s connected-app search path. |
| Return every row where Status = Blocked | Verify | Property filtering and exhaustive retrieval require testing. |
| Enumerate hundreds of rows without omissions | Verify | Recent evidence shows structured-query gaps. |
| Run unattended exact database automation | Poorer fit | Completeness and mutation reliability become primary requirements. |
Is Notion MCP free?
Notion does not document a separate per-MCP-call fee for the hosted endpoint. Real cost comes from the Notion plan, the AI client or API, model-token usage, and operational work such as retries, review and database-validation calls.
| Cost layer | What it means |
|---|---|
| Notion MCP endpoint | No separate per-call MCP fee documented in the evaluated material. |
| Notion workspace | Normal Notion plan and workspace requirements still apply. |
| AI client | Claude, ChatGPT, Cursor or another client can have its own subscription/usage terms. |
| Model API | Input and output tokens are billed when using API-priced models. |
| MCP schema/context | Tool definitions and retrieved Notion content consume model context. |
| Operational cost | Retries, human review and database completeness checks add time and tokens. |
Does Notion MCP require an API token?
The current hosted Notion MCP uses OAuth, not a manually pasted Notion integration token. Older setup guides refer to NOTION_TOKEN because the legacy self-hosted package uses token-based integration credentials.
OAuth improves setup and permission fidelity because the hosted service acts as the authenticated Notion user. That same design expands the security boundary when the user has broad workspace access.
Is Notion MCP safe?
Notion MCP has strong permission fidelity, but its effective access boundary can be broad because tools act with the authenticated user’s Notion permissions. A high-privilege account can expose sensitive workspace knowledge and authorize meaningful edits.
| Security area | Finding |
|---|---|
| Authentication | OAuth for hosted Notion MCP. |
| Permission fidelity | Strong; existing Notion permissions remain enforced. |
| Permission scope | Potentially broad because the connection can access what the user can access. |
| Read exposure | Retrieved workspace information becomes AI-client/model context. |
| Write exposure | Pages, comments and structured workspace objects can be modified. |
| Prompt injection | Workspace/imported content can contain untrusted instructions. |
| Enterprise governance | Admins can approve/block AI apps and manage allowed connections. |
| Telemetry visibility | Hosted implementation is proprietary; exact MCP-specific server telemetry is not independently source-auditable. |
Grade C reflects authenticated access to private workspace data plus meaningful write authority. It is a capability-risk classification, not a claim that the official Notion server is malicious.
What Notion data can the MCP access?
The MCP can access content available to the authenticated Notion user. Permission inheritance prevents the server from bypassing Notion access controls, but a user with executive documents, internal strategy, HR material, private databases or client pages creates a correspondingly broad retrieval surface.
Can Notion content create prompt-injection risk?
Yes. Retrieved workspace text becomes model context, so malicious or misleading instructions embedded in customer feedback, imported research, public submissions or copied external material can influence an agent. Human confirmation remains important for consequential writes.
How reliable is Notion MCP?
Notion MCP is practical for interactive knowledge work, but 2026 reports document OAuth failures, blocked post-authentication calls, database-query gaps and a remote tool-schema regression that affected both Anthropic and OpenAI clients.
| Evidence | Practical effect | Status/context |
|---|---|---|
| Cloudflare blocked Cursor tool calls after auth | Tool requests returned 403 despite successful authentication | April 2026 · Issue #252 |
| Hosted database query path missing | Roughly 500-row workflow could not enumerate/filter reliably | April 2026 · Issue #256 |
| OAuth state persisted across reconnects | Connection became unusable across browsers/devices | April 2026 · Issue #268 |
| Claude.ai callback returned server error | OAuth approval completed but connection failed | April 2026 · Issue #269 |
| Claude Code remained unauthenticated | OAuth flow did not produce a working MCP connection | April 2026 · Issue #277 |
| Property filtering absent from semantic search | Exact database retrieval could systematically miss rows | April 2026 · Issue #278 |
| Top-level schema combinators rejected | Anthropic and OpenAI could reject the whole tool set | August 2026 · Issue #340 · later closed |
| Safari authorization window blank | OAuth flow could fail in a specific macOS/Safari state | August 2026 · Issue #345 |
Individual GitHub reports do not prove universal failure. They do show enough cross-client friction that an unattended production workflow deserves connection, query-completeness and retry testing.
How many tokens does Notion MCP use?
MCPVerdict estimates the hosted Notion MCP tool-schema overhead at roughly 4K–8K tokens per session. No trustworthy current raw tools/list capture was available for the hosted service, so this range is explicitly an estimate rather than an official Notion token count.
The legacy local implementation had a 2026 report of roughly 21K tool-definition tokens per session. MCPVerdict does not reuse that number for the hosted server because the remote implementation is architecturally different and is explicitly designed to return smaller, relevance-focused Markdown context.
What consumes context?
| Context component | Estimated / observed size | Meaning |
|---|---|---|
| Hosted tool schemas | ~4K–8K tokens estimated | Fixed-ish MCP tool-definition overhead; exact tokenizer/client behavior varies. |
| Legacy local schemas | ~21K reported | Older API-derived implementation; not a hosted-server measurement. |
| Semantic search results | Variable | Relevant retrieval can reduce unnecessary context. |
| Page fetches | Variable | Long Notion documents can still add thousands of useful tokens. |
| Realistic MCPVerdict evaluation | ~45K input + 10K output | Estimated 10-task knowledge/database test, not a Notion benchmark. |
How much of each model’s context window does a 4K–8K schema use?
The percentages below isolate the estimated MCP tool-schema load. They exclude the conversation, retrieved Notion pages, tool results, project context, reasoning tokens and model output.
| Current model | Documented context | 4K–8K Notion schema share | Input / output price per 1M | 4K–8K schema input cost | 45K in + 10K out evaluation |
|---|---|---|---|---|---|
| GPT-6 Astra OpenAI | 1.05M | 0.38%–0.76% | $10 / $50 | $0.04–$0.08 | $0.95 standard short-context |
| GPT-6 Sol OpenAI | 1.05M | 0.38%–0.76% | $2 / $10 | $0.008–$0.016 | $0.19 standard short-context |
| GPT-6 Luna OpenAI | 1.05M | 0.38%–0.76% | $0.1 / $0.5 | $0.0004–$0.0008 | $0.0095 standard short-context |
| GPT-5.6 Sol OpenAI | 1.05M | 0.38%–0.76% | $4 / $20 | $0.016–$0.032 | $0.38 promotional standard |
| Claude Opus 5.5 Anthropic | 1M | 0.40%–0.80% | $4 / $20 | $0.016–$0.032 | $0.38 standard API |
| Claude Fable 5.1 Anthropic | 1M | 0.40%–0.80% | $10 / $50 | $0.04–$0.08 | $0.95 standard API |
| Claude Sonnet 5 Anthropic | 1M | 0.40%–0.80% | $2 / $10 | $0.008–$0.016 | $0.19 standard API |
| Claude Haiku 4.5 Anthropic | 200K | 2.00%–4.00% | $1 / $5 | $0.004–$0.008 | $0.095 standard API |
| Gemini 3.7 Flash | 1,048,576 | 0.38%–0.76% | $0.75 / $3.75 | $0.003–$0.006 | $0.071 intro pricing through Dec 31, 2026 |
| Gemini 3.1 Pro Preview | 1M | 0.40%–0.80% | $2 / $12 | $0.008–$0.016 | $0.21 ≤200K prompt tier |
| Gemini 3.1 Flash-Lite | 1M | 0.40%–0.80% | $0.25 / $1.5 | $0.001–$0.002 | $0.026 standard paid tier |
| Grok 4.7 xAI | 500K | 0.80%–1.60% | $2 / $6 | $0.008–$0.016 | $0.15 short-context |
| DeepSeek V4.1 Flash DeepSeek | 1,048,576 | 0.38%–0.76% | $0.15/$0.60 off-peak · $0.30/$1.20 peak | $0.0006–$0.0012 off-peak | $0.013 off-peak · $0.026 peak off-peak; peak is 2× |
| Mistral Medium 3.5 Mistral | 256K | 1.56%–3.12% | $1.5 / $7.5 | $0.006–$0.012 | $0.14 standard API |
| Mistral Small 4 Mistral | 256K | 1.56%–3.12% | $0.15 / $0.6 | $0.0006–$0.0012 | $0.013 standard API |
4K–8K is an estimated hosted schema range. Exact token counts depend on the MCP client, current tool inventory, schema serialization and the target model’s tokenizer. Capture the live tools/list payload for an exact setup-specific measurement.
How much does a realistic Notion MCP evaluation cost on current AI models?
A 45K-input + 10K-output evaluation costs about $0.0095 on GPT-6 Luna, $0.19 on GPT-6 Sol or Claude Sonnet 5, $0.38 on GPT-5.6 Sol or Claude Opus 5.5, and $0.95 on GPT-6 Astra or Claude Fable 5.1 at the listed standard rates.
The calculation is (45,000 × input rate / 1,000,000) + (10,000 × output rate / 1,000,000). It measures API inference, not consumer subscription pricing. Caching, batch/flex tiers, long-context surcharges, regional processing and tool-specific fees can change actual billing.
| Provider example | 45K input + 10K output | Interpretation |
|---|---|---|
| GPT-6 Luna | $0.0095 | Low per-token cost for the reference workload. |
| Mistral Small 4 | $0.0128 | Low-cost open model API tier. |
| DeepSeek V4.1 Flash | $0.0128 off-peak · $0.0255 peak | Time-dependent official pricing. |
| Gemini 3.1 Flash-Lite | $0.0263 | Low-cost Gemini paid tier. |
| Grok 4.7 | $0.15 | Short-context rate; long prompts can cost more. |
| GPT-6 Sol / Claude Sonnet 5 | $0.19 | Balanced frontier-agent cost for this workload. |
| GPT-5.6 Sol / Claude Opus 5.5 | $0.38 | Higher standard token price. |
| GPT-6 Astra / Claude Fable 5.1 | $0.95 | Premium frontier pricing for this workload. |
Pricing snapshot: September 27, 2026. The detailed table above includes 15 current models and uses current official provider pricing where available.
How do you connect Notion MCP?
Add https://mcp.notion.com/mcp as a remote MCP server in a compatible client and complete Notion OAuth. The hosted path does not require installing the old NPM server or pasting a Notion integration token.
How do you add Notion MCP to Claude Code?
Register the remote HTTP endpoint, then authorize it from Claude Code’s MCP interface:
claude mcp add --transport http notion https://mcp.notion.com/mcpOpen Claude Code, run /mcp, select Notion, and complete OAuth. Verify one read operation before enabling a workflow that performs edits.
How do you add Notion MCP to Cursor?
Add the official remote endpoint in Cursor’s MCP settings, then complete the browser OAuth flow:
{
"mcpServers": {
"notion": {
"url": "https://mcp.notion.com/mcp"
}
}
}How do you connect ChatGPT to Notion?
Notion currently lists ChatGPT Pro among popular AI applications for Notion MCP. Connect Notion through the supported ChatGPT connector/MCP interface, authenticate with Notion, and verify the resulting permission scope before using sensitive workspace content.
How do you connect Notion MCP to OpenCode?
Use OpenCode’s remote-MCP configuration and point it to the official Notion endpoint. OpenCode supports OAuth-capable remote MCP servers; verify the current client instructions because its CLI/config surface can change independently of Notion.
Authenticate → confirm tools appear → search a known page → fetch it → test a controlled write → test database completeness separately. A green connection indicator is not enough for production trust.
Is Notion MCP better than the Notion API?
Notion MCP fits agentic knowledge retrieval and conversational content work; the Notion API fits deterministic application logic that requires explicit filters, pagination and controlled request behavior. The interfaces solve different jobs.
| Requirement | Notion MCP | Direct Notion API |
|---|---|---|
| AI-agent integration | Native fit | Requires an application/tool layer |
| Semantic workspace retrieval | Strong | Requires custom retrieval logic |
| Natural-language knowledge search | Strong | Must be implemented |
| User OAuth permission inheritance | Strong | Depends on integration architecture |
| Markdown-oriented agent context | Strong | Developer controls transformation |
| Exact property filtering | Verify supported path | Explicit programmatic control |
| Pagination control | Less direct | Explicit |
| Exhaustive row processing | Requires validation | Easier to enforce |
| Deterministic application workflow | Conditional | Strong fit |
Who is Notion MCP a good fit for?
Knowledge-first workflows
- Search project documentation and decisions.
- Retrieve context while coding or researching.
- Summarize meeting notes and research collections.
- Create technical docs, specs and project pages.
- Update workspace knowledge with human review.
- Search content exposed through Notion’s connected-app path.
Automation-first workflows
- Exact large-database enumeration.
- Property-filter-heavy automation without completeness testing.
- Broad unattended writes from a high-privilege account.
- Workflows that cannot tolerate OAuth/client interruptions.
- Use cases where incomplete search results could be mistaken for complete data.
- Deployments requiring verified MCP 2026-07-28 compliance.
For knowledge held in conversations rather than workspace pages, compare Slack MCP.
What should you test before production use?
Test Notion MCP with a normal low-risk member account and non-sensitive content before authorizing a high-privilege production account. Separate knowledge retrieval from database completeness because success on one does not prove success on the other.
| Test | Pass condition |
|---|---|
| Known decision search | The correct historical decision is retrieved with supporting context. |
| Requirements search | Relevant project requirements are found consistently. |
| Multi-page summary | Important source information is not silently omitted. |
| Exact property query | Every known matching row is returned. |
| Large database | More than 100 rows can be processed completely. |
| Duplicate detection | Existing duplicates are identified reliably. |
| Controlled update | Only the specified records/pages change. |
| Permission boundary | An inaccessible page remains inaccessible. |
| Reconnect behavior | OAuth refresh/reconnect recovers without corrupting state. |
MCPVerdict’s reference test budget is approximately 40–55 minutes and 45K input + 10K output tokens. These figures are evaluation estimates, not official Notion benchmarks.
What are the current Notion MCP technical details?
Show the technical profile
| Attribute | Evaluated value |
|---|---|
| Publisher | Notion |
| Type | First-party hosted remote MCP |
| Canonical endpoint | https://mcp.notion.com/mcp |
| Version | Hosted service; no public semantic version |
| Evaluation date | September 27, 2026 |
| Authentication | OAuth |
| Permission model | Authenticated user’s existing Notion permissions |
| Hosted implementation open source | No |
| Legacy local implementation | MIT; no longer actively maintained/supported |
| Write capability | Yes |
| Estimated hosted schema overhead | ~4K–8K tokens |
| Reported legacy schema footprint | ~21K tokens in a 2026 issue |
| Scan grade | C |
| Overall score | 77/100 |
| Confidence | Moderate |
| MCP 2026-07-28 | Unknown / not publicly verified |
Notion MCP evidence and sources
These direct sources support the findings below. Documentation describes supported behavior; issue reports describe individual observations. Neither establishes a universal success rate.
- Hosted setup and permissions: Notion documents MCP connections and notes that tools act with the connected user’s Notion permissions. Enterprise admins can restrict approved clients. Read source ↗
- Hosted versus local support: The official local repository directs users toward the remote server and says active support is prioritized there. Read source ↗
- Reported client failure: An April 2026 report describes successful authentication followed by Cloudflare 403 responses in Cursor. This is a user report about a specific environment, not proof of a universal current failure. Read source ↗
Historical incident summaries without a verified original source are omitted from this evidence list. Scores and token estimates elsewhere in this profile remain the supplied editorial assessment, not independently reproduced benchmarks.
Which external sources support this Notion MCP evaluation?
The links below separate first-party product documentation, operational GitHub evidence and current AI-model pricing sources. External links open in a new tab so the evidence can be checked without losing the profile.
Official description, supported AI clients, permissions, governance and use cases.
Official developer documentation for the hosted MCP endpoint and client setup.
Official hosted-tool documentation; useful because the tool surface changes over time.
Official legacy/self-hosted repo; now tells users to prefer Remote Notion MCP.
Official security guidance for untrusted content, agent actions and external connections.
April 2026 Cursor report: authentication succeeded, then tool calls returned Cloudflare 403.
April 2026 report involving a roughly 500-row database and missing structured query capability.
April 2026 report of reconnect state persisting across browsers/devices.
April 2026 hosted MCP callback failure after OAuth approval.
April 2026 Claude Code connection remaining in an authentication-required state.
April 2026 evidence that semantic search did not provide reliable property-based database filtering.
August 2026 remote schema regression affecting Anthropic and OpenAI clients; issue later closed.
August 2026 Safari authorization-flow report.
Current GPT-6 and GPT-5.6 API token prices used in the model-cost table.
Current Claude Sonnet 5 pricing.
Current Claude Opus 5.5 pricing.
Current Claude Fable 5.1 pricing and cache-read update.
Claude Haiku 4.5 pricing and availability.
Gemini API prices used in the model-cost table.
Grok 4.7 context window and short-context token pricing.
DeepSeek V4.1 Flash 1M context and peak/off-peak token prices.
Mistral Medium 3.5 and Small 4 token pricing.
Writing framework used for direct answers, question-led H2s, contextual flow and semantic precision.
What are the most common Notion MCP questions?
Does Notion support MCP?
Yes. Notion operates an official hosted MCP server at https://mcp.notion.com/mcp and documents connections for AI clients including Claude, Cursor and ChatGPT.
Is Notion MCP free?
Notion does not document a separate per-call MCP fee. Normal Notion plan requirements, AI-client costs and model-token charges still apply.
What is the Notion MCP URL?
The canonical hosted endpoint is https://mcp.notion.com/mcp.
Does Notion MCP need a token?
The current hosted server uses OAuth. The older local package uses Notion integration credentials, which is why legacy instructions reference NOTION_TOKEN.
Can Claude Code use Notion MCP?
Yes. Add the remote HTTP endpoint with claude mcp add --transport http notion https://mcp.notion.com/mcp, then authorize through the MCP interface.
Can Cursor use Notion MCP?
Yes. Cursor is explicitly named by Notion as a popular MCP client. Add the hosted endpoint and complete OAuth.
Can ChatGPT connect to Notion?
Yes. Notion currently lists ChatGPT Pro among popular AI apps for Notion MCP. Availability and UI details can vary with the ChatGPT product surface.
What are the main Notion MCP tools?
The hosted surface includes semantic search, fetch/read, page creation/update/movement, database operations, comments, users and teams. The exact remote list changes over time.
Is Notion MCP better than the Notion API?
Not categorically. MCP is stronger for agentic knowledge retrieval; the direct API is stronger when application logic requires explicit filters, pagination and deterministic processing.
How many tokens does Notion MCP use?
MCPVerdict estimates the hosted schema at roughly 4K–8K tokens. The exact count requires capturing the current tool schema and counting it with the target model tokenizer.