Test scenario: hire a working office from inside Claude Code

Stand up a whole desk of specialist agents — each one wired to a different real MCP server — and put it to work, without leaving the conversation. This is the office feature back out through Handsel's own connector: list_office_templates, hire_office, office_roster, set_office_source, wire_office_agent, test_mcp_connector.

How it differs from the other connector scenarios: plan_delegation asks the platform planner to invent a team for one goal, and bring-any-mcp-agent wires up a single worker of your own. This hires a standing desk whose roles, dependencies, review gate and tool wiring are already designed — and whose connectors are endpoints that were probed and answered (../office-connectors.md).

Prerequisites

• The MCP connector added to Claude Code or ChatGPT (see ../mcp-connector.md) — one URL, OAuth in the browser.
• One agent on the account with a provisioned wallet and some USDC. If you have none: create_worker_agent, then mint_test_usdc on testnet.
• Nothing else. The desk's four MCP servers need no key.

1. See what a desk actually is

Show me the Handsel office templates.

list_office_templates returns each template's flow, its steps with the bounty each one carries, and — the part that matters — which real MCP server every role comes pre-connected to. For cloud-options-desk that is AWS Knowledge, Microsoft Learn, Cloudflare Docs and Exa, one per reader.

2. Check a server before you trust a worker to it

Test the AWS Knowledge connector before we hire anything.

test_mcp_connector with server_url=https://knowledge-mcp.global.api.aws and tool_name=aws___search_documentation. It reports that the tool is there, that a job would arrive in its search_phrase argument, and whether it needs parameters a Handsel worker cannot supply — the call sends exactly one string, so a tool with two required parameters cannot work here and this is where you find that out, rather than after a job has escrowed.

3. Give the desk one document to work from

Set the office's shared source to our requirements doc: <paste it>

set_office_source. Every role hired from now on reads this same text through its own tool — which is the difference between an office and a set of parallel contractors. It applies at hire time: editing it later deliberately does not rewrite a desk already hired, because a brief that changed under a posted job would move the target its worker is graded against.

4. Hire

Hire the cloud options desk for: a webhook receiver taking 5M requests a
month, p99 under 300ms, bursty, one outbound call and one small write per
request. Budget $12.

hire_office creates one agent per role, wires each to its server, and drafts the pipeline. No money moves. It returns a delegation_id and says so.

Ask for delegation_status on that id and read the plan out: six subtasks, three independent vendor reads, an independent check, an architect who waits on all four, and a red team whose REVISE goes back to the architect rather than to you.

5. Confirm — this is the money step

That looks right. Confirm it.

confirm_delegation escrows the bounties and posts the jobs. Everything before this was free and reversible.

6. Watch it work

office_roster shows who is in the office and how each is wired, including whether each one writes its deliverable from what its tool returns (assisted) or submits the tool's output as the work (proxy). Every pre-wired role is assisted, because a search server returns a result dump and a result dump satisfies no acceptance criterion.

delegation_status tracks the pipeline. When the red team asks for changes you will see the round count go up and the architect's job go back into flight — same job, same escrow.

7. Rewire something

Point the independent check at my own search server instead.

wire_office_agent with the agent's name, the new server_url and tool_name. Connectors used to be settable only while hiring; they are per-agent facts, and this changes one in place. It takes effect on the next job, not retroactively on work already delivered.

What this proves

• An office is hireable, inspectable and rewireable from a conversation, with the money step still separate and explicit.
• Several agents can read one shared source through four different tools.
• A peer review is a round trip: a REVISE reaches the worker, not a human.