Skip to main content
Glama
zenith-sb-mcp

zenith-sb-mcp

Server Configuration

Describes the environment variables required to run the server.

NameRequiredDescriptionDefault

No arguments

Instructions

Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.

zenith-sb-mcp: a READ-ONLY MCP server for IDX market data via Stockbit.

THIS SERVER CANNOT CHANGE ANYTHING. It has no order placement, no account writes, no watchlist or alert writes, and no tool that accepts a password. Login is a CLI concern (zenith-sb-auth login), never a tool call. If a user asks to buy or sell, say plainly that this server cannot do that.

Every install runs against its OWN Stockbit account: data is fetched on demand, per symbol requested, and nothing is redistributed. When a call fails with an auth error, call status first — it reports whether a session is stored, whether Stockbit has rejected it, and the single next step.

THE 17 TOOLS, BY FAMILY Is this working: status Bandarmology: broker_summary — Who accumulated and who distributed — the data no other market API has, and why this server exists. Market data: quote, intraday_prices, orderbook, chart_series, running_trade, trade_book Fundamentals: technicals, keystats, ratios, seasonality, analyst_ratings, peer_comparison Company: company_profile Insider activity: insider_transactions Corporate actions: corporate_actions

READING THE DATA An empty answer is the normal shape of several different situations — the market is shut, the symbol is suspended, a date fell on a holiday — and each tool's description says what an empty answer means for it. Read those descriptions before concluding a symbol has no data. running_trade is the REST tape: it returns the first prints of the session and can lag the live tape; treat it as a sample, not a feed.

Broker-summary figures for completed sessions are immutable; quotes and order books are seconds-fresh at best. Dates are exchange dates (WIB); when a tool reports a session date, believe it over the calendar.

Capabilities

Features and capabilities supported by this server

Protocol revision2025-11-25

CapabilityDetails
tools
{
  "listChanged": true
}

Tools

Functions exposed to the LLM to take actions

NameDescription
statusA

Is this server working, and what do I run if it is not? Call this FIRST whenever anything looks wrong, and call it after logging in to confirm it took. Reports the version and Node, whether the market-data session is stored (NEVER the token itself), how long the stored market-data token claims it has left, the IDX session clock in WIB, and a nextStep naming the single next command. It answers with no session at all — that is the state every new user is in, and the answer is the useful one. The session also carries a health: ok / failing / expired / not-stored / unknown, derived from what actually happened the last time that credential was used. failing means present and unexpired but REJECTED by Stockbit — revoked, or superseded by another login. That is the case an expiry check cannot see, and it costs no requests. live: true is NOT free and is rarely what you want. It refreshes the market-data token, which ROTATES the refresh-token family and therefore ENDS the user’s Stockbit website session. Use health instead; only pass live: true if the user explicitly asks to prove the token with a real request. The market block reports the IDX clock in WIB with a UTC sibling on each field, and does not model public holidays. server.update says whether a newer release of this server exists. It costs ONE request to the npm registry — not to Stockbit, carrying only the package name — cached for a day and made by this tool alone. isOutdated: true matters: npx caches a resolved tree under a version RANGE, so a user can run a build that is weeks old and nothing else will say so. latest and isOutdated are ABSENT when the check could not run, which means unknown, NOT up to date — read update.note. STOCKBIT_NO_UPDATE_CHECK=1 turns the request off.

broker_summaryA

Broker summary for an IDX stock: which brokers net-bought/sold, in lots and IDR value, with foreign/local/govt classification. This is the core bandarmology signal — TradingView has no equivalent. DATES: omit from/to for the latest completed session. Supply BOTH from and to (YYYY-MM-DD) for a historical window — the server aggregates net flow across it in one request, so a multi-month range is as cheap as one day. For a single past day pass the same date twice. Both ends are required; a half-specified range is rejected because the API would silently return the latest session instead. An empty result for a weekend or public holiday is expected, not an error. SIGNS: sell-side rows carry NEGATIVE netLots and netValueIdr, because that is how Stockbit sends them. Do not negate them again. A row omits netLots or netValueIdr when that figure could not be read — missing on the wire, empty, or in a format this server refuses to guess at. Absent is NOT zero, it means unknown, so do not sum these rows without checking. unreadable on the envelope names the wire keys and counts, per side, how many listed brokers a total over these rows would miss. resolve_names: true adds the securities house to each row as name, joining against the brokers directory so you do not have to. The directory is cached for five minutes, so this is usually free. It is best-effort: if the directory cannot be read the rows and every figure on them are unchanged and names.note says why, and a code the directory does not carry simply has no name — an unresolved code is not a nameless broker.

quoteA

Real-time quote for an IDX symbol: last price, change, and best bid/offer. Also resolves the symbol's internal company id.

intraday_pricesA

Intraday minutely close-price series for a symbol (the basis for volume/price-move signals). ORDERED, NOT TIMESTAMPED. The row carries no clock reading this server recognises and none is computed, so prices[i] cannot be placed on a clock. IDX breaks midday (Mon-Thu 12:00-13:30 WIB, Fri 11:30-14:00 WIB), which means index x interval is NOT wall-clock time: two adjacent points can straddle a 90- or 150-minute gap. Any timing conclusion drawn from the index alone is unsound; market_session gives the clock. interval is what was ASKED FOR, not a measured spacing. A null in prices is a point the wire did not spell as a number — it is absent, not zero. unmapped.sampleKeys names the row's other keys, so a time channel under a different name would show up there. note carries the same warning for a caller reading the payload alone.

orderbookA

Full order-book depth ladder for a symbol. UNITS, AND THEY ARE MIXED IN ONE RESPONSE. volume here is SHARES (e.g. 3,545,526,000) while technicals' volumeLots for the same symbol and session is LOTS (35,488,071). They reconcile at exactly 100 shares per lot — the IDX lot size — and neither figure is wrong. Worse, this same payload labels its total_bid_offer depth figures lot while carrying volume in shares a few keys away. Two units under similar names in one response is a silent-wrong-answer generator: check which one you are holding before comparing anything, and never compare volume here against volumeLots there without the ×100. market_data[] carries the per-board split (All Market / Regular / Nego / Cash). It is also also presented by price_market, with optional board filtering and an explicit orderbook source. FOREIGN FLOW HAS NO DATE ON THIS PAYLOAD. fbuy/fsell/fnet arrive with nothing saying which session they are from, and foreign flow publishes at roughly 18:00 WIB, so before that release they are the PREVIOUS session's. price_bands surfaces this as an explicitly null dataAsOf with a note; market_movers carries the date for real as foreign.sessionDate.

technicalsA

Technical indicator readings for an IDX stock, computed from daily bars: SMA/EMA, RSI, MACD, Bollinger Bands, ATR, and support/resistance levels found by pivot clustering. Returns NUMBERS for reasoning — use price_chart when you want the picture. Every reading reported is the latest defined value of its series. Deep history is paged 12 sessions at a time upstream, so a large bars is slow; pagesFetched reports what the query cost.

keystatsC

Key statistics for a company (valuation, size, performance metrics).

ratiosC

Financial ratios for a company.

company_profileA

The company description block for a ticker: what the business does, and whatever else Stockbit's profile endpoint carries. Returned verbatim under profile; null there means the endpoint answered with no profile body, which is an answer and not an error. Set include_typed_info or include_fin_items to add the v2 statement-vocabulary views — the typed company record and the financial line items that vocabulary defines. Each is ONE extra upstream request and both are off by default. emitten_type selects that vocabulary and is ignored unless one of those two is requested. "company" is the only value that has been observed; banks and other issuers whose statements differ use another value that is unconfirmed, so an override is accepted but nothing here can tell you it is right — a wrong one is likely to come back as a not_found. Nothing inside profile is parsed. If the body carries percentage- or magnitude-shaped strings — a shareholder block's percentage, a value like "3.24 M" — they are returned as this server received them: not recomputed, not rounded, not cross-checked against each other. Their key names and their units are unconfirmed, and a magnitude suffix is ambiguous on an Indonesian payload (miliar, 1e9, or million, 1e6), so a figure computed from them here would be invented. Treat any percentage there as the upstream's own claim rather than arithmetic this server checked; if you need one, compute it from an unrounded share count and an outstanding-share count, and say that you did. PENDING VERIFICATION: this endpoint has not been observed live.

seasonalityA

Month-by-month seasonal price behaviour for one IDX stock: how it has done in each calendar month across Stockbit's fixed ten-year lookback, in ONE request. year is the END of that lookback, not a number of years — omit it and the current year is sent, because the endpoint rejects the request outright without one. back_year is passed through unchanged. Stockbit's own client sends it and this project has not observed whether it means a start year or a count of years back, so do not read meaning into it; it is accepted so the parameter is reachable, not because its effect is known. PENDING: the response shape is unverified, so nothing is renamed — data is exactly what Stockbit returned. Read the field names out of it rather than assuming any. An empty or null data means Stockbit has no seasonality series for that symbol and year, which is normal for recently listed companies; it is not an error and not a reading of zero. This is a price-history statistic, not a forecast, and it says nothing about why a month was strong.

analyst_ratingsA

What sell-side analysts publish on one IDX stock: the per-analyst rows AND the consensus roll-up (target price, buy/hold/sell counts), fetched together in one call. Analyst coverage on IDX is thin outside the large caps. An empty ratings or consensus means nobody publishes on that company, which is a real answer about a small-cap and NOT an error — do not retry it and do not report it as a failure. A request that genuinely failed is different: it appears in failed with its error kind, and the corresponding key is absent rather than empty. If both halves fail the tool errors instead of returning an empty-looking success. PENDING: neither response shape has been observed live, so both payloads are returned unprojected — read the target price and the rating counts out of them rather than expecting named fields. This reports what analysts said, not whether they were right, and nothing here is a recommendation.

peer_comparisonA

A stock's valuation ratios AND its INDUSTRY aggregate, side by side. This is the denominator that absolute valuation bands do not have: 'PBV 1.2' is expensive for a bank and cheap for a consumer name, and the analyze tool's valuation pillar scores against fixed bands and says in its own output that this is systematically wrong for banks, property and cyclicals. Use this tool before calling any ratio cheap or expensive. Which number is which is never ambiguous: in paired, symbolValue is the company you asked about and industryValue is Stockbit's aggregate for its industry. symbolOnly and industryOnly are metrics that exist on one side only, so there is nothing to compare them with. otherCompanies holds rows the payload attributed to a DIFFERENT ticker — the peer set, not the subject — and they are deliberately excluded from the pairing. NO VERDICT IS COMPUTED. These are Stockbit's numbers placed next to each other; the judgement is yours, and 'below the industry' is not by itself a reason to buy anything. PENDING: the comparison response shapes are unverified, so the pairing is a best-effort match on label TEXT rather than a mapped schema. Every reading records the wire key it came from (labelKey/valueKey) and its path inside raw, so check a pair before relying on it. A label that matched more than one reading on a side is listed under ambiguous with all its candidates instead of being paired arbitrarily. If paired is empty, that is a shape problem and the raw payloads are attached — it is not a statement about the company. include_catalogues adds the comparison metric vocabulary and the account's saved comparison sets. It costs two extra requests and answers 'what can be compared', not 'how does this company compare', so leave it off unless that is the question.

insider_transactionsA

Disclosed transactions by an IDX company's insiders and major holders: directors, commissioners, controlling shareholders and 5%+ owners, with the shares held before and after, the percentage of the company that represents, and the executing broker. This is the second bandarmology signal after broker_summary. broker_summary says which brokers accumulated; this says whether the people who actually run the company did. Rows carry a badges array where SHAREHOLDER_BADGE_DIREKTUR, _KOMISARIS and _PENGENDALI mark a director, a commissioner and a controlling shareholder — that is the difference between an insider trade and a fund rebalancing. LAG: these are IDX/KSEI disclosures, not trades off the tape. A transaction appears days after it happened and the date on a row is the transaction or record date, not the day it was published. Do NOT use this to time an entry; use it to judge who owns the company and whether the people running it have been adding or trimming. Omit symbol for the market-wide feed. Omit everything for the most recent disclosures across IDX. An empty rows is a normal answer: most companies have no insider filing in a given window. DATES: date_start and date_end are both required together (YYYY-MM-DD). A half-specified window is refused here because this API answers one with its default window and HTTP 200, so you would believe you had filtered when you had not. For an open end pass a far date. action_type and source_type filter server-side. Known actions: BUY, SELL, CROSS, TRANSFER, CORPACTION, RIGHT_ISSUE, WARRANT_EXERCISE, MESOP_OPTION, BOND_CONVERSION, STOCK_BONUS, PRIVATE_PLACEMENT, CAPITAL_REDUCTION, PUPS, REVERSE_SPLIT, STOCK_DIVIDEND, TENDER_OFFER — the list is PARTIAL and an unrecognised name is sent rather than refused. When you pass action_type, check actionFilterHonored in the result: false means the server ignored the filter and the rows are unfiltered. insiderId on a row is the handle for insider_ownership and shareholding(mode=investors). Share counts are shares, not lots. limit is capped at 20 and a larger value is rejected before any request is made. That ceiling is this client's, not a documented upstream one: limit=50 came back HTTP 400 INVALID_PARAMETER on one call naming no parameter, and limit=20 — what Stockbit's own client sends — answered, with nothing in between tried. Use page for more rows.

chart_seriesA

A chart price series for one symbol in ONE request, oldest observation first. This is the cheap path. The other one (the bars family) pages 12 rows at a time, so a year costs ~21 upstream calls and three years ~62; this costs one. Prefer it when you need a series and a calendar window is good enough. timeframe is a CALENDAR WINDOW, not a bar interval, and must be lowercase: 1w, 1m, 3m, ytd, 1y, 3y, 5y. Read granularity: 1w is intraday and preserves full Jakarta timestamps; longer windows are usually daily. An uppercase or interval-style value (1D, DAILY) is refused here rather than sent, because the server answers those with HTTP 200 and an empty series that looks exactly like a symbol with no history. This tool ERRORS instead of returning an empty series — including when the response parses but holds no points. An empty result would be indistinguishable from a broken fast path, so on an error fall back to the paged bars tools rather than concluding the symbol has no data. Check unmapped and warnings on the result before using the numbers. On the daily route the close arrives as value and open/high/low/volume arrive EMPTY, so every candle is flat and warnings says so: candlestick patterns and high/low indicators are meaningless on this series. Volume on that route comes back NULL, not 0 — a field that arrived empty on every bar is absent, and warnings names every such field. mapped, extraKeys and sample show which wire keys were used and what else the point carried. raw=true returns THIS ROUTE'S payload untouched instead of bars — the same request, with nothing projected. Use it only to discover real field spellings when unmapped or extraKeys is non-empty; it is not a fallback and returns no bars. raw=true used to read a DIFFERENT endpoint. It called /charts/:symbol while the projection reads /charts/:symbol/daily, and handed the first the second's vocabulary — so the escape hatch could not reproduce the payload it exists to diagnose. Measured 2026-09-01 the two do not share a vocabulary: /charts/:symbol REFUSES 1w with 400 "Kurun waktu tidak valid" and answers {chart_points: []} for everything else, while the daily route accepts 1w and answers the rich payload. That is why 1w errored and 1m came back empty — the route was wrong, not the value. Fixed: raw now reads the same route as the projection.

running_tradeA

The running-trade tape: individual prints as they cross the exchange. Omit symbol for the market-wide tape, pass it to follow one ticker. action filters to one side (BUY = buyer-initiated, SELL = seller-initiated); omitted means the endpoint's own default, which is not necessarily ALL. grouped=true reads a different endpoint that returns the tape aggregated rather than print by print. It is not a display option: the two return different payloads. WHICH SESSION AM I LOOKING AT. Both views carry data.date and it is the answer — read it rather than assuming today. Measured 2026-09-01 after the ~18:00 WIB broker release, both served that same day. Broker-derived and foreign-derived figures across this API publish at roughly 18:00 WIB, so before that release a payload carrying YESTERDAY's date is correct and not yet published — it is not stale data and not a bug. market_movers' foreign.isShown is the explicit flag for the same thing and flips false to true across that release. INTRADAY BROKER ATTRIBUTION DOES NOT EXIST TO BE RETURNED. IDX closed broker codes on live running trade on 6 December 2021. So is_broker_exists: false with empty buyer/seller is an honest answer about the exchange, not a parse failure or a missing field — check that flag per row rather than concluding the projection broke. THIS TAPE IS NOT LIVE. Measured against the live API during an open session, it runs about EIGHT TO TEN MINUTES BEHIND and refreshes in bursts — its head sat unchanged for over three minutes while the lag grew second for second, and the staleness is at Stockbit's origin, not in a CDN. Do not answer 'what just traded' from it. src/live/tape.ts exists because of this. READ THIS BEFORE ASKING IT ABOUT THE CLOSE. Measured 2026-08-28: the window always starts at the SESSION OPEN, at most 100 rows come back however large limit is, and offset/page are accepted but move nothing. So on a busy ticker this tool can only ever show the first 100 prints of the day — the last trade before the close is NOT reachable through it at any argument. A limit above 100 is refused here rather than silently truncated. For the rest of the session use broker_flow_intraday, which covers 09:00 to 16:14 at one-minute resolution with per-broker value and volume — but note it serves the PREVIOUS session until the ~18:00 WIB broker release, so it does not answer 'what just traded' either. NOTHING here does; docs/LIVENESS.md is the one place that says what each source is worth and why. order_by is required upstream and is NOT a time ordering: 1 is chronological from the open with real trade sizes, 2 and 3 both return lot-ascending rows that were every one a 1-lot trade when measured. None of the three returns the most recent prints and none returns the largest. 1 is the default because it is the only one that answers a question anyone asks. An empty result is NORMAL outside 09:00-16:00 WIB — the tape is a live feed and nothing crosses when the market is shut. Call market_session before treating emptiness as a fault. PENDING VERIFICATION: this response shape has not been observed live.

trade_bookA

Traded volume broken down by price level for a session. mode=BIG_MONEY restricts to large prints, mode=OVERALL covers everything; omitted sends no mode. data_modes excludes auction phases: EXCLUDE_PRE drops the pre-opening auction, EXCLUDE_POST the post-closing one. Both auctions match at a single price and can distort the distribution badly, so excluding them changes the picture rather than trimming it. chart=true reads a different endpoint returning the chart form of the same idea, not a rendering of this one. group_by is REQUIRED: without it every call answers 400 "Group by is required" whatever else is sent, which is why this tool could not be called at all until the parameter existed. PASS group_by=1 — measured 2026-09-01, that returns the by-price book this tool is named for. 2 is also accepted but answered empty on a closed market, so what it groups by is not established; 0 is read as absent and 3 is refused. Only those four have been tried, so your value is sent verbatim rather than checked against a list. chart=true is a different endpoint and has never been seen to demand it, so it is not required there. An empty result is normal before the session's first print. It is a per-session view: there is no date argument, so this is today. PENDING VERIFICATION, and precisely which part: the RESPONSE has now been seen live — a book of price levels, each with buy/sell/pre_open/post_close/total blocks — but mode and data_modes have never been observed to be accepted, because until group_by existed no call from this project ever reached this endpoint at all. Treat those two as unverified guesses and check the payload rather than assuming a filter was applied.

corporate_actionsA

Corporate actions of ONE kind: dividends, rights issues, RUPS (shareholder meetings), bonus shares, splits, reverse splits, tender offers, warrants, public expose, IPOs, or the economic-events calendar. Pass symbol for one issuer's history of that action, or omit it for the market-wide list. An empty rows with rowsFrom: "data" is a real answer, but it means NONE IN THE PERIOD THIS FEED COVERS — nothing in the response says where that period starts, so an empty page is 'none on record here' and is NOT evidence the issuer never did it. action_type is a closed list and an unknown one is rejected here rather than sent, because the endpoint answers an unknown kind with an empty page that reads like a quiet calendar. warrant is a corporate action on the ISSUER's own share count. A structured call warrant — the separately listed instrument a securities house issues against a stock under its own ticker — is not one, and nothing here ties such a ticker back to its underlying. An empty warrant list therefore answers for the issuer alone: it is not evidence that no warrant trades on the symbol. suspectDates appears only when a row's own dates are in an order that cannot have happened — a RUPS record date after the meeting it gates, a cum date after its ex-date, a payment before the record date. That is UPSTREAM data being wrong, not a failure here: the row is still returned exactly as Stockbit sent it, and each entry names both wire keys and both values so you can judge the pair yourself — the values NORMALIZED to YYYY-MM-DD, not the wire strings, so read the row itself to see what was actually sent. The key spellings are candidates, so a suspicion can be a misread field rather than a bad date. Absent means nothing was out of order — it does NOT mean the dates were checked and cleared, because most rows carry neither side of any pair. For dividends prefer dividend_calendar, which also covers stock dividends.

Prompts

Interactive templates invoked by user choice

NameDescription

No prompts

Resources

Contextual data attached and managed by the client

NameDescription

No resources

TDQS

A3.5/5.0

Scored across 17 tools

Disambiguation3/5

The price-data cluster (quote, intraday_prices, chart_series, technicals, orderbook) overlaps heavily, and an agent must read long prose to know which one to reach for; the descriptions do cross-reference each other, which helps somewhat. Worse, several descriptions instruct the agent to prefer or fall back to tools that do not exist here (price_chart, price_bands, price_market, market_movers, market_session, broker_flow_intraday, the `bars` family), which actively muddies selection. broker_summary vs insider_transactions vs running_trade vs trade_book are more cleanly separated by the text.

Naming Consistency4/5

Names are uniformly snake_case with no camelCase intrusions and no mixed casing, following a consistent noun-phrase convention (company_profile, analyst_ratings, running_trade, corporate_actions). A handful of bare single nouns (status, quote, technicals, keystats, ratios, orderbook) deviate from the descriptor pattern, but the style is predictable overall.

Tool Count4/5

17 tools is at the upper end but each maps to a genuinely distinct market-data endpoint (quotes, series, depth, broker flow, insiders, fundamentals, corporate actions, diagnostics), so none is obviously filler. The count only feels heavy because several descriptions are extremely long rather than because the surface is padded.

Completeness2/5

The set reads as a partial slice of a larger server: descriptions repeatedly depend on market_session, price_chart, price_bands, price_market, market_movers, shareholding, insider_ownership, dividend_calendar and broker_flow_intraday, none of which are present. There is also no symbol/issuer search or resolution tool despite tickers being the entry point for nearly every other call, and no financial-statement endpoint except an opt-in inside company_profile. An agent following the documented workflows will hit dead ends.