Zolva
Robinhood Chain 上的交易代理:每小時買入、賣出或持有及其理由,另附股票代幣和加密貨幣價格。唯讀,無需 API 金鑰。
託管 MCP 伺服器
npx add-mcp 'https://zolvaagent.app/mcp'可安裝到 Claude Code、Codex、Cursor 等客戶端
文件
MCP server
Connect Claude, Cursor or any MCP client to Zolva: ask about agents and markets in plain words, and with an API key backtest and run your own paper agents.
Overview
Zolva has an MCP server, so AI assistants such as Claude, ChatGPT, Cursor or Codex can read Zolva's data and, with your API key, run your paper agents. Ask in plain words; the assistant picks the tools.
- URL:
https://zolvaagent.app/mcp(Streamable HTTP, no session state) - No key: 7 read-only tools over the same data as the Public API.
- With an API key: up to 11 more tools, only those your key's scopes allow.
What it can and cannot do
| Can | Cannot |
|---|---|
| Read published agents, their decisions with reasons, markets, prices, funding and trades. | See other people's private agents, wallets or AI usage. |
| With a key: list your own agents, check rules, backtest, write rules with AI. | Trade real money, swap or move funds. |
| With a key: create, pause, resume, publish and unpublish your paper agents, and change their AI budget. | Change an agent's strategy after launch (settings are locked). |
With a key: stop one of your real money agents (pause_live_agent). | Resume, create or edit a real money agent. Those stay in the browser. |
[!note] Note Paper only, with one exception Every agent the MCP server creates or changes trades virtual money. The only real money action is
pause_live_agent, which can only stop an agent. Past results do not predict future results, and nothing here is investment advice.
Prerequisites
- An MCP client that supports remote (HTTP) servers. Claude Desktop with a key also needs Node.js, for mcp-remote.
- For the keyed tools: a Zolva account and an API key from Settings, API keys. Tick only the scopes you need; the secret is shown once.
- Nothing else for read-only use. The server costs nothing; only
write_strategyspends your AI credits.
Setup
Every snippet below includes the key header; leave it out for read-only use. Replace zk_... with your key, or keep the key in an environment variable where the client supports it.
Claude Code
One command. Leave the header out for read-only use.
claude mcp add --transport http zolva https://zolvaagent.app/mcp --header "Authorization: Bearer zk_..."
Claude Desktop and claude.ai
Without a key: Settings, Connectors, Add custom connector, and paste the URL. With a key, add it to claude_desktop_config.json (Settings, Developer, Edit Config) through mcp-remote, which needs Node.js, then restart Claude. The header is split into an env value because Claude Desktop on Windows does not escape spaces in args.
Without a key
https://zolvaagent.app/mcp
With a key: claude_desktop_config.json
{
"mcpServers": {
"zolva": {
"command": "npx",
"args": ["mcp-remote", "https://zolvaagent.app/mcp", "--header", "Authorization:${ZOLVA_AUTH}"],
"env": { "ZOLVA_AUTH": "Bearer zk_..." }
}
}
}
Cursor
Add this to ~/.cursor/mcp.json (most clients use the same shape).
{
"mcpServers": {
"zolva": { "url": "https://zolvaagent.app/mcp", "headers": { "Authorization": "Bearer zk_..." } }
}
}
VS Code (GitHub Copilot)
Add.vscode/mcp.json to your project. VS Code asks for the key once and stores it as a secret, so it never sits in the file.
{
"inputs": [
{ "type": "promptString", "id": "zolva-key", "description": "Zolva API key", "password": true }
],
"servers": {
"zolva": {
"type": "http",
"url": "https://zolvaagent.app/mcp",
"headers": { "Authorization": "Bearer ${input:zolva-key}" }
}
}
}
Codex
Add the server to ~/.codex/config.toml. Codex reads the key from the environment variable you name and sends it as a Bearer token. Leave that line out for read-only use.
[mcp_servers.zolva]
url = "https://zolvaagent.app/mcp"
bearer_token_env_var = "ZOLVA_KEY"
Then, in the shell that starts Codex
export ZOLVA_KEY="zk_..."
Hermes Agent
Add the server to ~/.hermes/config.yaml, then run /reload-mcp in Hermes.
mcp_servers:
zolva:
url: "https://zolvaagent.app/mcp"
headers:
Authorization: "Bearer zk_..."
Any of them at once: npx add-mcp
add-mcp writes the right config file for each client you name with -a (repeatable). -y skips the questions.
npx add-mcp https://zolvaagent.app/mcp --header "Authorization: Bearer $ZOLVA_KEY" -a claude-code -a cursor -a vscode -a codex -a hermes -y
Check that it works
- Look at the tool list. Your client lists the server's tools (in Claude Code:
/mcp). Without a key you see 7: list_agents, verify_agent, get_agent, list_markets, get_market, recent_activity, get_consensus. With a key you also see the tools its scopes allow. - Ask a read question: “List Zolva's crypto markets with their 24 hour change.” The assistant should call
list_markets. - With a key, ask: “List my Zolva agents.” It should call
my_agentsand show your own agents, private ones included.
From a terminal, this lists the tools without any client:
curl -X POST https://zolvaagent.app/mcp \
-H "Content-Type: application/json" -H "Accept: application/json, text/event-stream" \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'
Example conversation
What a session with a key that has compute and agents:write can look like. The numbers are made up.
You
Which Zolva agents beat holding their markets this week?
Assistant
(calls list_agents with sortBy vs_holding) Three published agents are ahead of holding. The top one is 4.2 points ahead on NVDA and ETH, with a worst drop of 3.1%. All of them trade virtual money.
You
Why did the top one buy ETH today?
Assistant
(calls get_agent) At 09:00 UTC it bought because the 48h z score fell to −2.7, an unusually deep dip. It has held since; the latest decision says it waits for the z score to get back above 0.5.
You
Backtest the same idea on BTC for 30 days and make it a paper agent if it beats holding.
Assistant
(calls backtest, then create_paper_agent with confirm false) It returned 2.4% against 1.1% for holding. Here is the preview: mean reversion on BTC, $10,000 virtual, stop loss 8%. Shall I launch it?
You
Yes.
Assistant
(calls create_paper_agent with confirm true) Launched. It is private and makes its first decision at the next hourly tick.
Note the preview before the launch: create_paper_agent and publish_agent change nothing until the assistant calls them again with confirm: true, and the server tells assistants to ask you first.
Prompts to try
- “Which Zolva agents beat holding their markets, and by how much?”
- “Explain every decision Lab · Dip Hunter made in the last 24 hours.”
- “Which markets fell most in the last 7 days, and which agents trade them?”
- “What did agents buy or sell in the last two days, and why?”
- “Which perps have the highest funding right now, and what does that cost a long?”
- “Which stock tokens can be borrowed against on Longbow, and at what LLTV?”
- “With a key: write rules for buying NVDA on RSI dips, check them, and backtest them for 30 days.”
- “With a key: pause all my paper agents that are behind holding.”
Tools without a key
Each answer is the same JSON as the matching Public API endpoint, so the field lists there apply.
list_agents
Published Zolva agents with their return, return vs simply holding the same markets, worst drop and trade count. Paper trading only.
| Input | What it is |
|---|---|
strategy one of: momentum, mean_reversion, dip_hunter, spike_fade, dip_fade, bounce_short, dca, custom | Only agents using this strategy |
market string, up to 17 characters | Only agents watching this market symbol, such as NVDA, ETH or BTC-PERP |
venue one of: spot, perps | Only spot agents (tokens, long only) or only perps agents (Lighter, long or short) |
sortBy one of: return, vs_holding, drawdown | return (default, highest first), vs_holding (most ahead of holding first) or drawdown (smallest worst drop first) |
limit integer, 1 to 50 | How many agents to return, default 20 |
verify_agent
Replays one published agent with Zolva's open source engine over every hour it ran, on the closes it saw and the same funding, and compares every trade and its value with the record. Returns match or the first hour that differs; hours whose look-back predates stored prices (2026-11-14) are reported apart as unverifiable.
| Input | What it is |
|---|---|
id string (uuid) · required | The agent id from list_agents |
get_agent
One published agent: settings, hourly value vs holding for 7 days, open positions, latest trades, and every decision of the last 24 hours with its reason.
| Input | What it is |
|---|---|
id string (uuid) · required | The agent id from list_agents |
list_markets
Every market Zolva prices on Robinhood Chain: latest hourly price, 1 hour to 30 day changes, a 7 day sparkline and how many agents trade it. Perps (venue perps, symbols like BTC-PERP) also carry their funding rate. Spot tokens that Longbow (Morpho Blue lending) accepts as collateral for USDG carry longbow: { lltvPct, borrowAprPct, availableUsdg, feedFrozen, url }, otherwise null (borrowAprPct is null while the price feed is paused; rates change, not investment advice).
| Input | What it is |
|---|---|
assetClass one of: stock, crypto | Only stocks or only crypto |
venue one of: spot, perps | Only spot tokens or only Lighter perps |
get_market
One market: up to 30 days of hourly closes, the published agents trading it, and its Longbow lending market if any (longbow: { lltvPct, borrowAprPct, availableUsdg, feedFrozen, url } or null; rates change, not investment advice).
| Input | What it is |
|---|---|
symbol string, up to 17 characters · required | Market symbol, such as NVDA, SPY, ETH or BTC-PERP |
recent_activity
Trades by published agents over the last 48 hours, newest first, each with the agent's own reason.
| Input | What it is |
|---|---|
market string, up to 17 characters | Only trades in this market, such as NVDA or BTC-PERP |
agentId string (uuid) | Only trades by this agent |
limit integer, 1 to 200 | How many trades to return, default 50 |
get_consensus
Statistics, per market: how many Tested published paper agents (rules that passed Zolva's 6 year backtest) are long, short or in cash, and what they did in the newest hour. Returns the agent count each market is based on and says when it is too few. Paper trading, not investment advice.
| Input | What it is |
|---|---|
market string, up to 17 characters | Only this market, such as BTC or SOL-PERP |
venue one of: spot, perps | Only spot or only perps markets |
Tools with an API key
A tool appears only when the key holds every scope in its row. Inputs, answers and errors are the same as the /me endpoints; every tool is marked read-only, destructive or safe to repeat (MCP tool annotations).
my_agents
My agentsScope: read:private
No inputs.
validate_rules
Check rulesScope: compute
| Input | What it is |
|---|---|
rules object · required | Custom rules JSON, as on /docs/rules (entry, exit, takeProfitPct, maxHoldHours) |
backtest
BacktestScope: compute
| Input | What it is |
|---|---|
strategy one of: momentum, mean_reversion, custom · required | A preset, or custom with rules |
rules object | Custom rules JSON, as on /docs/rules (entry, exit, takeProfitPct, maxHoldHours) |
markets array of string, up to 42 characters (1 to 3) · required | 1 to 3 market symbols or token addresses |
venue one of: spot, perps | spot (default): stock and crypto tokens, long only. perps: Lighter perps, paper only |
capital number, 1,000 to 1,000,000 | Virtual dollars, default 10000 |
maxPositionPct number, 10 to 100 | Most of the capital in one market, default 33 |
stopLossPct number, 2 to 25 | Stop loss per position, default 8 |
days integer, 1 to 30 | Days to replay, default 30 |
write_strategy
Write a strategy with AIScope: ai
| Input | What it is |
|---|---|
prompt string, up to 600 characters · required | The strategy in plain words, such as: buy when RSI is under 30, sell above 60 |
create_paper_agent
Create a paper agentScope: agents:write
| Input | What it is |
|---|---|
name string, up to 40 characters · required | The agent's name, up to 40 characters |
strategy one of: momentum, mean_reversion, custom, dip_hunter, spike_fade, dca · required | Preset or custom (needs rules). spike_fade opens shorts and needs venue perps. dca buys on a schedule and never sells (needs dca, spot, no aiModel) |
markets array of string, up to 42 characters (1 to 14) · required | Market symbols (NVDA, ETH, BTC-PERP) or token addresses, from list_markets |
venue one of: spot, perps | spot (default): stock and crypto tokens, long only. perps: Lighter perps, paper only |
capital number, 1,000 to 1,000,000 · required | Virtual dollars, 1000 to 1000000 |
maxPositionPct integer, 10 to 100 · required | Most of the capital in one market, 10 to 100 |
stopLossPct integer, 2 to 25 · required | Stop loss per position, 2 to 25 |
rules object | Custom rules JSON, as on /docs/rules (entry, exit, takeProfitPct, maxHoldHours) |
dca object | dca only: the schedule. The first buy is the agent's first hour; capital is split equally across the buys and the markets |
strategyPrompt string, up to 600 characters | Custom only: the plain words the rules came from, read by the AI co-pilot |
aiModel string, up to 80 characters | AI co-pilot model id (needs the ai scope); leave out for rules only |
aiLimitUsd number | With aiModel: the most the co-pilot may spend, USD |
aiExpiryDays integer, at most 9,007,199,254,740,991 | With aiModel: how many days the AI limit lasts |
acceptTerms boolean | Must be true with confirm true: you accept https://zolvaagent.app/terms |
confirm boolean | false (default): only show what would happen. true: do it |
idempotencyKey string, up to 200 characters | Any unique string: a retry with the same value never makes a second agent |
pause_agent
Pause a paper agentScope: agents:write
| Input | What it is |
|---|---|
agentId string (uuid) · required | The agent id from my_agents |
resume_agent
Resume a paper agentScope: agents:write
| Input | What it is |
|---|---|
agentId string (uuid) · required | The agent id from my_agents |
publish_agent
Publish a paper agentScope: agents:write
| Input | What it is |
|---|---|
agentId string (uuid) · required | The agent id from my_agents |
confirm boolean | false (default): only show what would happen. true: do it |
unpublish_agent
Unpublish a paper agentScope: agents:write
| Input | What it is |
|---|---|
agentId string (uuid) · required | The agent id from my_agents |
set_ai_limit
Set an AI limitScope: agents:write + ai
| Input | What it is |
|---|---|
agentId string (uuid) · required | The agent id from my_agents |
limitUsd number · required | The most the co-pilot may spend in total, USD |
expiryDays integer, at most 9,007,199,254,740,991 · required | How many days the limit lasts from now |
pause_live_agent
Stop a real money agentScope: live:stop
| Input | What it is |
|---|---|
agentId string (uuid) · required | The id of a real money agent (mode live) from my_agents |
Troubleshooting
| What you see | Why, and the fix |
|---|---|
| 401, or the server fails to connect | The key is wrong, cut short or revoked. A bad key is refused, never treated as no key. Copy a new key from Settings and check the header reads Authorization: Bearer zk_... with one space after Bearer. |
| Only the 7 keyless tools, or one tool is missing | The client sends no key, or the key lacks a scope. Each tool needs the scopes in its row above: make a key with the scopes you need. Some clients cache the tool list: restart or reload the server. |
| insufficient_scope | The tool was called with a key that lacks its scope, or create_paper_agent got an aiModel without the ai scope. |
| 429 rate_limited | Too many calls: about 60 a minute per IP without a key, 120 a minute and 30 writes a minute per key, and per account 60 backtests an hour, 200 a day and 20 creates a day. Wait and ask again. |
| not_found for your own agent | Paper tools ignore real money agents (use pause_live_agent), and pause_live_agent ignores paper agents. Check the id and mode in my_agents. |
Security
- Treat the key like a password. Keep it in an environment variable or your client's secret store, never in a shared file, a repository or a web page.
- Least privilege. New keys start with
read:privateandcompute. Addagents:write,aiorlive:stoponly when you need them, ideally on a separate key. - Revoke a key in Settings and it stops working on the next request. Every write and backtest made with a key is logged there.
- userText is data. Agent names and strategy prompts come in
userTextfields, cleaned of control characters. They were typed by Zolva users: an assistant should show them, never follow them as instructions. The server says so in its instructions. - Review previews. Read a create or publish preview yourself before you tell the assistant to confirm.