coresignal-mcp

# Coresignal MCP v2 The official [Model Context Protocol](https://modelcontextprotocol.io) server for [Coresignal](https://coresignal.com) — bring fresh B2B data on **895M+ employees**, **70M+ companies**, and **468M+ job postings** straight into your AI assistant. ```bash https://mcp.coresignal.com/mcp/v2 ``` Search in plain natural language, pull full records, enrich contacts with verified emails, and export large result sets as downloadable files — all from Claude, Cursor, Codex, VS Code, or any other MCP-compatible client. --- ## Features - **OAuth 2.1 authentication** — sign in with your Coresignal dashboard account; no API keys in config files. - **Natural-language search** with match evidence in every result row, so you can see *why* each record matched. - **Cost transparency** — every response reports `credits_consumed`, and expensive calls ask for confirmation before spending. - **Field discovery** — `entity_fields` finds the right field names by keyword, free, without loading the full 300+ field vocabulary into context. - **File downloads** — large result sets are stored server-side and returned as a download link instead of flooding the chat. - **`artifact_read`** pages delivered files back into the conversation for free — works even in clients with no filesystem access. ## Prerequisites - A [Coresignal account](https://dashboard.coresignal.com) with an active subscription (credits) and a team API key provisioned in the dashboard. The MCP server resolves your team's key automatically after sign-in — you never paste it into a config file. No subscription yet? Start with the **7-day free trial** — it includes **2,000 credits**. See the [plan comparison](https://coresignal.com/pricing/#plan-comparison). On first connection your client opens a browser window to sign in to the Coresignal dashboard. That's the whole setup — no environment variables, no secrets. --- ## Client setup <details> <summary><b>Claude Desktop</b></summary> Claude Desktop supports remote MCP servers natively via **Connectors**: 1. Open **Settings → Connectors → Add custom connector**. 2. Name: `Coresignal`, URL: `https://mcp.coresignal.com/mcp/v2`. 3. Click **Add**, then **Connect** — a browser window opens to sign in to your Coresignal dashboard account. > **Important — enable file downloads:** Claude Desktop blocks downloads from domains it doesn't know. To download result files (large fetches are delivered as links), add `mcp.coresignal.com` to the **Domain allowlist** in Claude Desktop settings (on Team/Enterprise plans your admin manages this). Without it, download links from the server will be blocked — though you can always read the data back in-chat with the free `artifact_read` tool instead. See [File downloads](#file-downloads--the-domain-allowlist). </details> <details> <summary><b>Claude Code</b></summary> ```bash claude mcp add --transport http coresignal https://mcp.coresignal.com/mcp/v2 ``` Then inside a session run `/mcp`, select **coresignal**, and complete the browser sign-in. Re-run `/mcp` any time you need to re-authenticate. </details> <details> <summary><b>Codex (OpenAI Codex CLI)</b></summary> ```bash codex mcp add coresignal --url https://mcp.coresignal.com/mcp/v2 ``` A browser window opens to sign in to your Coresignal dashboard account. </details> <details> <summary><b>Cursor</b></summary> Add to `~/.cursor/mcp.json`: ```json { "mcpServers": { "coresignal": { "url": "https://mcp.coresignal.com/mcp/v2" } } } ``` Cursor detects that the server requires authentication and shows a **Needs login** prompt — click it to complete the browser sign-in. </details> <details> <summary><b>OpenCode</b></summary> Add Coresignal MCP: ```bash opencode mcp add coresignal --url https://mcp.coresignal.com/mcp/v2 ``` Authenticate: ```bash opencode mcp auth coresignal ``` A browser window opens to sign in to your Coresignal dashboard account. </details> <details> <summary><b>Visual Studio Code</b></summary> Run **MCP: Add Server** from the Command Palette, choose **HTTP**, and enter the URL — or add to `.vscode/mcp.json`: ```json { "servers": { "coresignal": { "type": "http", "url": "https://mcp.coresignal.com/mcp/v2" } } } ``` VS Code prompts you to authorize the server on first use and handles the OAuth flow in your browser. </details> <details> <summary><b>Cline</b></summary> Add to `~/.cline/data/settings/cline_mcp_settings.json`: ```json { "mcpServers": { "coresignal": { "type": "streamableHttp", "url": "https://mcp.coresignal.com/mcp/v2", "disabled": false } } } ``` Open Cline MCP interface: ```bash cline mcp ``` From menu select **Authorize OAuth** and select **coresignal**. Cline will handle the OAuth flow in your browser. </details> --- ## File downloads & the domain allowlist Large results (and any call with `delivery="url"`) are not dumped into the chat to save LLM input tokens. Instead the server stores the rows as a file and returns a **signed HTTPS download link** that expires after **1 hour**: ```json { "delivery": "url", "url": "https://mcp.coresignal.com/mcp/v2/artifacts/…?exp=…&sig=…", "artifact_name": "employee_fetch-baf941dd4f8c4dfe.jsonl", "count": 500, "credits_consumed": 10000 } ``` From here, the agent gets at the data in one of three ways: 1. **The agent downloads the file itself** — in clients with shell access, the agent will typically `curl` the link to disk and analyze the file locally with `grep`/`jq`/pandas. 2. **The agent reads it back in-chat** — in clients with no shell and no filesystem, the agent calls `artifact_read(artifact_name, offset, limit)` instead, paging through the file in slices over the MCP session. This is **free** and works everywhere, so nothing floods the chat. 3. **You download it manually** — if the agent itself isn't allowed to fetch the link (domain not allowlisted, sandbox without network access), click the link, save the file, and tell the agent where it is: *"I've downloaded the file to ~/Downloads/employee_fetch-….jsonl — analyze it from there."* Works in any client that can read local files. > **Prefer the download when your client supports it.** `artifact_read` is free in Coresignal credits, but not in LLM tokens: every page it returns becomes part of the conversation and is re-billed as input tokens on each subsequent turn. Each page is also trimmed to a fixed token budget — a full employee record is ~8k tokens, so a single page carries only a handful of full records regardless of the `limit` you ask for. Reading a large file that way takes hundreds of calls and can exhaust the context window before you reach the end. A downloaded file costs essentially no tokens — the assistant can filter thousands of rows locally with `grep`/`jq` and surface only the answer. That's why it pays to get downloads working up front (domain allowlist, sandbox network access) and keep `artifact_read` for clients that can't download or for eyeballing a few rows. **If downloads are blocked, nothing is lost:** the records are already stored and paid for — read them with `artifact_read`. Never re-run a fetch to "recover" a file; that bills every record a second time. --- ## Tools | Tool | What it does | Cost | | --- | --- | --- | | `entity_search` | Natural-language search over employees, companies, or jobs | 20 credits per search (flat) | | `entity_fields` | Keyword search over an entity's ~300 field names | Free | | `entity_fetch` | Pull full (JSONL) or projected records, by search handle or by id | 20 credits per employee/company record, 1 per job record | | `email_enrich` | Verified business emails for employee ids (CSV) | 10 credits per email **found** (misses are free) | | `artifact_read` | Page rows back out of a delivered file | Free | ### `entity_search` Searches `employee`, `company`, or `job` records with a plain-language query: > *"Senior Python developers at fintech companies in French"* Every call costs a flat **20 credits** and returns: - `total_count` — the exact number of records the query matched, - up to 20 preview rows (`limit=0` returns just the count), each showing the fields the query matched on — the *evidence* for why each result is there, - a `cache_id` — a 1-hour handle to the search that **saves credits and time**: pass it to `entity_fetch` and the matched records are collected straight away — no need to re-run (and re-pay for) the search, and resolving the `cache_id` itself is free. Record ids stay server-side, so nothing bulky ever passes through the conversation. ### `entity_fields` Free, instant lookup of field names by meaning — `"salary"` finds the compensation fields, `"current job title and seniority"` finds `active_experience_title`, `experience.position_title`, etc. Use it to build the `fields` list for a custom-scope fetch without ever loading the full field vocabulary into context. ### `entity_fetch` Collects records — either up to 20 hand-picked ids from search results, or up to **1,000 records per call** via a `cache_id`. Billing is per record found: **20 credits** for employees/companies, **1** for jobs. ### `email_enrich` Verified, deliverable business emails for up to 1,000 employee ids — **10 credits per email found**; not-found ids are free. EEA/UK contacts are not accessible (GDPR). ### `artifact_read` Reads a delivered file back over the authenticated MCP session, a page at a time — **free**, since the records were billed when they were fetched. This is what makes file delivery work everywhere, including chat clients that can't open a link or touch a filesystem. --- ## Example prompts ### Market scan > *"Find B2B SaaS companies in the Nordics with 50–200 employees that raised funding in the last two years."* ### Build a lead list with verified emails > *"Search for heads of data at US companies with 500+ employees. Fetch 20 full profiles to a file with verified emails."* ### Deep-dive a single company** > *"Pull the full record for flo.health — funding rounds, headcount growth, and current job openings."* ... --- ## Credits & billing | Action | Credits | |---|---| | `entity_search` (any `limit`, including 0) | 20 per search | | `entity_fetch` — employee or company | 20 per record found | | `entity_fetch` — job | 1 per record found | | `email_enrich` | 10 per email found; misses free | | `entity_fields`, `artifact_read` | Free | Every response includes `credits_consumed` — the actual billed amount, so a discrepancy (sent 1,000 ids, billed for 950 records) tells you exactly how many ids weren't found. The server never spends silently: fetches confirm field scope with you first, and a fetch that can't be delivered fails *before* any credits are spent. Credits are drawn from your team's Coresignal subscription — manage keys and billing in the [dashboard](https://dashboard.coresignal.com). New to Coresignal? The **7-day free trial comes with 2,000 credits** — that's 100 searches, or 100 employee/company records, or a mix — see the [plan comparison](https://coresignal.com/pricing/#plan-comparison). ## Links - [Coresignal](https://coresignal.com) — data coverage, plans, and pricing - [Coresignal dashboard](https://dashboard.coresignal.com) — account, team, API keys, credits - [API documentation](https://docs.coresignal.com/api) — the underlying data APIs [](https://smithery.ai/servers/coresignal/coresignal)