toolspoke
Start free
DocsOverview

Start here

Reference

Build and run

DocsStart here

Toolspoke documentation

Toolspoke is one gateway between your team's AI agents and the tools those agents need. Install a tool once, decide which part of it may be used and by which agents, and give every agent the same address.

What the gateway is for

An agent that needs to touch GitHub, Notion, Stripe and an internal API normally needs four credentials, and whoever set it up now holds four credentials on their laptop with no record of what was done with them. Toolspoke moves those credentials to one place, puts a per-project policy in front of them, and gives the agent a single MCP endpoint instead.

The agent never sees a credential. It sees the operations its project allows and its own key was granted, it asks the gateway to run one, and the gateway does the rest: resolve who the agent belongs to, check that both still permit it, execute the call, and record what happened. Every call writes a row carrying the person, the agent, the tool, the arguments, the response, the duration and the cost.

From sign-up to first call

  1. Create a workspace

    Signing up creates your account and a workspace around it, with a Default project inside and 5,000 credits on the Free plan. Nothing is billed and no card is asked for.

  2. Install a toolkit

    Pick something from the catalogue, or build your own from an API spec, an MCP server or a CLI. A toolkit is installed into one project and carries the operations that toolkit can perform.

  3. Store a credential

    The toolkit asks for whatever it needs to authenticate: an API key, a token, or a sign-in where the provider supports one. Credentials are encrypted with AES-256-GCM before they are written and are held per project.

  4. Choose what the project allows

    A project switches an installed toolkit's actions on and off individually, for everyone in it at once. Installing a toolkit does not commit the project to every operation it exposes.

  5. Mint an agent key

    An agent key belongs to the person who created it and reaches only the projects it was granted. It is shown once and stored as a SHA-256 hash, so it cannot be read back off the server.

  6. Connect a client

    Claude, OpenAI, OpenCode, VS Code and Cursor can sign in against the gateway and be issued their own token. Anything else takes the key in an Authorization header from a config file.

  7. The agent calls a tool

    It searches the catalogue, reads one input schema, and runs one tool. The gateway checks the project's policy and the key's grant again on the way through, executes the call, spends the credit, and writes the row the activity feed reads.

What an agent receives

The first-party catalogue is 1,143 individual actions across 102 connectors. Advertising all of them with their input schemas works out at roughly 73,000 tokens of an agent's context before it has read its first instruction, so the gateway does not. It advertises a small bridge instead, and the agent searches for what it needs. That token figure is our own estimate from this catalogue rather than a benchmark.

tools/list, on a workspace with everything installed
search_tools           find a tool by what it does, its input schema and its kind
list_toolkits          every toolkit installed here, names only
use_tool               run one read-only tool by its exact name
run_tool               run one tool that creates or updates
run_destructive_tool   run one tool that deletes
read_guide             advertised only when a reachable toolkit has a guide

A client that already knows a tool's real slug__action name can still call it straight through tools/call. The bridge is a cheaper way in, not a wall.

What these docs cover

Start here

Reference

Build and run