Web Research To Docs

Competitor watch, cited research reports, paper alerts and market maps, filed where you read.

An MCP server with 10 workflows across Firecrawl, Tavily, Notion, Slack, Google Docs, Google Sheets, GitHub and Linear. Each workflow is a prompt your agent runs as a slash command, over the 19 tools it needs and no others.

uv tool install https://github.com/r28ai/web-research-to-docs-mcp/releases/download/v0.1.0/web_research_to_docs_mcp-0.1.0-py3-none-any.whl
claude mcp add research -- web-research-to-docs-mcp

It installs with uv from this repository's release, with no git and nothing to build; nothing but Charter and the libraries it uses comes from PyPI. To update, run the install line from the latest release. If a desktop app cannot find web-research-to-docs-mcp, give it the full path from which web-research-to-docs-mcp (where web-research-to-docs-mcp on Windows).

Then ask your agent to connect your apps, or run /mcp__research__setup.

Connect your apps

Ask the agent to connect one ("connect Linear"). It tells you where to get that app's key and the command that stores it, and the next call works, with no restart. The agent never asks for a key in the chat.

Or connect everything this server uses from a terminal:

web-research-to-docs-mcp login            # each app in turn
web-research-to-docs-mcp login firecrawl  # just one
web-research-to-docs-mcp status           # what is connected

Tokens and keys go to your operating system's keychain (macOS Keychain, Windows Credential Manager, the Secret Service on Linux), and are checked with one read-only call to the app's own API before they are kept. Every key, token and OAuth client is yours: we register no app with any of these services, and nothing passes through a server of ours, because there isn't one.

A variable set in your client's config always wins over the keychain.

Workflows

Every prompt takes one optional argument, details: the repo, team, channel, customer or date range you mean, so the agent does not have to ask. In Claude Code, put it in quotes, or only its first word arrives:

/mcp__research__competitive_intel_digest "competitors acme.com and globex.com"

Reads run without asking. Before anything that creates, sends, changes or deletes, the prompt tells the agent to show you the call and wait.

6 of the 10 workflows need no Google or Granola credential.

Other clients

Claude Desktop: install uv if you have not, since Claude Desktop starts the server with it, then open the .mcpb from the latest release. Claude asks for any keys in its own settings and keeps them in your keychain. The first start takes a few seconds longer, while uv installs it.

VS Code (.vscode/mcp.json): VS Code asks for each key the first time the server starts and stores it securely. Leave out any you stored with login.

{
  "inputs": [
    {
      "type": "promptString",
      "id": "firecrawl-api-key",
      "description": "Firecrawl: API key",
      "password": true
    },
    {
      "type": "promptString",
      "id": "tavily-api-key",
      "description": "Tavily: API key",
      "password": true
    },
    {
      "type": "promptString",
      "id": "notion-api-key",
      "description": "Notion: Integration secret (ntn_\u2026)",
      "password": true
    },
    {
      "type": "promptString",
      "id": "slack-bot-token",
      "description": "Slack: Bot token (xoxb-\u2026)",
      "password": true
    },
    {
      "type": "promptString",
      "id": "google-client-secret",
      "description": "Google: OAuth client secret",
      "password": true
    },
    {
      "type": "promptString",
      "id": "github-token",
      "description": "GitHub: Personal access token",
      "password": true
    },
    {
      "type": "promptString",
      "id": "linear-api-key",
      "description": "Linear: Personal API key",
      "password": true
    }
  ],
  "servers": {
    "research": {
      "type": "stdio",
      "command": "web-research-to-docs-mcp",
      "env": {
        "FIRECRAWL_API_KEY": "${input:firecrawl-api-key}",
        "TAVILY_API_KEY": "${input:tavily-api-key}",
        "NOTION_API_KEY": "${input:notion-api-key}",
        "SLACK_BOT_TOKEN": "${input:slack-bot-token}",
        "GOOGLE_CLIENT_SECRET": "${input:google-client-secret}",
        "GITHUB_TOKEN": "${input:github-token}",
        "LINEAR_API_KEY": "${input:linear-api-key}",
        "GOOGLE_CLIENT_ID": ""
      }
    }
  }
}

Cursor (.cursor/mcp.json) starts it the same way:

{
  "mcpServers": {
    "research": {
      "command": "web-research-to-docs-mcp"
    }
  }
}

Codex (~/.codex/config.toml) starts a turn without waiting for a server unless it is required, and then the agent has none of its tools. required = true makes the session wait for it, and startup_readiness = "catalog" waits for its tool list rather than just its connection:

[mcp_servers.research]
command = "web-research-to-docs-mcp"
required = true
startup_readiness = "catalog"
startup_timeout_sec = 30

Name the server research. A host builds each tool's name from that key, and a longer one can push a tool past the 64 characters a function name allows.

Built with Charter

Every tool here is a Charter declaration: a Pydantic schema saying where each field goes on the wire. Charter's runtime builds the request, attaches and refreshes the credential, and trims the response before the model reads it. It runs in your process, with no proxy and no telemetry.

The 19 tool schemas come to 38,591 tokens.

The same tools work in your own agent, without MCP:

from charter.adapters.openai import to_openai_tools
from charter_packs_mcp import FAMILIES

tools = FAMILIES["research"].tools()
definitions = to_openai_tools(tools)   # or charter.adapters.langchain

Need an API that isn't here? Write a pack: your coding agent writes the declarations, and Charter's conformance suite checks them.

License

Apache 2.0.