form4api-mcp

Production-grade SEC Form 4 insider trading data for any MCP-compatible AI assistant — amendment-aware, 10b5-1 clean, with Form 144 + institutional 13F-HR overlay, plus congressional STOCK Act trades and insider/Congress convergence — 38 tools + 6 ready-made research prompts

npm version

An MCP server that exposes the hosted Form4API REST API to Claude Desktop, Cursor, Windsurf, VS Code, Codex CLI, and any other MCP-compatible client. Configured once, your LLM can answer questions about insider trading, institutional positioning, and intent-to-sell filings directly during research sessions.

Four data-quality claims no scraping-based alternative can make:

• 🛡 Amendment-aware — Form 4/A amendments are reconciled automatically. No double-counting when an insider corrects a filing.
• 🎯 10b5-1 clean — every transaction flagged as pre-scheduled (10b5-1 plan) or discretionary. Cluster signals exclude planned trades by construction.
• 📜 Form 144 intent-to-sell — 118K+ Form 144 filings indexed. Catch insider sales ~2 days before they hit Form 4.
• 🏛 Institutional × insider join — every transaction carries the current 13F-HR ownership context (top-3 holders, AUM trend) in the same response: no second call, no client-side join. Among the self-serve SEC data APIs we've surveyed, none return both sides in one query — sec-api.io and Kaleidoscope both ship 13F and insider data as separate endpoints.

Quick install

1. Get a free API key

Go to www.form4api.com → Sign in → Dashboard. Free plan includes 500 requests/day, no credit card required.

2. Add to your MCP client

Claude Desktop — edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "form4api": {
      "command": "npx",
      "args": ["-y", "form4api-mcp"],
      "env": {
        "FORM4API_KEY": "YOUR_API_KEY"
      }
    }
  }
}

Restart the client. The tools appear automatically.

Claude Code (CLI):

claude mcp add form4api -- npx -y form4api-mcp

…then set FORM4API_KEY in your shell or in ~/.claude/mcp.json.

Cursor — edit ~/.cursor/mcp.json (user-level) or .cursor/mcp.json (workspace-level):

{
  "mcpServers": {
    "form4api": {
      "command": "npx",
      "args": ["-y", "form4api-mcp"],
      "env": {
        "FORM4API_KEY": "YOUR_API_KEY"
      }
    }
  }
}

Restart Cursor. The tools appear automatically.

Windsurf — edit ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "form4api": {
      "command": "npx",
      "args": ["-y", "form4api-mcp"],
      "env": {
        "FORM4API_KEY": "YOUR_API_KEY"
      }
    }
  }
}

Restart Windsurf. The tools appear automatically.

VS Code — edit .vscode/mcp.json (workspace-level). Note: VS Code uses the servers key (not mcpServers):

{
  "servers": {
    "form4api": {
      "command": "npx",
      "args": ["-y", "form4api-mcp"],
      "env": {
        "FORM4API_KEY": "YOUR_API_KEY"
      }
    }
  }
}

Restart VS Code. The tools appear automatically.

Codex CLI — config is TOML at ~/.codex/config.toml:

[mcp_servers.form4api]
command = "npx"
args = ["-y", "form4api-mcp"]
env = { FORM4API_KEY = "YOUR_API_KEY" }

Verify it works

Ask your LLM to run the verify_setup tool — it confirms your API key is valid and the MCP server is reachable, or returns the exact fix steps.

Example: "Run the verify_setup tool to confirm the MCP is configured correctly."

Try before you commit a key

get_public_stats is a keyless tool — it works with no FORM4API_KEY set. Try it first to preview live data coverage before signing up:

FORM4API_KEY="" npx form4api-mcp

Once you like what you see, sign up for a free key at www.form4api.com → set FORM4API_KEY → all tools unlock.

3. Or run directly

FORM4API_KEY=YOUR_API_KEY npx form4api-mcp

Available tools (38)

Search

Form 4 insider trading

Signals + sentiment

Form 144 + institutional

Congress + convergence

Utility

Prompts (6)

Beyond the 38 tools, this MCP ships 6 prompts — ready-made research recipes that a client can list (prompts/list) and load (prompts/get) so you don't have to hand-assemble the right tool sequence yourself. Each one tells the LLM exactly which SEC Form 4 / Form 144 / 13F-HR tools to call, in what order, and how to read plan-gated results.

These map to the recipe workflows scraping-based Form 4 MCPs don't offer — each one leans on data this MCP alone exposes (10b5-1 flags, Form 144, 13F-HR join, per-insider return scoring). Plan requirements are honored honestly: prompts that touch Business-plan tools (get_signals, get_sentiment, get_form144, get_holdings, get_managers) or Pro-plan tools (get_insider_career_summary, get_insider_scorecard) instruct the LLM to surface the structured upgrade_required response rather than silently failing.

In Claude Desktop, prompts surface as a / slash-command picker; in Claude Code or other MCP clients, ask the assistant to "use the insider_monitor prompt for NVDA" (or similar) and it will fetch and follow the recipe.

Example prompts

Configured? Ask your LLM:

Quality-led (these require our amendment-aware, 10b5-1 clean, joined dataset):

• "Show me cluster buy signals from this week — discretionary only, no planned trades"
• "Which companies have insiders buying while institutional ownership is increasing this quarter?"
• "Are there any Form 144 filings at NVDA suggesting upcoming sales?"
• "What's the monthly insider sentiment for TSLA over the last 6 months, with 10b5-1 plans excluded?"
• "Berkshire Hathaway's top 13F-HR holdings — what did they add or trim this quarter?"

Standard insider research:

• "What insider trades happened at NVDA in the last 30 days, excluding 10b5-1 plans?"
• "What is Tim Cook's career insider-trading summary?"
• "Show me all open-market purchases over $1M at Tesla in 2026"
• "What has the CFO of Microsoft been doing with their shares this year?"

Why this MCP vs scraping-based alternatives

Some MCPs in this space scrape free public sites (e.g. openinsider.com) for Form 4 data. That's fine for a quick prototype but the data layer they give your LLM has structural gaps:

If your LLM session is the start of a real research workflow that eventually wants production delivery, picking the MCP that has a graduation path matters.

Beyond MCP — when you need more

The MCP is the easiest entry point. When your workflow grows past LLM-mediated research, the rest of the Form4API platform is right behind it:

• Webhooks — HMAC-signed, exponential backoff, dead-letter queue, auto-disable on persistent failure. For production pipelines, not just LLM chats.
• Python SDK — pip install form4api (PyPI)
• JS / TypeScript SDK — npm install form4api (npm)
• Dashboard — usage, billing self-serve, key rotation, webhook health, billing history.

The MCP wraps the same backend as all of the above — every fact your LLM cites can be re-fetched programmatically through any of these channels with the same key.

Plans

24 of the 38 tools work on the free plan, and every tool that is free today stays free. New premium capability gets tiered as it ships; nothing that already works on your key is taken away later.

For a bulk historical pull, use the REST /v1/transactions/export endpoint (Business) rather than paging — it streams the whole filtered set as CSV in one request.

What your agent sees at a paywall

A gated call never surfaces a bare HTTP error. The MCP returns a structured upgrade_required payload so the model can explain the situation and the fix in one turn:

{
  "error": "upgrade_required",
  "required_plan": "business",
  "current_plan": "Free",
  "message": "This endpoint requires the Business plan or higher. Your current plan is Free.",
  "unlocks": "The Business plan includes cluster-buy signals and sentiment scores, 13F institutional holdings and managers, Form 144 notices, bulk CSV export, and 250,000 requests/day."
}

message is the API's own explanation, kept verbatim — it names the specific limit or parameter that stopped the call, which is usually what the model needs to suggest a working alternative. The same shape is returned when a Pro-only parameter is used on an otherwise free tool, so the model can simply retry without that filter.

upgrade_url is included only when the API supplies one; it is omitted otherwise.

Data coverage

• 1.3M+ Form 4 transactions from SEC EDGAR
• 596K+ filings across all reporting companies
• 118K+ Form 144 notice-of-proposed-sale filings (Business+)
• 43M+ Form 13F-HR holdings across 114K+ filings, $71T+ AUM in the latest complete quarter (Business+)
• 3.3+ years of history (since 2023-04)
• 10b5-1 plan flag on every transaction
• Amendment-aware — Form 4/A reconciled
• Congressional STOCK Act trades (Pro+) — House Clerk PTR + Senate eFD, digital filings, amounts always shown as disclosed ranges (amountLow/amountHigh), never a fabricated midpoint, plus disclosureLagDays on every trade (up to 45 days under the STOCK Act)
• Real-time ingestion — new filings within minutes of SEC publication

Install as a Claude Desktop Extension (DXT)

A manifest.json is included at the repo root for one-click install via the Desktop Extensions (DXT) format. Once Claude Desktop supports .dxt files natively, pack and install with:

npx @anthropic-ai/dxt pack
# Produces form4api-mcp.dxt — open it in Claude Desktop to install

Until then, use the standard claude_desktop_config.json method described in Quick install above.

How tools stay in sync with the backend

This MCP is split between two layers:

• Hand-written tools in src/tools/*.ts (transactions, signals, sentiment, form144, holdings, …) — these carry the LLM-discriminator descriptions (amendment-aware, 10b5-1 clean, etc.) that make this MCP pick correctly over alternatives.
• Auto-generated tools in src/tools/_generated.ts — produced from https://api.form4api.com/openapi/v1.json by npm run codegen. Every new backend endpoint that lands in the OpenAPI spec flows in here automatically. CI runs npm run codegen:check on every PR and fails the build if the committed file drifts from what the live spec would produce, so the MCP is never silently behind the backend.

To add a new generated tool: ship the endpoint on the backend, regenerate (npm run codegen), commit src/tools/_generated.ts, publish. No tool-wrapper code needed.

The 6 recipe prompts live in src/prompts/recipes.ts — also hand-written, not generated. They add no new backend surface area; each one is a client-side template that tells the LLM which existing tools to call and in what order.

Links

• Form4API — API homepage
• Documentation — Full REST API reference
• Dashboard — Manage your API key, view usage, configure webhooks
• Status — live uptime, database, and ingestion-queue health
• npm — npm package
• mcp.so — MCP server directory listing