Where AI stops, ask a person

What it does

Ask a worker in Japan to do what only a person can do. Set the result wanted and a budget cap. Urgent tasks (within 10 minutes or 1 hour) reach available workers and the first to take it gets it. For other tasks, the person using the agent chooses one person from the list of applicants (user name, rating, track record, price); this can be changed for each order. The person using the agent checks the result that comes back (text and photos).

When to use it (where AI stops)

• Phone calls: call a shop or an office to ask, tell, or check for free slots
• On site: go to a place, check it and take photos
• Shopping: buy something on request (how the price is paid is agreed between the two sides)
• Review: a person checks what AI made
• Try-outs: use an app, a service or a shop and give feedback or survey answers (posting public reviews cannot be asked for)
• Data and translation: data creation (collecting photos, labelling, transcription), annotation, translation
• Read the categories with list_categories.

How to ask, by kind of task: Making phone calls, On-site checks and photos, Shopping, Checking AI-made work, Try it and write feedback

Steps (finished inside the AI chat)

1. Build (start_order): pass what should be done and it returns, as questions, the items still missing for that type of task. The agent asks the user those questions one at a time.
2. Show the person (preview_order): once complete, it returns the order sheet (what to do, deadline, budget cap, first price offered to workers, the most to be paid, done-when conditions). The agent shows it to the user exactly as returned.
3. Place (place_order): only after the user clearly says yes. Urgent tasks (within 10 minutes or 1 hour) reach available workers at once and the first to take it gets it. For an order decided by choosing, show the list from list_applicants to the user after applications have been collected, and decide the one person they clearly chose with choose_applicant.

4. Status (get_dispatch_status): being offered, price going up, taken, or returned.
5. Check the result (review_result): the person using the agent checks the returned result (text and photos) and their answer (checked, or redo) is recorded. Recording the check does not complete the deal: it moves on to completion when a person presses "OK" on the site (complete_url in the answer) or automatically after a fixed time (24 hours after the result came back).

How to connect

MCP URL: https://latentdatabank.com/mcp (Streamable HTTP, JSON-RPC 2.0). For English descriptions, errors and order sheets, connect to https://latentdatabank.com/mcp?lang=en or send Accept-Language: en.

Settings for each client (Claude Code, Claude connectors, Claude Desktop, Cursor, ChatGPT, Gemini CLI, REST) are on How to connect.

Connect by logging in (Claude, ChatGPT)

1. Add the MCP URL https://latentdatabank.com/mcp as a connector in Claude or in ChatGPT.
2. When the connect screen opens, log in with a company account and press Continue (sign up first if needed; it is free).
3. You can see the connected apps, and disconnect them, on the "AI agent keys" screen (agentkey).

Calling /mcp with neither a key nor a login returns 401 with WWW-Authenticate (Claude and ChatGPT then open the login screen). Existing keys (ldbk_…) still work as before.

What works without a key (the REST entry point for programs)

list_categories, estimate_job, start_order, preview_order, add_candidates need no key and no account over the REST entry point for programs (https://latentdatabank.com/api/agent/…): look, get a guide price, build the order sheet. An order sheet built without a key is deleted after 24 hours. Calls without a key are limited per device to 20 a minute, 300 a day, and 20 new order sheets a day.

MCP (/mcp) needs a login or a key (once connected, the tool list shows every tool). Over REST without a key, the entry points that place orders or touch money return key_required.

With add_candidates you only pass candidates found on the web to an order sheet that is not yet placed. Passing them sends nothing to anyone; candidates are notified only when a registered user places the order.

When you reach preview_order over REST without a key, place_requires in the result holds two URLs for placing the order. site_url: the user signs up (free) or logs in and places that same order on the site. register_url: a person at the company signs up and creates a key. After that, pass the same order_id with the key and the order carries over.

Claude Code, connecting by logging in (the login screen opens the first time you use it):

claude mcp add --transport http latentdatabank https://latentdatabank.com/mcp?lang=en

Create a key (to place, follow and review)

For clients that cannot log in, or from your own programs, keys (ldbk_…) work as before.

1. Log in with a company account (or sign up).
2. In the top-right menu open "AI agent keys" (agentkey), choose a name and a type (live or test) and create the key.
3. The full key is shown only once, right after you create it. Copy it into the settings below. Up to 3 keys per company; you can stop or remake a key on the same screen.
4. To use place_order, a person at the company turns on "confirmed in the LLM screen" on the automation screen (off by default; they set a cap per task and the categories allowed).

Send the key as Authorization: Bearer KEY.

Claude Code

claude mcp add --transport http latentdatabank https://latentdatabank.com/mcp?lang=en --header "Authorization: Bearer YOUR_KEY"

Claude Desktop (claude_desktop_config.json)

{
  "mcpServers": {
    "latentdatabank": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://latentdatabank.com/mcp?lang=en", "--header", "Authorization:${LDB_AUTH}"],
      "env": { "LDB_AUTH": "Bearer YOUR_KEY" }
    }
  }
}

ChatGPT (OpenAI API, Responses)

{
  "model": "your-model",
  "tools": [{
    "type": "mcp",
    "server_label": "latentdatabank",
    "server_url": "https://latentdatabank.com/mcp?lang=en",
    "headers": { "Authorization": "Bearer YOUR_KEY" },
    "require_approval": "never"
  }],
  "input": "Please have someone call the hair salon by the station and ask for free slots on Saturday afternoon. My cap is 3000 yen"
}

REST (from other tools)

curl -X POST -H "Content-Type: application/json" \
  -d '{"request":"Call the shop and ask for free slots today","category":"call","fields":{"due_within":"1h","budget_max_yen":3000}}' \
  "https://latentdatabank.com/api/agent/orders?lang=en"

curl -H "Authorization: Bearer YOUR_KEY" "https://latentdatabank.com/api/agent/workers?category=dev&limit=5"

Entry points: GET https://latentdatabank.com/api/agent/categories POST estimates POST orders GET orders/{id} POST orders/{id}/preview (no key needed up to here) POST orders/{id}/place raise_limit cancel repeat series_cap series_stop GET invoices GET dispatch/{id} POST dispatch/{id}/review workers jobs job_drafts invitations offers. Answers are always JSON (success carries mode; failure is {"error":{"code","message","hint"}}).

Prices and caps

• The requester decides the price (the budget cap). The agent works out a guide price from the minutes the task takes and how hard it is, and asks: "This task takes about N minutes, so the guide price is JPY X. Is that OK?" If the cap is too low, it explains: "For this task it takes about N minutes, so it needs at least JPY Y." The first price offered to workers comes from the market for past requests and never exceeds the budget cap.
• A guide price is available from estimate_job (a guide only; the range is widened where there are few past requests). Tasks asked through the agent interface start from JPY 100.
• If nobody takes the task, the price goes up at fixed intervals (every 1 minute for 10-minute tasks, 5 minutes for 1-hour tasks, 30 minutes for today or a date) until it reaches the budget cap. The user pays the price at the moment a worker takes the task.
• If nobody takes it even at the cap, we stop offering it and tell you. The user chooses to raise the cap and continue (raise_order_limit, which confirms the new cap and the new steps once more) or to stop (cancel_order).
• AI tools do not move money (they do not place card holds either). We sit in between, hold the payment and pass it to the person who did the work. There are two ways: bank transfer (a person at the company transfers from their own bank; place_order returns the amount, account and due date) or card (place_order does not place the order but returns a site page URL; the card hold is placed when a person presses "Place this order" there).
• For a card order decided by choosing, choose_applicant keeps it on card. payment.pay_url in the answer is the site page where a person presses "Pay by card"; the payment is made only then. The due date (payment.pay_by) and the automatic cancellation 7 days after it are the same as for bank transfer.
• A deal is completed and the money passed to the worker only when a person presses "Complete" (OK) on the site, or by the fixed time rules (24 hours after the result came back; 7 days after delivery).

Where a person must confirm

• Placing: only after a clear "yes" from the person who saw the order sheet (confirm_token and human_confirmed: true). The token lasts 30 minutes, works once, and stops working if the sheet changes.
• Raising the cap or stopping: both only after asking the user. The agent cannot choose a cap itself.
• Checking the result or asking for a redo: only after the answer of the person who saw the result (human_confirmed: true). A check recorded in the AI chat is only a record; completion and payment happen when a person presses the button on the site or by the time rules.
• Sign-up, login, keys and switches: signing up, logging in and pressing Continue on the connect screen, creating a key and turning on switches such as "confirmed in the LLM screen" are done by a person.
• Phone numbers, street and room numbers and door codes go into the fields shown only to the person who takes the task; they never appear in the public request or the work list.

What it cannot do

• Emergencies that concern life or safety (call the emergency services).
• Anything against the law, deceiving the other party, or nuisance sales calls.
• Procedures that only the person themselves may do (where the counter does not allow a proxy).
• Requests that book a person by the hour or direct how the work is done step by step (every task is contract work defined by a result and a deadline).
• A promise that a task will always be taken. We have only just started and there are not many workers yet; if nobody takes it, we ask as described under "Prices and caps".
• Returning a worker's contact details or legal name. Only public profile fields are returned.

Tools

Listing details for MCP directories

Test keys

Choose "test" when you create a key and it returns sample data instead of real workers (mode is test). Drafts made with a test key are kept apart from real ones and cannot be posted or sent (deleted after 30 days). place_order does not really place anything either. Use it to check your connection.

Limits

• Up to 60 calls a minute per key (429 beyond that). REST without a key: 20 a minute and 300 a day per device.
• A wrong or stopped key gets 401. Keys are stored only as hashes. Calls without a key are counted without storing the IP address itself.
• The key screen shows how many times each key was used and when it was last used.