# Agents

> Connect Claude Code, Claude Desktop, Cursor or your own agent to the same evidence over MCP, with your own key.

Everything in the console is also available as tools over the
[Model Context Protocol](https://modelcontextprotocol.io), so an assistant can
answer "what happened to coins like this one?" with measured numbers instead
of a guess.

The tools read evidence and manage your own exit coach and alerts. **They do
not trade.** An agent connected to QuantBase cannot buy, sell, move funds or
sign anything for you.

## 1. Get a key

In the [console](https://app.quantbase.live), open
**Settings → API keys & MCP** and choose **New key**. Copy it straight away:
it is shown once and stored only as a hash. Keys start with `qmcp_`.
**Revoke** retires a key on the next request.

The key carries your plan. Free keys can call the free tools below; the Meme
plan unlocks the rest.

## 2. Connect

The server is `https://mcp.quantbase.live/mcp`, streamable HTTP, with the key
as a bearer token.

### Claude Code

```bash
claude mcp add --transport http quantbase \
  https://mcp.quantbase.live/mcp \
  --header "Authorization: Bearer $QUANTBASE_KEY"
```

### Claude Desktop, or any `mcpServers` config

```json
{
  "mcpServers": {
    "quantbase": {
      "type": "http",
      "url": "https://mcp.quantbase.live/mcp",
      "headers": { "Authorization": "Bearer YOUR_KEY" }
    }
  }
}
```

### Cursor: `.cursor/mcp.json`

```json
{
  "mcpServers": {
    "quantbase": {
      "url": "https://mcp.quantbase.live/mcp",
      "headers": { "Authorization": "Bearer YOUR_KEY" }
    }
  }
}
```

The transport is stateless, so a server update does not drop your
connection, but a connected client is not told about new tools. Reconnect to
refresh the list.

For the agent's system prompt, load
[quantbase.live/agents](https://quantbase.live/agents/): it tells the agent
what the tools mean and what it must not claim.

## 3. The meme tools

### Free with any key

| Tool | What it returns |
|---|---|
| `meme_status` | What the launch record holds: launches, graduation rate, price checks, creators |
| `meme_base_rates` | The 14-day base rates: organic graduation, up at 1 h and 24 h, pool alive at 1 h |
| `meme_graduations` | The newest graduations, with the curve at 2, 10, 30 and 120 minutes and the creator's record |
| `meme_features` | Everything knowable about one launch at 2, 5, 10, 30 or 120 minutes after it, and nothing later |
| `meme_dataset` | Rows of point-in-time features with outcome labels, up to 500 a call |
| `meme_metadata` | A launch's socials and website, and its name, ticker and image matches |
| `meme_cluster` | One funding cluster: its wallets, launches, graduations and rug proxy |
| `meme_attention` | Paid DexScreener boosts seen, and whether they came before or after graduation |
| `account_status`, `plans` | Your plan and quotas; what each plan includes |

Lens evidence for one coin is a plain HTTP call, no key needed:
`GET https://quantbase.live/lens/<mint>`. See [Lens](/docs/lens/#the-api).

### Meme, Pro and Team plans

**Feed and evidence**

| Tool | What it does |
|---|---|
| `dex_ideas` | The graduate feed: coins worth $150K+ in a funded pool at two minutes, with kind, flags, block reason and their group's track record. `hours` from 1 to 48 |
| `dex_scorecard` | How those coins did: hit 2x, under entry at 6 h, rugged, by kind, coin size and pool size. Up to 30 days |
| `dex_coin` | `action="candles"`: OHLCV at 1m, 5m or 15m. `action="like_this"`: how coins of the same kind, size and pool did at 1 h, 6 h and 24 h |
| `dex_watchlist` | Coins you are eyeing but do not hold, with price, 24 h change, value and pool. Up to 100 |

**Exit coach**

| Tool | What it does |
|---|---|
| `dex_watch` | `add` a coin you bought; `list` open coins with every rule's state; `sold_half` / `sold_all` record a sale you made; `history` shows closed coins and how much of each peak you kept; `close` stops watching |
| `dex_rules` | `get` or `set` your exit-rule defaults, with their bounds. Applies to coins added afterwards |
| `dex_daily` | Your day: alerts, exits, peak kept, P&L where you gave amounts, what is open |

**Alerts**

| Tool | What it does |
|---|---|
| `alerts_inbox` | Your alerts, newest first; mark them read |
| `alerts_telegram` | Link status, a one-hour link code, unlink, and which alert kinds go to Telegram |

**Rule lab**

| Tool | What it does |
|---|---|
| `meme_rule_explore` | Score up to 25 launch rules on the discovery half of graduates. Free of trials, records nothing |
| `meme_rule_test` | Judge one rule on the holdout half, after fees and slippage, deflated for every rule you have tried. Counts as a trial |
| `meme_rule_trials` | Your rule trials so far |

A rule that passes the referee is a measured result on past data, not a
prediction.

## Example questions

- "What graduated in the last hour, and which of them did the creator buy out?"
- "For mint `...pump`, how did coins of the same kind and size do six hours later?"
- "I bought `...pump` at $0.000182 with $25. Watch the exit."
- "Which of my coaches is closest to a rule firing?"
- "How much of each peak did I keep this month?"
- "Who funded this creator, and how did their other launches end?"

A good agent answers with the sample size and the date range, and says when a
sample is too small to mean anything.

## What agents cannot do

- **No trading.** No tool places an order for a customer. The coach records
  sales you report; it never makes one.
- **No money actions.** Live trading, live meme runs, the operator's DEX desk
  and wallet funding run on the operator's own credentials. The server
  refuses them from every other account and from every browser session, so
  they are not customer features even when they appear in `tools/list`.
- **No advice.** Nothing the tools return is a buy or sell recommendation.

## The research server

The same endpoint serves the backtesting tools of the
[research server](/docs/research-overview/): splits, backtests, the trial
ledger and the statistics. The full schema of every tool is in the
[tool reference](/docs/tools/).