This directory runs its own MCP server. Point a client at one URL and an agent can search all 26 GTM APIs, read any provider in full, compare two, list the vendor MCP servers, open a costed stack and pull the markdown of any skill file. Same records as the pages, same verified dates, no key to create.
Streamable HTTP, stateless, read only. Everything it returns is already published on these pages, so there is nothing to authenticate.
https://frontal.so/gtm-apis/mcp # Claude Code claude mcp add --transport http gtm-directory https://frontal.so/gtm-apis/mcp # Anything that takes an mcpServers config block { "mcpServers": { "gtm-directory": { "url": "https://frontal.so/gtm-apis/mcp" } } }
Older clients are answered on their own protocol version. There is no SSE stream and no session to keep, so a client that only sends requests and reads replies works unchanged.
Every result carries the source URL and the date each figure was read, so an agent can cite the page rather than restate a price on its own authority.
Worth knowing before you wire it into something that spends money.
One dataset sits behind the pages, llms.txt, data.json, the skill files and this server. A price can be wrong in all five at once, but it cannot be wrong in only one. That is the same reason the skill files are generated rather than written.
A different thing to ours: these call the vendor's API with your key and your credits. Transport, auth, what each one can do and what it costs, read from each vendor's own documentation between 12 September 2026 and 19 October 2026.
The most complete GTM MCP published so far. It covers the whole motion rather than a slice of it: search, enrich, write contacts, build lists, create sequences and send one-off email, all through the same OAuth connection.
Two separate servers doing two different jobs. The remote one gives an AI client read and write access to contacts, deals and engagements. The local one drives the developer platform CLI for scaffolding and deploys.
Two tools by default, web_search_exa and web_fetch_exa, with an advanced search tool and a multi-step Agent tool you enable by appending them to the URL. The Agent tool is the one that does list building and enrichment.
Every LeadMagic plan ships the MCP server alongside the API and the terminal client, with no separate charge or tier gate. It exposes the same 15+ enrichment endpoints the REST API does, drawing on the same credit pool.
Published as one third of Attio's developer platform, next to the REST API and the SDK. Because Attio lets you define custom objects, the MCP surface reflects your schema rather than a fixed contact-and-deal model.
Clay's server runs its find-and-enrich tools, and the Functions a workspace admin switches on for MCP, on behalf of a signed-in user. Admins set a credit budget for every rep who connects and can give reps MCP-only access, with no view of the tables, keys or workflows behind it.
Company and people search and counts, dictionaries for industries, codes, countries, regions, departments, seniorities and technologies, and your account's compliance and intent settings. Compliance is enforced by the server: do-not-call numbers are masked and geo-restricted records excluded.
Plain-language expert tools that answer a question or build a Crunchbase search, structured tools for exact lookups and searches, and list tools that save results back to Crunchbase. Answers draw on company, funding, investor, people, acquisition and prediction data, as far as the signed-in user's access reaches.
Four tools: list campaigns, list a campaign's lead lists, find a lead by email or LinkedIn URL, and upload up to 1,000 LinkedIn profiles to a campaign as a new lead list. The first three only read; the upload creates a list.
One connection covers every LinkedIn account you manage in Expandi. An assistant builds lead lists, drafts campaigns, works the inbox and explains why outreach stalled, inside the same daily limits and warm-up as the rest of Expandi.
Three ways in: keyless, for Search, Scrape and Parse within daily limits; a browser sign-in; or an API key as a bearer token. Signing in and a key reach the same full tool surface on your team's plan.
Search people and companies in plain language, enrich them with emails and phone numbers one at a time or as a batch, and export the results as CSV or JSON without leaving the chat. The assistant is meant to show results and confirm before it spends anything on enrichment.
Discover, Domain Search, the Email Finder and Verifier, enrichment and leads management, so an assistant can find a contact, verify the address and save it to a Hunter list in one conversation. The key goes in an X-API-Key header, or a client that supports it can use the one-click OAuth flow instead.
Every Instantly V2 endpoint is a tool, generated from the same spec as the API reference: accounts and warmup, campaigns and subsequences, leads and lists, the inbox, analytics, webhooks, enrichment and workspace settings. The scopes on the API key decide what an agent can reach.
Find people and companies, check that someone still works where you think, and pull titles and verified contact details in plain language. Claude, ChatGPT and Codex connect over OAuth through Lusha's listing; any other client sends the API key in an x-api-key header.
Pipedrive's own server reads and writes deals, persons, organizations, leads, products, notes and activities, converts leads into deals and reads pipeline stages. It inherits the user's permissions, and every change lands in the change log.
Eight tools: person and company enrichment, bulk enrichment of up to 25 at a time, people and company search, filter suggestions and an account check. A search returns contact previews with their verification status, and the credits for the real values are only spent on enrichment.
Seventy tools across sequences, contacts, the inbox, tasks and the Jason AI SDR, 31 annotated read-only and 39 as changes. A scoped personal API key limits an agent to the areas you grant; an OAuth connection gets the full catalog.
An admin configures a server once in the org and any MCP client then connects per user: read, create, update and delete records under field-level security and sharing rules, and expose Apex actions, Flows, named queries and Prompt Builder templates as tools. The Salesforce DX MCP Server, in beta, covers developer work.
Thirteen read-only tools over accounts, people, opportunities, users and conversations, recordings included. It follows Salesloft's permission model, so a user only sees through it what they can already see in Salesloft.
Campaign insights, deliverability and health diagnostics, and lead, account and performance data inside Claude Desktop, reached through the mcp-remote bridge with the API key in the URL. Smartlead's guide lists reads and diagnostics rather than write actions.
tavily-search for real-time web search with filters and tavily-extract to pull content from pages, through a remote server that takes the API key in the URL or signs you in over OAuth, with a keyless option as well.
Search job postings and companies, read a company's technographics and buying intents, and manage company lists, saved searches and webhooks. Tools that change something are flagged so the client can ask before running them; read-only ones run without asking.
The official open-source server, installed from npm and run on your own machine. It validates single addresses or batches, uploads CSV files for bulk validation and AI scoring, finds addresses by name and domain, checks recent activity, and reports credits and API usage.
Two of the 26 providers here document no MCP server. That is not a mark against them: a sourcing or verification API called once per row in a pipeline has little to gain from the protocol.
The same test we run before building anything on it.
The agent-driven account research stack chains Exa, Apollo and Attio over MCP, costs $0.09 an account and ships with a skill file you can hand to an agent. It is the shape this page is describing, with the arithmetic attached.
Apollo prohibits model training on data pulled through its MCP and asks you to confirm training is off in your client first. That is an obligation on you with no technical enforcement behind it, so check it, and check it again whenever anyone changes client.
The hard part is the line between what an agent should do on demand and what belongs in a pipeline that runs every night. Frontal builds both, inside the tools you already run.
Worked with 300+ companies. A senior operator runs your account.
An endpoint an AI client connects to, which lists the tools it can call and then runs them. For a GTM API it means an assistant can search, enrich or write a record without anybody writing an integration first.
Yes, and it is the first thing on this page: seven read-only tools over Streamable HTTP, no auth, answering from the same dataset that draws every page. Point a client at it and ask for a provider by name.
13 of the 26 in the directory, and the number moves most months. Each provider review states whether there is one and where it lives, read from the vendor's own documentation.
No, and treating it as one is the common mistake. An MCP call happens because somebody asked, which is the wrong shape for a nightly job over tens of thousands of rows with retries and a warehouse write. Use the REST API for the pipeline and MCP for the conversation.
