TRANXACT as an MCP server
An AI assistant or agent that speaks the Model Context Protocol can read and write what an administrator has published, at <your TRANXACT address>/api/v1/mcp, with an API key whose protocol is MCP or both. A second, governed endpoint (/api/v1/mcp/run) is for a licensed Claude worker taking on a task an agent delegated.
The published MCP endpoint
The same publication that serves the REST API serves MCP: an exposure with Offer as an MCP tool ticked becomes tools for any client granted it. The endpoint is POST <your TRANXACT address>/api/v1/mcp, MCP over streamable HTTP, request and response (there is no event stream; GET answers 405).
- On Platform › Published API › Exposures, tick Offer as an MCP tool on what the model may use, and write What to tell a model: a sentence the tool’s description carries.
- On API Clients, register the assistant with Protocol MCP (or both), Acts as a named user, and optionally What to tell an MCP client (the server’s instructions). Issue key.
- Grant it the exposures on API Grants, and name where it runs on Allowed Software.
The tools
| 도구 | When it is offered |
|---|---|
| list_<code> | the grant may read: search (q), limit, offset, sort |
| get_<code> | the grant may read: one record by id |
| create_<code> | the grant may create |
| update_<code> | the grant may change |
| delete_<code> | the grant may delete |
<code> is the exposure’s Published as name, in lower case with other characters as underscores. tools/list answers exactly the tools this key may call; initialize, ping and notifications are answered too.
Connecting a client
# Claude Code (terminal or VS Code)
claude mcp add --transport http tranxact https://demo.tranxact.ai/api/v1/mcp \
--header "Authorization: Bearer tx_assistant_xxxxxxxxxxxxxxxx"
# any MCP client that takes a JSON configuration (.mcp.json and similar)
{ "mcpServers": { "tranxact": { "type": "http",
"url": "https://demo.tranxact.ai/api/v1/mcp",
"headers": { "Authorization": "Bearer tx_assistant_xxxxxxxxxxxxxxxx" } } } }
Any MCP client that can call an HTTP server and send an Authorization (or X-API-Key) header can connect. Keep a configuration holding a key out of version control.
What governs it
- Exactly the REST rules: the key, the client, Allowed Software, the rate limit, the exposure and the grant, then the same route the screen uses as the Acts as user.
- A write is a write to a live business system: the same validation, numbering, posting rules and audit trail apply as when a person does it.
- Every call, tools/list and tools/call included, is on the API Call Log.
Delegated agent work: /api/v1/mcp/run
When one of the company’s agents hands a task too large for one turn to a licensed Claude worker (Agentic AI › Operate › Delegated Tasks, Claim), the worker connects to POST <your TRANXACT address>/api/v1/mcp/run with a short-lived worker token minted at the claim, not with an API key. It may do exactly what the agent may: the agent’s own tool grants, plus get_delegation (the brief), find_endpoints, call_api (accepted only where it matches one of the agent’s granted tools), check_approval, attach_file and submit_result, which hands the result back and ends the token.
- The token lasts 60 minutes unless the company sets otherwise, can be renewed while the task is open, and is revoked on submit, release, cancellation, expiry or the end of the run.
- Every call goes through the agent’s guardrails, autonomy level and approvals, and is recorded against the task number in the agent’s tool calls and the audit trail, with the worker named.
- The card vault is never reachable, and identifier fields (card and bank account numbers, national IDs, passports and the like) are withheld from every result.
- Claim shows the address and a ready-to-paste configuration with the token filled in.