> SpecStream is the live web data engine for teams and agents. This is the Markdown copy of https://specstream.ai/docs/mcp. Every page, with what SpecStream does and costs: https://specstream.ai/llms.txt

# Connect an agent with MCP

SpecStream's MCP server gives an agent live web data as structured records, every value with its source URLs. It is a remote endpoint over Streamable HTTP with authenticated requests. A client that only runs local servers can connect through the [stdio bridge](#stdio).

[Setup guides for Claude, ChatGPT, Gemini, Grok, Cursor, and more](https://specstream.ai/docs/mcp/clients.md) give the connection steps for each client. Building on OpenAI's APIs? See [Use SpecStream from the Responses and Agents APIs](https://specstream.ai/docs/mcp/openai-api.md).

## Connection details

| Server URL | `https://api.specstream.ai/mcp` |
| --- | --- |
| Transport | Streamable HTTP |
| Protocol | MCP `2026-07-28`: stateless requests and `server/discover`. Clients that still start with `initialize` get `2025-11-25`. |
| Authentication | Sign in with OAuth, or your SpecStream API key as a Bearer token |

Add a remote MCP server in your client's settings. Set the URL above and configure its authentication header. The exact configuration format depends on your client.

text

```text
Authorization: Bearer YOUR_SPECSTREAM_API_KEY
```

## Apps that take a server URL

Lovable, Gumloop, Relevance AI, Lindy, Manus, Genspark, and most agent builders: add a remote MCP server, paste `https://api.specstream.ai/mcp`, and sign in when asked. If the app only takes headers, use the Bearer line above.

In Replit, one click adds SpecStream to Replit Agent, and you sign in when the Project Editor asks. No key goes in the link.

[Add to Replit](https://replit.com/integrations?mcp=eyJkaXNwbGF5TmFtZSI6IlNwZWNTdHJlYW0iLCJiYXNlVXJsIjoiaHR0cHM6Ly9hcGkuc3BlY3N0cmVhbS5haS9tY3AifQ==)

## Clients that only run local servers

Some clients can only start a server on your computer and talk to it over stdio. For those, SpecStream publishes `specstream-mcp-stdio` on PyPI. It forwards to the same endpoint, in whichever protocol version the client speaks, so your workspace and your agent share the account, permissions, and balance.

Run it with `uvx`, which fetches the package the first time. Add the entry to the client's MCP settings, and put your API key in the client's private environment or secret settings as `SPECSTREAM_API_KEY`, never in a prompt or a shared config file.

JSON

```json
{
  "mcpServers": {
    "specstream": {
      "command": "uvx",
      "args": ["specstream-mcp-stdio"],
      "env": { "SPECSTREAM_MCP_URL": "https://api.specstream.ai/mcp" }
    }
  }
}
```

Prefer to install it? Use Python 3.12 or newer, run `pip install specstream-mcp-stdio`, and set `"command": "specstream-mcp-stdio"` with an empty `args` list.

## OAuth clients

Clients that support MCP sign-in connect in one click: SpecStream accepts a client ID metadata document (an HTTPS client ID), and dynamic client registration for clients that haven't moved to metadata documents. You can also register an OAuth client with read, prepare, or execute access. Review its consent request before approval. Registered clients can be listed and revoked from account settings.

## Tools

There are tools for everything in the workspace: requests, results and their sources, lists and maps, notes, projects, destinations, schedules, developer access, and balance. [Explore the tool groups in the workspace](https://specstream.ai/workspace/agents).

Your client discovers the current tools with `tools/list`. Use `compile` to interpret an everyday-language request and `spec_options` to see the supported fields. Tools also cover saved revisions, project permissions, follow-up work, schedules, sharing, exports, and billing corrections.

New work uses `quote_capacity_work`, then `authorize_capacity_work` with the quote ID and hash, the approved maximum in `confirm_max_cents`, and a stable request key. `list_find_rows` returns the results, each value with its source URL and read time.

## Prompts

The server offers prompts for the jobs people start most. A client lists them as slash commands or a prompt picker. A prompt only writes the opening message: the agent still reads the request back, gets a quote, and waits for your approval before anything runs.

-   **Find live web data** (`find_data`) takes what you need in plain words and an optional budget.
-   **Add facts to my records** (`enrich_records`) takes your records, pasted or linked from a Google Sheet, Excel workbook, or Airtable view, and the facts to add.
-   **Keep data current** (`keep_current`) takes what to watch, how often, and where new and changed records should go.
-   **See what changed** (`see_changes`) takes a list's name and shows each changed fact, before and after, with the page where the new value was seen.

## Events

A client can subscribe a callback and hear when a find finishes, a monitor finds something new, or records are added to a list, through `events/subscribe`. [MCP Events](https://specstream.ai/docs/webhooks.md#mcp-events) covers the subscription, the signature, and the retries. ChatGPT uses them.

## Give each agent only what it needs.

An agent uses the same identities, permissions, and balance as the workspace. Give it only the project access, operations, and spending limits the task needs.

An API key has a total spending limit. Delegated access can restrict the agent to specific operations or an exact reviewed request, with a per-run limit, monthly limit, and expiry.

Interpreting or saving a request doesn't authorize research. New work still needs a quote and an approved maximum, either reviewed by you or covered by an explicit delegation.

## Pick up the results yourself.

Open the same saved work in the workspace to read sources, map locations, compare results, and organize projects.

Agents that read before they act can start with [llms.txt](https://specstream.ai/llms.txt): how to connect, how a request works, and the money rules, in plain text. An agent with no account can [pay per request](https://specstream.ai/docs/agent-payments.md) instead.

## Where it is listed

SpecStream is in the [official MCP registry](https://registry.modelcontextprotocol.io/v0.1/servers?search=ai.specstream) as `ai.specstream/specstream`, and on [Smithery](https://smithery.ai/servers/specstream/specstream), [LobeHub](https://lobehub.com/mcp/specstream), [ModelScope](https://www.modelscope.ai/mcp/servers/SpecStream/SpecStream), and [MCPBundles](https://www.mcpbundles.com/skills/apispecstreamaimcp).

[Prefer direct HTTP requests?](https://specstream.ai/docs/api.md) · [Understand quotes and billing](https://specstream.ai/docs/billing.md)