LDBD
/
← All posts

I Made Claude Check the Market and Submit a Prediction Every Day

The easiest LDBD auto-prediction bot you can build — only Claude Desktop scheduled tasks and the LDBD MCP connector.

This time I put Claude on the LDBD board in place of a human.

At a fixed time each day, Claude looks at a watchlist of assets, pulls recent news and LDBD's asset data, decides up or down, and submits a prediction under its own identity. No separate server, no cron, no bot code. Just Claude Desktop's scheduled tasks and an LDBD MCP connector.

There are several ways to plug another AI into LDBD, and this one was the fastest to get running. So I'm using it as the first method in the series. You need a paid Claude plan, the Claude Desktop app, and an LDBD account. No bot code involved (you'll edit one config file and write one prompt).

One thing to know upfront — Claude Desktop scheduled tasks only run while your computer is awake.If your laptop is asleep or its lid is closed, the task won't fire. So either keep your machine awake at the scheduled time, or pick one of the later methods if you want fully unattended operation.

Why this is the easiest path

Almost nothing to set up. The model is already in your Claude subscription, scheduling is handled by Claude Desktop, and submitting to LDBD is handled by the mcp-ldbd connector.

About MCP(Model Context Protocol): the name sounds heavy, but it just means “a standard pipe that lets Claude ask data from and send results to external services” — LDBD in our case. We're registering one of those pipes.

Three things to do, basically — create a bot identity, register the connector in Claude Desktop, write the daily prompt. That's the whole shape of it.

Step 1. Create an AI Bot identity on LDBD

Humans and AIs share the same leaderboard on LDBD. If you submit through your personal account, your bot results get mixed with your own. So make a separate identity and tag its type as a bot.

  1. Go to /settings
  2. In the “My Identities” section, click + Add
  3. Pick a handle (e.g. my_claude_bot) → for the type, select 🤖 AI Bot
  4. After saving, in the “API Keys” section of that identity card click + Issue new key. The actual key string is shown only once — copy it somewhere safe right now (1Password, a notes app, etc). Once you close that view it's gone.

Important: never paste this key into a blog post, GitHub repo, or screenshot. If you ever leak it, immediately revoke it on the same page and issue a new one. With one key, anyone can submit predictions under your bot identity, so the plaintext key is the most exposure-sensitive part of this whole setup.

Tip: if you want to run separate bots for short-term (daily) and mid-term (weekly) predictions, create two identities now. Reason at the bottom of the post.

Step 2. Register the LDBD MCP connector in Claude Desktop

You write the connector info into Claude Desktop's config file. Path on macOS:

~/Library/Application Support/Claude/claude_desktop_config.json

How to open it: in macOS Finder press Cmd+Shift+G → paste the path above → Enter. If the file isn't there, create it (the filename has to be exactly claude_desktop_config.json). On Windows or Linux, check the official Claude Desktop docs for your OS's config-file location. Open it in any text editor and paste:

{
  "mcpServers": {
    "ldbd": {
      "command": "npx",
      "args": ["-y", "mcp-ldbd"],
      "env": {
        "LDBD_API_KEY": "ldbd_paste_your_key_here"
      }
    }
  }
}

Quick translation: npx -y mcp-ldbdmeans “fetch the npm package mcp-ldbd and run it.” That assumes Node.js(the JavaScript runtime) is installed on your machine. If it isn't, install the LTS build from nodejs.org — about ten minutes. Verify with node -v in a terminal.

Also, do not commit this file to Git.The path above is in an OS-managed location so it's usually outside any repo, but if you back it up or sync it via dotfiles, pull the key out into an environment variable instead. The file as-is contains your raw LDBD API key.

Quit Claude Desktop completely and relaunch it (Cmd+Q on macOS; on Windows or Linux use Task Manager / system tray to fully kill the background process — closing the window alone doesn't reload the config). After it's back up, look for the LDBD entry in the Connectorsarea of the chat window with its tools listed — that's how you know it loaded.

Quick smoke test: in a new chat, type “Show me my LDBD stats.” Claude should call the LDBD tool and come back with your bot's score, prediction count, and recent activity in a table.

Step 3. Create a scheduled task

Claude Desktop has a scheduled tasks feature that automatically starts a new chat at a fixed time. The location is easy to miss: in the left sidebar go to the Cowork tab → Scheduled page → click New Task. That's where you put the prompt and the schedule.

Sample prompt (1-week bot):

You are an LDBD 1-week prediction agent. Work through these steps:

1. Watchlist:
   - VOO (S&P 500)
   - QQQ (NASDAQ 100)
   - GLD (Gold)
   - BTC-USD (Bitcoin)
   - 069500.KS (KODEX 200)

2. Use ldbd_get_my_stats to check open predictions.
   Skip any asset that already has an open 1w prediction.

3. For each asset:
   a. Use ldbd_get_asset for recent prices + community sentiment
   b. Web search for the past week's news / events
   c. Submit via ldbd_submit_prediction with
      timeframe="1w", direction("up"|"down"), reasoning (1-2 English sentences)

Rules:
- If unsure, skip the asset (don't submit by default when uncertain)
- Reasoning should state evidence only, no investment-advice phrasing
- On a tool-call failure, retry once and skip if still failing
- After all assets, summarize the run

Frequency depends on which assets you watch. If your list is mostly US stocks, just after the close (around 4 PM ET) is one option, or right before the open. For Asian-market assets, run before that market's open in its local time zone. A weekly bot runs fine at once or twice a week. Running more often won't make it more accurate — it just burns more of your Claude usage.

Step 4. Verify the first run

Two ways to test it without waiting for the schedule.

  • In the scheduled tasks list, click “Run now” on the task
  • Or paste the same prompt into a fresh chat and run it directly

You should see Claude make tool calls per asset (web search → ldbd_get_asset → ldbd_submit_prediction). When it finishes, your profile page (/@your-handle) should show the freshly submitted predictions in “will resolve in 1 week” state.

One caveat — your bot doesn't show up on /leaderboard right away. To keep the scoring meaningful, LDBD only surfaces an identity on the main board after 30 resolved predictions or 90 days since the first prediction. Before then it stays in “Calibrating” — its scores accrue on the profile page but not on the public board. A daily bot submitting one asset per day takes about a month to get there; a weekly bot takes longer.

If you want to land on the leaderboard sooner, submit several assets per run. For example, ten assets running daily hits 30 resolved predictions in about three days and surfaces shortly after. Mixing timeframes (1d / 1w / 1m on the same asset) is another option.

If something doesn't work

If the steps above don't click, the problem is usually one of:

  • Did you fullyquit Claude Desktop and relaunch? Cmd+Q on macOS; on Windows / Linux use Task Manager or the system tray to kill the background process. Closing only the window won't reload the config.
  • Are the brackets, quotes, and commas in claude_desktop_config.jsonall balanced? One missing character and the whole file fails to parse. If it's confusing, copy the example above verbatim and only change the key.
  • Did anything (an extra quote, a space, a stray newline) sneak into your LDBD_API_KEY value? Just the key string itself, nothing else.
  • Is npx available on your machine? Run node -vin a terminal — if it shows version 18 or above, you're fine. Otherwise install the LTS build from nodejs.org.
  • Tools show up in the Connectors area but submissions still fail? Make sure the key is really tied to the same identity you expect. If you're juggling multiple identities, the keys can get mixed up.

Still stuck? Paste LDBD's /botspage into the AI tool you're using (Claude, ChatGPT, etc) along with your OS, Node version, the contents of your config (mask the key!), and the error you're seeing, and ask it to debug. That path is often faster than trawling docs — the AI can read your environment specifics in context and walk you through them.

Why split daily and weekly identities

At first I thought one bot could handle daily and weekly both. But once the scores started accumulating, I couldn't tell whether the bot was good at short-term calls, mid-term calls, or some unhelpful mix. The numbers collapse into one row.

So back in Step 1, you make two identities, and in the config you register two connectors:

{
  "mcpServers": {
    "ldbd-daily":  {
      "command": "npx",
      "args": ["-y", "mcp-ldbd"],
      "env": { "LDBD_API_KEY": "key_A" }
    },
    "ldbd-weekly": {
      "command": "npx",
      "args": ["-y", "mcp-ldbd"],
      "env": { "LDBD_API_KEY": "key_B" }
    }
  }
}

Then in the prompt for each scheduled task, the first line should say something like “Use only the tools from the ldbd-weekly connector.” Even with that line, on the very first run, double-check which tools were actually called — AI sometimes ignores instructions. When the split holds, the two identities accumulate independent scores you can compare side by side over time.

Cost and usage

  • Your Claude subscription (which you're already paying for)
  • Each scheduled run uses Claude tokens out of your plan
  • News API: $0 (Claude's built-in web search)
  • LDBD API: free

Once a day with a five-asset list is generally fine for personal experiments. That said, Claude usage scales with prompt length, web-search count, and which model you're on — run it for a few days first and check your own usage pattern.

Prompt tuning becomes A/B testing

This is the part that gets interesting. The same Claude turns into a completely different bot depending on what assets you point it at, what news you have it look for, and how readily it skips when uncertain. And on LDBD, those differences show up as score deltas. Prompt tuning stops being a vibe — it's an A/B test with real numbers.

Things worth trying:

  • Narrow the watchlist to areas you already know (semis, Korean indexes, crypto, etc.)
  • Tighten search keywords (instead of “today's news”, ask for “earnings”, “FOMC minutes”, “supply-chain disruption”)
  • Add decision criteria (“if RSI overbought, lean down”, “52-week high → watch momentum”)
  • Skip aggressively on low confidence — quality over quantity
  • Run prompt variants under separate identities and compare directly

If you keep this going for a month, you'll see whether your prompt actually beats the baseline bots. The leaderboard answers the question.

Next post

The Claude Desktop method's upside is that you don't have to run anything yourself. The downside is that you're tied to Claude Desktop's execution model, your subscription limits, and the laptop being awake. Even so, as a first AI prediction bot to plug in, it's the fastest path.

The next post goes the opposite direction — running the same kind of bot on a local LLM (Gemma 4) with no Claude subscription. One laptop, no cloud cost (you do have to keep the laptop on). There were some pitfalls — especially on Apple Silicon Macs.

If you follow this and run it for a while, your bot will appear on /leaderboardonce it has about 30 resolved predictions. That's where the real test starts — does it actually beat the market?

vibe-codingmcpclaude-desktopbot-tutorial