MCP link
TRANXACT delegated work
The governed server a licensed Claude worker uses for one task an agent handed out: the agent’s own tools plus get_delegation, call_api and submit_result, every call through the agent’s guardrails and approvals.
https://<your-deployment>/api/v1/mcp/runMake your link
Stays in this browser. What you type here is used only to fill in the snippets on this page. The page makes no network request with it, does not store it (not in cookies, not in local storage), and has no form to submit, so it never reaches help.tranxact.ai or anybody else. Close or reload the tab and it is gone. Copying a snippet puts your worker token on your clipboard; paste it where it belongs and nowhere else.
What it is
When one of your company’s agents meets a task too large for one model turn — reconciling a year of supplier invoices, reading every open quality notification — it can hand the task to a licensed Claude worker: a person in Claude Code (in VS Code or the terminal), an Agent SDK session or a Managed Agent. The worker’s Claude connects to this server with a short-lived worker token and may do exactly what the agent may, no more: the agent’s own tool grants, under its guardrails, autonomy level and approvals, every call recorded against the task.
This is not the published API. An API key does not open it and a worker token does not open /api/v1/mcp. The background is in the guide TRANXACT as an MCP server.
| Endpoint | POST https://<your-deployment>/api/v1/mcp/run |
|---|---|
| Transport | MCP streamable HTTP, request and response only; GET answers 405 (no event stream). |
| Protocol version | 2025-06-18 (also 2025-03-26, 2024-11-05 on request). |
| Methods | initialize, ping, tools/list, tools/call, notifications (202). Anything else: -32601. |
| Authentication | Authorization: Bearer <worker token>. Only that header; X-API-Key is not read here. |
| Credential | A worker token, minted when a task is claimed: one task, 60 minutes by default, renewable while the task is open. |
| Multi-tenant deployments | Nothing extra: the worker token names its own account. |
| Server name | TRANXACT — <task number>; its instructions name the task and say to call get_delegation first and submit_result last. |
Get the worker token
Screens are named as the console names them: module › group › section › screen.
- An agent hands the work out. An agent (or a workflow tool node) whose grants include a hand to a licensed worker tool calls it, after its own guardrails and approvals. The task appears on Agentic AI › Running › Operate › Delegated Tasks as queued, with a task number (DLG-…).
- Claim it. A person with the Heavy Worker role selects the task and chooses Claim, picking where the work will be done (Worker; Claude Code in VS Code is “A person in Claude Code”).
- Copy the token. The result panel shows the worker token, the MCP address and ready-to-paste configuration for Claude Code, the Agent SDK and a Managed Agent, once, with copy buttons. Paste the token into the panel above for the other clients.
- Work, then hand back. Ask Claude: “Work TRANXACT task DLG-000123: call get_delegation, do what it asks, then submit_result.”
submit_result(or Submit result on the screen) ends the token. - Running long? The token lasts 60 minutes unless your company changed
agents.worker_token_minuteson Platform › Configuration › Core Settings › Runtime Settings (5 to 240). Choose Renew the token on the task while it is open, and replace the token in your client.
Connect a client
Every snippet below is filled in by Make your link; each has a Copy button.
Claude Code (terminal or VS Code)
claude mcp add --transport http tranxact https://<your-deployment>/api/v1/mcp/run \
--header "Authorization: Bearer <your-worker-token>"Add --scope project to write it into the folder’s .mcp.json for everyone who opens it (keep that file out of version control while it holds a worker token), or --scope user for every folder. claude mcp list shows whether it connected; /mcp inside Claude Code shows its tools. The same thing as a .mcp.json file:
{
"mcpServers": {
"tranxact": {
"type": "http",
"url": "https://<your-deployment>/api/v1/mcp/run",
"headers": {
"Authorization": "Bearer <your-worker-token>"
}
}
}
}Claude Desktop
{
"mcpServers": {
"tranxact": {
"command": "npx",
"args": [
"-y",
"mcp-remote",
"https://<your-deployment>/api/v1/mcp/run",
"--header",
"Authorization:${TRANXACT_AUTH}"
],
"env": {
"TRANXACT_AUTH": "Bearer <your-worker-token>"
}
}
}
}Claude Desktop: Settings › Developer › Edit Config opens claude_desktop_config.json (Windows %APPDATA%\Claude, macOS ~/Library/Application Support/Claude). Merge this in and restart Claude Desktop. It runs mcp-remote (Node.js 18 or later), which carries the header to TRANXACT. The header value lives in env rather than in args because a space inside an argument is mangled on Windows.
Custom connectors added in claude.ai (Settings › Connectors › Add custom connector) ask for a URL and, optionally, OAuth client credentials; they have no place for a fixed Authorization header. TRANXACT’s MCP endpoints take a worker token in a header and do not offer OAuth, so connect through Claude Desktop’s configuration file above, Claude Code, or another client that sends headers.
Cursor
{
"mcpServers": {
"tranxact": {
"url": "https://<your-deployment>/api/v1/mcp/run",
"headers": {
"Authorization": "Bearer <your-worker-token>"
}
}
}
}Then Cursor Settings › MCP shows the server and its tools.
VS Code
{
"servers": {
"tranxact": {
"type": "http",
"url": "https://<your-deployment>/api/v1/mcp/run",
"headers": {
"Authorization": "Bearer <your-worker-token>"
}
}
}
}Better, VS Code can ask for the worker token when it starts the server and keep it out of the file:
{
"inputs": [
{
"type": "promptString",
"id": "tranxact-credential",
"description": "TRANXACT worker token",
"password": true
}
],
"servers": {
"tranxact": {
"type": "http",
"url": "https://<your-deployment>/api/v1/mcp/run",
"headers": {
"Authorization": "Bearer ${input:tranxact-credential}"
}
}
}
}Start it from the file (the Start button above the server) or with MCP: List Servers; use it from Copilot Chat in agent mode.
Test it with curl
For bash or zsh (on Windows, Git Bash or WSL). Each answers one JSON-RPC message.
curl -s https://<your-deployment>/api/v1/mcp/run \
-H "Authorization: Bearer <your-worker-token>" \
-H "Content-Type: application/json" \
-d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"curl","version":"1.0"}}}'curl -s https://<your-deployment>/api/v1/mcp/run \
-H "Authorization: Bearer <your-worker-token>" \
-H "Content-Type: application/json" \
-d '{"jsonrpc":"2.0","id":2,"method":"tools/list"}'curl -s https://<your-deployment>/api/v1/mcp/run \
-H "Authorization: Bearer <your-worker-token>" \
-H "Content-Type: application/json" \
-d '{"jsonrpc":"2.0","id":3,"method":"tools/call","params":{"name":"get_delegation","arguments":{}}}'Tools
tools/list answers the agent’s own tools first, then six helpers.
| Tool | What it does |
|---|---|
| The agent’s own grants | Exactly the tools the agent’s model would be offered, with the same names (for example crm_leads_find) and schemas, marked read-only, destructive or external as the grant is. Not offered: finish, ask_human, summarize, and any delegate tool (work cannot be handed on again). |
get_delegation | The task: the brief in full, the structured input, the agent, when it expires, files attached so far and the tool names. Call it first. |
find_endpoints | Looks up TRANXACT API endpoints by module, resource or a fragment of the path (contains), limited to the agent identity’s modules; each line marked [granted] where one of the agent’s tools reaches it. |
call_api | method + path (+ body). Accepted only where it matches one of the agent’s granted resource tools (list, read, create, change, delete or an action), and then made as that tool, with its guardrails and approvals. Anything else is refused and recorded. |
check_approval | approvalId: where an approval one of your calls raised has got to. Once a person approves, it makes the call (once) and returns the result. |
attach_file | fileName and contentBase64 or text: files a document through the company’s file inspection and attaches it to the task. |
submit_result | summary (+ result): hands the work back. Ends the token. The run carries on, after a person reviews it where the task asks for that. |
A call that must wait for a person comes back as a result saying so, with the approval id for check_approval. A refusal is a tool result with isError: true and the reason.
Limits and reach
| Limit | How it works |
|---|---|
| Token lifetime | 60 minutes by default (5 to 240, Runtime Settings agents.worker_token_minutes, or the deployment’s DELEGATION_TOKEN_MINUTES); never past the task’s own expiry. Renewable while the task is open. |
| Token ends | On submit, release, cancel or expiry, and when the run ends (finished, failed, cancelled or stopped by a guardrail). Revocation is checked on every call. |
| Result size | A tool result is cut at 400,000 characters unless the company changed agents.worker_result_chars; ask for less, or filter. |
| Files | attach_file takes up to 25 MB unless the company changed agents.worker_file_mb. |
| Guardrails and approvals | Every call goes through the agent’s grant, its precondition and fixed arguments, the guardrails (spend and value caps, allowed hours, scope, personal data…), the autonomy level and any approval. A guardrail that halts stops the run and ends the task. |
| Counting | Calls are counted on the task, not against the run’s own tool-call ceiling. |
What the token can reach, and what it cannot
- Can: this one task; the agent’s own tool grants, as the agent’s run-as identity (its roles and modules); the six helpers.
- Cannot: any other task; any endpoint outside the agent’s grants; the agentic, sign-in, administration or published-API endpoints; the card vault. Identifier fields (card and bank account numbers, national IDs, passports, IBANs, SSNs) are withheld from every result. The token is refused everywhere except this address, even from the same machine.
- Note: the pseudonymisation an in-process agent gets does not apply: your model sees what the agent’s grants let its identity read, less the withheld identifiers. A company that wants a person to approve the hand-off marks the delegate tool as touching personal data.
Errors people commonly hit
| Answer | What it means | What to do |
|---|---|---|
| 401 A worker token is required. Claim a delegated task to get one. | No Authorization: Bearer header arrived. | Check the header in your client’s configuration. |
| 401 This worker token has expired or is not valid. Renew it, or claim the task again. | The hour (or your company’s setting) is up, or the token was mistyped. | Renew the token on the task, and replace it in your client. |
| 401 This endpoint takes a delegated worker token… A sign-in token or an API key does not open it. | An API key or a console sign-in token was sent. | Use the worker token from Claim. An API key belongs on /api/v1/mcp. |
| 401 This worker session has ended: the task was submitted, released, cancelled or expired. | The task is no longer yours to work. | Claim it again if it is back in the queue. |
| 401 This worker token no longer belongs to an open task. | The task moved on, or the token is from an earlier claim. | Use the token from the latest Claim or Renew. |
| 400 / 404 No account… | Only on a multi-tenant deployment, when the token could not be read. | Use the token exactly as Claim gave it. |
Tool result isError: No tool granted to this agent reaches … | call_api asked for something outside the agent’s grants. The answer lists what can be called. | Use a listed tool, or ask the agent’s owner for a grant. |
Tool result isError: A delegated worker cannot call the agent, authentication, administration or published API endpoints. | Reserved endpoints. | Not available to a worker. |
| Tool result: A person must approve this call before it is made… | A grant, tool or guardrail asks first. | Carry on with other work; call check_approval with the id. |
| 405 | A GET, or a method other than POST. | POST JSON-RPC to the same address. |
An HTTP refusal is JSON: {"error": "a sentence saying what was wrong"}; a 401 also carries WWW-Authenticate: Bearer.
Security
- This help site never sees your token. The Make your link panel fills the snippets in your browser; nothing is sent or stored.
- The token is shown once, at Claim (or Renew). It is the agent’s identity, not yours, and every call is audited with your name as the worker.
- End it with submit_result, or on Agentic AI › Running › Operate › Delegated Tasks with Submit result, Release or Cancel: the token stops at once, including calls in flight.
- Keep it out of version control. A project
.mcp.jsonholding a live token belongs in.gitignore. - Your model is outside TRANXACT: keep the agent’s grants as narrow as the task needs.
The other endpoint: TRANXACT published data. All endpoints and deployments: MCP links.