FOR AI AGENTS
Four channels that actually work today, from no sign-up required to calling the REST API directly from your own script. Everything here is shipped and verifiable — you can run every command yourself.
If you are an AI agent (or writing integration code for one), this page lists every channel Privora currently exposes, with the install steps and known gotchas for each. Not sure which one to use? See the comparison table below.
All four channels sit on the same underlying capability (GET /agent/skills and POST /agent/skills/execute) — they differ only in protocol shape and whether sign-up is required.
| Channel | Sign-up required? | How to install | Best for |
|---|---|---|---|
| Official ClawHub skill package | Privora account + Bearer Token | clawhub install privora-cn-quant |
Already using a ClawHub-compatible client (Claude, OpenClaw, etc.) |
| MCP Server (stdio) | Optional (anonymous mode without a token) | npm install -g the website tarball | Native MCP clients: Claude Code, Cursor, Windsurf, Cline, Codex CLI |
| Bearer Token + HTTP API | Privora account + Bearer Token | curl or any HTTP client, direct to the REST API | Your own script or agent framework, maximum flexibility |
| Anonymous read-only | Not required | No install — call the public endpoints directly | Trying the platform before deciding whether to register |
Different install steps, but all four ultimately hit the same GET /agent/skills + POST /agent/skills/execute dispatcher.
clawhub install privora-cn-quant (full capability set) or clawhub install privora-alert (real-time alerting only)LG_AGENT_TOKEN and LG_AGENT_BASE_URL=https://privora.cn as the plugin promptsBoth packages share the same underlying scripts — which one you install only changes how many skills show up in the catalog, not how you call them.
Wraps the same GET /agent/skills / POST /agent/skills/execute pair as native MCP tools, for stdio MCP clients such as Claude Code, Cursor, Windsurf, Cline, and Codex CLI.
npm install -g https://privora.cn/downloads/privora-mcp-server-0.1.2.tgzcurl -fsSLO https://privora.cn/downloads/privora-mcp-server-0.1.2.tgz
curl -fsSLO https://privora.cn/downloads/privora-mcp-server-0.1.2.tgz.sha256
sha256sum -c privora-mcp-server-0.1.2.tgz.sha256
npm install -g ./privora-mcp-server-0.1.2.tgz{
"mcpServers": {
"lg-agent": {
"command": "lg-agent-mcp-server",
"env": {
"LG_AGENT_BASE_URL": "https://privora.cn",
"LG_AGENT_TOKEN": "lgatk_your_token_here"
}
}
}
}LG_AGENT_BASE_URL defaults to http://localhost:3000 — a website install MUST set it explicitly to https://privora.cn, or every call fails to connect. LG_AGENT_TOKEN is optional: without it the server still lists all 12 tools, but what actually succeeds depends on the anonymous allow-list.For where each client (Claude Code / Cursor / Windsurf / Cline / Codex CLI) keeps this config, and the one-line add commands, see User Guide · AI Agent access. Note that Codex CLI uses TOML, not the JSON above.
Always 12 tools: 5 protocol-level meta-tools (list_skills / describe_skill / execute_skill / execute_confirm_required_skill / whoami) plus 7 first-class tools (dataasset_list, dataasset_get, dataasset_schema_get, dataasset_metadata_get, dataasset_data_get, dataasset_data_get_realtime, marketplace_item_list). Every tool carries an annotations disclosure hint (read-only / destructive) a host may use, and may ignore.
The most direct channel: once you have a token, call the REST endpoints with any HTTP client — no plugin required.
export LG_AGENT_TOKEN="lgatk_your_token_here"
export LG_AGENT_BASE_URL="https://privora.cn"
curl -s "$LG_AGENT_BASE_URL/agent/skills" \
-H "Authorization: Bearer $LG_AGENT_TOKEN"
curl -s -X POST "$LG_AGENT_BASE_URL/agent/skills/execute" \
-H "Authorization: Bearer $LG_AGENT_TOKEN" \
-H "Content-Type: application/json" \
--data-binary '{"skillId":"dataasset.list","params":{}}'X-Agent-Mode header — the Node proxy layer injects it automatically based on your Bearer token, and adding it yourself can get the request rejected.No registration, no token: call the public read-only endpoints directly. A fixed set of 10 read-only skills (asset listing, schema, marketplace items, etc.) is always allowed, rate-limited per IP across three buckets. Good for finding out what the platform can do before deciding whether to register.
curl -s "https://privora.cn/agent/skills"Every one of these is a verified, real failure — not a hypothetical.
Subscribing to a marketplace item gives your own team a new, cloned asset id — not the publisher's id shown on the marketplace page. Re-resolve the id from the subscribe response every time; never carry one over from a screenshot or chat log.
Omitting filter_op on dataasset.data.get defaults to a substring match (LIKE '%v%'), not exact. Pass filter_op=eq explicitly for an exact match.
The {id} in URL paths must be the numeric asset ID; passing an asset name returns a 500 with an unhelpful message.
The default "Read market & asset data" preset already includes metadata reads. Subscribing to marketplace items (to get a clonedAssetId) needs the separate "Read data & manage marketplace subscriptions" preset, ticked explicitly at token creation.
Paper-trading execution scopes are reserved for internal platform use and are auto-issued when you bind a strategy. They never appear as checkboxes on the token-create page — there is nothing to look for.
Since 1.0.55 the shell helpers accept the official domains and loopback by default; any other host must be declared in LG_AGENT_ALLOWED_HOSTS or the call is refused locally before the token is sent. This is a breaking change for existing self-hosted setups — the refusal names the host and the variable. (Domain choice itself is a non-issue: the helpers default LG_AGENT_BASE_URL to privora.cn, they pointed at lg-data.cc before 2026-09, and both serve this API path directly — measured, no redirect.)
GET /api/public/agent/capabilities is a trimmed public view that omits the entire paper-trading surface. Use GET /agent/skills for the real, scope-aware catalog.
The Node proxy layer injects X-Agent-Mode automatically from your Bearer token. Setting it manually can get the request rejected.
curl -d 'non-ascii text' silently corrupts non-ASCII bytes on Windows Git Bash. Use --data-binary @file instead.
paper.* scopes never appear on the create page — the platform auto-issues them when you bind a strategy.@privora/mcp-server has not been published to the npm registry. Install from the website tarball instead: npm install -g <tarball URL>.LG_AGENT_BASE_URL. It defaults to http://localhost:3000; a website-installed MCP Server or standalone script must set it explicitly to https://privora.cn.Not ready to register? Call GET /agent/skills anonymously first and see what comes back. When you are ready, create a token to unlock everything.