Skip to content
actionlake
Start free

MCP server documentation

/docs

Short on purpose.

The actionlake MCP server lives at one URL and gives an agent eight tools: help, find_actions, get_action, do, ask, check, cost and account. Add it to Claude Code with one command, or to any MCP client that speaks Streamable HTTP, and sign in with OAuth 2.1 or a bearer key. Everything is also on REST.

Start freeRead the quickstart
Or connect from Claude Code; no key needed, the consent page creates a free account
claude mcp add --transport http actionlake https://connect.actionlake.com/mcp
what the agent sees on connect
help {}

  actionlake · 8 tools · 1,000+ connectors · 289,000+ actions
  billed: do, ask
  free: everything else, and every cached answer

  find_actions   what can I do?
  get_action     how exactly do I do it?
  do             do it
  ask            answer a question across systems
  check          will this work?  (free)
  cost           what will it cost?
  account        tenant · keys · connections · plans
Illustrative: the shape of the help reply, not a transcript.
  1. 01QuickstartConnect an agent, run a public action, connect a system, read the receipt.four steps
  2. 02Connect an agentClaude, ChatGPT, Copilot, Cursor, Gemini, Grok and more: the exact steps for each.every MCP client
  3. 03API referenceThe REST twin: same names, same receipts, same price book.REST · openapi.json
  4. 04ConnectorsWhat agents can reach, and how a new system joins.1,000+ connectors

The surface

Eight tools. Named for what an agent wants.

Only two spend money, and they cannot be confused: do is a verb, ask is a question. The other six are free, always, including the dry run.

01helporientation, a topic, or "why did that fail?"free
02find_actionsplain-language search over 289,000+ actionsfree
03get_actionthe schema, output shape, cost and an examplefree
04doone action, or several in orderper upstream call
05askfetch from several systems and join them; the answer, not the rowsper upstream call
06checkdry run: inputs, permissions, connections, cost, what needs a confirmationfree
07costspend, forecast, cap headroom, the price bookfree
08accounttenant, keys, connections, plans, schedules, approvals, billingfree

The agent is the planner. ask takes structured steps and a join, not a sentence; find_actions, get_action and check exist to make writing that plan cheap. REST and MCP are one implementation; a test holds them to it.

Connect

If your agent speaks MCP, you are already integrated.

Claude, ChatGPT, Copilot, Cursor, Gemini, Grok, Codex and any other MCP client take the same URL. Send a key, or sign in on the consent page, which creates a free account on the spot.

  • OAuth 2.1 sign-in, or a key in Authorization or X-API-Key
  • 26 ways in, step by step, on Connect an agent
  • 8 tools, however many systems you connect
  • 100,000 free calls on the sandbox, no card
terminal
claude mcp add --transport http actionlake https://connect.actionlake.com/mcp
  1. 1Run the command once. Add --scope user before the name to use ActionLake in every project.
  2. 2Signing in instead of using a key? Leave the header off, start claude and run /mcp; the ActionLake consent page opens, and it can create a free account there.
  3. 3Ask your agent to call help. The eight tools are listed under actionlake.

Verified in ActionLake's code · every platform, step by step

For software

Machine-readable, on principle.

We sell infrastructure for software, so the documents that matter are fetchable by software, before you sign anything and without a key.

For AI agents

An agent can set itself up.

If you are an agent reading this: you do not need a person or a key to start.

  1. Connect to https://connect.actionlake.com/mcp with no credentials. help, find_actions and tenant.create need none.
  2. Call account {"op": "tenant.create", "name": "…"}. It returns a tenant and its key, shown once. Over REST it is POST https://api.actionlake.com/v1/account/tenant.create.
  3. Write the key into your own MCP config as Authorization: Bearer al_… (or X-API-Key) and reconnect. Connect an agent has the exact file for your client.
  4. Or ask your user to sign in: connect without a header, and the client opens the actionlake consent page, which signs them in or creates a free account in one click.

Connecting a vendor that uses OAuth, such as Xero, still needs a person once: register your own app with the vendor, sign in as an admin, and the tokens are stored as a connection. API-key vendors the agent connects itself with connections.create.

what the agent does
# 1  connect with no key: help, find_actions and tenant.create need nonemcp      https://connect.actionlake.com/mcp# 2  give yourself a tenantaccount  {"op": "tenant.create", "name": "my-agent"}         ✓ key: al_…   shown once - keep it# 3  put the key in your own MCP config, then reconnectheaders  {"Authorization": "Bearer al_…"}# or 3  have your user sign in instead: no header,#       the consent page signs them in or creates a free account

07Questions

Before the first call.

The questions people ask while they connect, with the answer an agent would give.

How do I add actionlake to Claude Code?

Run claude mcp add --transport http actionlake https://connect.actionlake.com/mcp, start Claude, and approve the consent screen. No key yet? The consent screen creates a sandbox with 100,000 free calls a month. Then say help {} to see the eight tools.

Does actionlake work with MCP clients other than Claude Code?

Any MCP client that supports remote servers over Streamable HTTP can add https://connect.actionlake.com/mcp and sign in with OAuth 2.1 (with dynamic client registration) or a bearer key. The Connect an agent page has the steps for each client, and says which ones we have not yet tested. Anything that can make an HTTPS request can use the REST API.

Do I need an API key to start?

No. Over MCP, the consent screen creates a sandbox account. Over REST, one unauthenticated call to tenant.create returns an account and a key with 100,000 free calls a month. The key is shown once, so keep it.

How does my agent know what an action costs before running it?

get_action shows each action's cost, and check runs a dry run: it validates inputs, permissions and connections, estimates calls and cost, and says what needs a confirmation. Both are free and send nothing to the vendor.

How do I limit what an agent can do?

Mint a key for it with allowed_actions (one action, a whole connector, a verb, read-only, or everything), a cap in calls, and a confirm policy: agent, human_for_writes or human. Revoking a key also revokes every key it minted.

What happens when a vendor rate-limits or fails?

A vendor's 401, 403, 429, 5xx or a timeout is not billed. Reads are cached per action, so repeats do not touch the vendor. Writes are never cached and never retried. Every error says what to do next.

Is there an OpenAPI spec and an llms.txt?

Yes: an OpenAPI description of the REST API, an llms.txt for agents, a public JSON price book and a machine-readable trust document. Each is linked in the Machine-readable section of this page, and none needs a key.

Ask one question across two systems.

Start freeSee pricing

100,000 free calls. No card. Hard caps on every key. Cached answers are free.