Indexera Finance Data

Use this skill when the user asks for financial data lookup, especially around securities, funds, indices,行情,财务指标、估值、持仓、分红、公告期数据、时间序列查询。

Goal

Turn a natural-language finance question into a safe MCP workflow:

1. normalize the instrument with resolve_instrument
2. discover the right dataset with search_table
3. fetch only the necessary fields with query_data
4. answer in plain Chinese with the selected instrument, table, fields, and scope made explicit

Bootstrap

Before the first real MCP request, check whether the local token file exists.

Resolve paths relative to the current skill root, not a hard-coded install location:

• .token.local.env

If the file does not exist, the agent must proactively ask the user to provide a token, then write it locally in this format:

• INDEXERA_FINANCE_MCP_BASE_URL=https://mcp.indexera.cn
• INDEXERA_FINANCE_MCP_ENDPOINT=https://mcp.indexera.cn/mcp
• INDEXERA_FINANCE_TOKEN

Every MCP request must send Authorization: Bearer from that file.

Treat that file as a local secret. Do not print the full token or commit the env file.

Hard Rules

1. Always resolve the instrument first

Even if the user provides something that looks like an exact code, still call resolve_instrument first.

Reason: user-entered code formats vary; the skill must normalize to the system canonical_code before table search and data query.

2. Prefer minimal, verifiable queries

• query only the columns needed to answer the question
• use the narrowest viable date/time window
• set a reasonable limit
• do not fetch wide tables unless needed

3. Let metadata drive decisions

search_table is now a pure metadata search over rebuilt metadata_table.search_text.

That search text is generated inside MCP from table fields + column names + column comments + name_alias.

Do not assume any runtime synonym expansion happens at query time.

search_table returns table description, column matrix, and sample rows. Use those to decide:

• whether the table really matches the user question
• which columns are likely valid in select
• which fields look like instrument code / trade date / report date / metric columns

4. Explain assumptions

If the skill auto-picks one candidate instrument or one table among several possibilities, say so briefly in the final answer.

5. Prefer fixed CLI over generated Python

When integrating this skill into OpenClaw or another agent runtime, prefer the fixed stdin/stdout runner under the current skill root:

• scripts/indexera_finance_stdio.py

Do not generate one-off Python code for each MCP request if the fixed runner is available.

Primary contract:

• input: one JSON object from stdin
• output: one JSON object to stdout
• success shape: {\"ok\": true, \"data\": ...}
• failure shape: {\"ok\": false, \"error_code\": ..., \"message\": ...}

Available MCP Tools

resolve_instrument

Input:

• query — name, alias, code fragment, ticker-like text, etc.

Output:

• instruments[]
• instrument_type
• display_name
• canonical_code
• exchange
• market

Use it to normalize every security/fund/index/entity reference before any downstream step.

search_table

Input:

• question or query
• optional limit

Output:

• tables[]
• schema_name
• table_name
• table_title
• table_desc
• columns
• sample_rows

Use it to find the best table for the requested metric, frequency, and time range.

query_data

Input:

• schema_name
• table_name
• select (required)
• filters (optional)
• order_by (optional)
• limit (optional)

Use it only after the table has been selected from metadata evidence.

Execution Workflow

Step 1: parse the user question

Extract five things from the request:

1. instrument hint
• e.g. 贵州茅台, 600519, 沪深300, 纳指ETF
2. data theme
• 行情 / 估值 / 财报 / 指数 / 基金持仓 / 分红 / 复权 / K线 / 资金流
3. metrics
• 收盘价 / 成交额 / PE / 营收 / 净利润 / 仓位 / 持仓市值
4. time scope
• 最新 / 最近N天 / 某个季度 / 某个年份 / 时间序列
5. result shape
• 单点值 / 时间序列 / 排名 / 对比

Step 2: resolve the instrument

Always call resolve_instrument.

Selection preference when multiple candidates exist:

1. exact code match after normalization
2. exact display-name match
3. exchange/market consistent with user wording
4. most obvious mainstream listing / benchmark candidate

If ambiguity remains material, ask the user; otherwise continue with the best candidate and say which one was chosen.

Step 3: search the table

Build a compact search phrase combining:

• normalized instrument or instrument type
• data theme
• requested metrics
• time/frequency cues

Good examples:

• A股 日线 收盘价 成交额
• 股票 财务 营收 净利润 季度
• 基金 持仓 季报
• 指数 成分 权重

Run search_table with a modest limit first.

Step 4: rank table candidates

Prefer tables whose metadata best matches all of the following:

1. title/description clearly matches the requested domain
2. columns include the expected instrument key and requested metrics
3. sample rows show the right grain (daily / quarterly / annual / holdings / constituent rows)
4. schema is appropriate (typically tushare for current phase)
5. access looks usable for the current use case

Step 5: construct the query

When building query_data:

• include only essential select columns
• prefer canonical instrument code as the main identifier filter
• add date/report filters when the question implies a period
• sort newest-first for “latest” queries
• use a small limit unless the user explicitly asks for a long series

Common select shape:

• identifier column
• date/report column
• requested metric columns

Common order_by shape:

• report/trade date descending

Step 6: answer the user

Return:

• what instrument was resolved
• which table was used
• the result itself
• any limitation or assumption

For short answers, keep the explanation tight.

For analytical questions, summarize the data before listing rows.

Query Construction Heuristics

Latest single-value request

Example: “茅台最新收盘价”

• resolve instrument
• search daily price table
• query trade_date, code, close
• sort by trade_date desc
• limit=1

Time series request

Example: “宁德时代近20个交易日收盘价和成交额”

• resolve instrument
• search daily market table
• query date + requested metrics
• filter instrument code
• sort by trade_date desc or asc as needed
• limit=20

Financial statement request

Example: “比亚迪2024年净利润和营收”

• resolve instrument
• search income/profit statement tables
• prefer report-period fields in sample rows
• query report date / end date + revenue + net profit
• filter company code and year

Holdings / constituent request

Example: “某基金最新前十大持仓”

• resolve instrument
• search holdings/portfolio table
• query holding name, code, weight/value, report date
• sort by report date desc and weight desc if allowed

Failure Handling

If resolve_instrument returns no candidates

• tell the user the instrument could not be resolved
• ask for a clearer name, market, or ticker format

If search_table returns multiple plausible tables

Pick the one whose metadata best matches the requested metric and grain.

If still ambiguous, briefly mention the chosen table and why.

If query_data fails validation

Retry once with a more conservative request:

• reduce select
• remove speculative filters
• remove speculative sorts
• keep only obvious identifier/date constraints

Do not loop repeatedly.

If the data is unavailable

Say whether the failure came from:

• unresolved instrument
• no suitable table found
• query validation mismatch
• empty data returned for the requested period

Response Style

• Default to Chinese.
• Be factual and concise.
• Do not expose secrets.
• When useful, cite the resolved canonical_code, table name, and date scope.

Quick Examples

For CLI usage and OpenClaw integration examples, also see:

• references/cli.md

For a larger scenario library, also see:

• references/examples.md

Example 1

User: 贵州茅台最近5个交易日收盘价

High-level tool plan:

1. resolve_instrument(query="贵州茅台")
2. search_table(query="A股 日线 收盘价")
3. query_data(...) using canonical code, trade_date, close, limit=5

Example 2

User: 600519 近四个季度营收和净利润

High-level tool plan:

1. resolve_instrument(query="600519")
2. search_table(query="股票 财务 营收 净利润 季度")
3. query_data(...) using canonical code and quarterly/report columns

Example 3

User: 沪深300有哪些可查询表

High-level tool plan:

1. resolve_instrument(query="沪深300")
2. search_table(query="指数 成分 权重 行情")
3. summarize matching tables without forcing a data query