Stockbit MCP

Bring Claude to your IDX trading desk.

Accepting donations: Saweria Link

Bandarmology, quotes, orderbook, fundamentals, your watchlists and portfolio — and, only if you switch it on, confirm-gated order entry — through your own Stockbit account, from Claude Desktop, Claude Code, Cursor or any MCP client.

English | Bahasa Indonesia

Broker-to-broker flow, rendered by the server. Synthetic data.

How it works (and why it is safe to run)

It is an HTTP client, not a bot. Every number it reports comes from a JSON endpoint in a closed route table — no headless browser scraping pages, no reading data off Stockbit's UI, no polling loop you did not start.

• One interactive login, captured from your own browser. You sign in on Stockbit's real page; the server reads the refresh token out of the response and stores it. Your password never touches this code.

• Three token domains, three separate stores. exodus (market data), carina (Stockbit Sekuritas), api-sekuritas (e-IPO). Logging out of one leaves the others alone, and the route table decides which credential each request may carry.

• The 24-hour access token is cached on disk, encrypted, and shared between processes. Because the refresh token rotates on every use, N clients each minting their own access token retire each other's credential. Same AES-256-GCM and mode 0600 as the file-backend refresh token — which on macOS is a genuine reduction, since there the refresh token is in the Keychain and this is not. STOCKBIT_NO_ACCESS_CACHE=1 turns it off. See SECURITY.md.

• A closed route table. 153 permitted request shapes across three hosts, enumerated in src/http/routes/. Anything not in that table cannot be requested — test/transport.test.ts asserts it, and every one of the 32 non-GET routes is admitted by a named decision record.

• Every log and every tool result is redacted. Tokens, PINs and bot tokens are matched by shape as well as by key name.

• A rate limit that behaves like a person. Three concurrent requests, 150 ms apart.

• Trading is a ladder you climb deliberately: off → stockbit-auth trading-enable --paper → --live. The environment can only move you down it. Orders are two steps with a human in the middle, and where your client supports elicitation you are asked directly, and your answer decides it — a model cannot skip that ask by asserting you already agreed.

flowchart LR
  C["Claude Desktop / Code / Cursor"] <-->|MCP over stdio| S["stockbit-mcp<br/>(your machine)"]
  S -->|market data| E["exodus.stockbit.com"]
  S -->|portfolio, orders| K["carina.stockbit.com"]
  S -->|IPO| A["api-sekuritas.stockbit.com"]
  S -.->|drawing only| B["your logged-in browser"]
  S -.->|paper mode| L["local ledger"]

Related MCP server: jv-idx-mcp

What this tool does not do

• No PIN handling by any tool. The six-digit trading PIN is typed at your terminal, used for one request and never stored. If anything asks you for it through an assistant, that is not this.

• No order without a ticket. By default, the write tools also need your confirmation. The only exception is capped --auto-confirm, which you must deliberately enable for live trading at a terminal; a model cannot enable it or widen its value cap. The tools take no price or quantity, so what reaches the exchange is exactly what the ticket described.

• No auto-resend, no auto-cancel. When an order's outcome is uncertain the server says so and stops. A resend is how one intention becomes two orders.

• Saved workflow recipes cannot write. Enforced by construction: a write tool is never added to the map recipes look names up in.

• No route outside the table. No day-trade or smart orders, no withdrawals, no deposits, no posting to the stream.

• No scraping, and no UI automation for data. Your own browser is used for three things and nothing else: the one-time login, drawing on your own chart, and opening Stockbit when you ask to look at it. Nothing is ever read out of the page.

• Nothing leaves your machine except to Stockbit, and to channels you configured.

• No short selling — IDX retail has none — and no financial advice.

Prerequisites

What it does

Bandarmology. broker_summary, broker_distribution, broker_activity, bandar_detector — who accumulated, who distributed, and who was on the other side of the tape. NET and GROSS, all four market boards, ten period windows including year-to-date in a single request. This is the data no other market API has, and it is why this project exists.

Market, company and fundamentals. Quotes, full orderbook depth, auto-rejection bands, movers, daily bars, seasonality, key statistics, ratios, financial statements, ownership, insider activity, corporate actions, analyst ratings and peer comparison.

One analysis engine. Indicators, 16 candlestick patterns, multi-timeframe alignment, 9 strategy presets, backtests with walk-forward validation, universe scans, and TradingView Pine generation — all over the same series grammar, so the Pine you paste into TradingView fires on the condition that was actually measured.

Your account. Watchlists and saved screens, read and edited, with every write verified by reading the account back.

Your chart. Read and draw on your real Stockbit chart, in your own logged-in browser.

Automation. Eight workflows, also offered as MCP prompts. An alert daemon that keeps watching while no client is open, delivering to a log, a desktop notification, a webhook and Telegram.

What price_chart returns. Synthetic data.

Why not a TradingView MCP?

Installation

Claude Code

claude mcp add --scope user stockbit -- npx -y stockbit-mcp

Claude Desktop — claude_desktop_config.json:

{ "mcpServers": { "stockbit": { "command": "npx", "args": ["-y", "stockbit-mcp"] } } }

On Windows, npx needs a shell:

{ "mcpServers": { "stockbit": { "command": "cmd", "args": ["/c", "npx", "-y", "stockbit-mcp"] } } }

Claude Desktop Extension — download the latest stockbit-mcp-*.mcpb from Releases and double-click it.

Cursor — ~/.cursor/mcp.json. Cursor stops at 40 tools and the default core profile is exactly 40, so nothing extra is needed — though that leaves no room for a second MCP server, and running one means a narrower list (STOCKBIT_TOOLS=market,bandarmology, say):

{ "mcpServers": { "stockbit": { "command": "npx", "args": ["-y", "stockbit-mcp"] } } }

VS Code — .vscode/mcp.json. Its cap is 128; the default fits with room to spare:

{ "servers": { "stockbit": { "type": "stdio", "command": "npx", "args": ["-y", "stockbit-mcp"] } } }

Windsurf — ~/.codeium/windsurf/mcp_config.json, same shape as Claude Desktop.

Codex CLI — ~/.codex/config.toml:

[mcp_servers.stockbit]
command = "npx"
args = ["-y", "stockbit-mcp"]

Claude Code plugin (adds six trading-desk skills):

/plugin marketplace add INo-xious/stockbit-mcp
/plugin install stockbit@stockbit-mcp

From source

git clone https://github.com/INo-xious/stockbit-mcp && cd stockbit-mcp
npm ci && npm run build
node dist/bin/stockbit-mcp.js

npm i -g stockbit-mcp avoids the npx cold start on every launch — at the cost of the automatic updates described next.

Staying up to date

Every configuration above uses npx -y stockbit-mcp with no version, and that is deliberate: npx re-resolves the newest release each time the server starts. A new version reaches you the next time your client launches it, with nothing to do.

That is measured, not assumed. With 1.1.0 already sitting in the npx cache, the next bare npx -y stockbit-mcp ran 1.1.1.

If you would rather not move. Pin a version and nothing changes under you:

{ "mcpServers": { "stockbit": { "command": "npx", "args": ["-y", "stockbit-mcp@1.1.1"] } } }

Or take patches and minors but never a breaking change — this project follows semver, so a new major is the only release that can break your setup:

{ "mcpServers": { "stockbit": { "command": "npx", "args": ["-y", "stockbit-mcp@^1"] } } }

To force a refresh right now, without waiting for a restart:

npx -y stockbit-mcp@latest --version

Which version you are actually running is always answerable — ask your assistant "is my Stockbit MCP working?" and status reports it, alongside what else is and is not set up.

Quick start

1. Install — one of the above.

2. Log in, once. Say "log me into Stockbit" and sign in in the browser window that opens. Or, at a terminal: npx -y -p stockbit-mcp stockbit-auth login. No browser? stockbit-auth import-har takes a login captured in any browser; stockbit-auth doctor diagnoses the rest.

3. Restart your client so it picks up the tools.

4. Ask: "Is my Stockbit MCP working?" — Claude calls status, which reports the version, which sessions exist (never the tokens), the trading mode, where the IDX trading day is in WIB, and the single next command if anything is missing. It answers with no session at all, which is where everyone starts.

5. Optional — practise first. npx -y -p stockbit-mcp stockbit-auth trading-enable --paper, then "buy 1 lot of BBRI on paper". No real money, no PIN, and the same protocol as the real thing.

Example prompts

"analyze BBRI" "who accumulated GOTO between 2026-07-01 and 2026-07-31" "broker distribution for BRMS" "technicals for BBRI" "chart BBRI with bollinger bands and a MACD panel" "give me Pine for BBRI with a golden cross alert" "alert me when BBRI's RSI drops below 30" "run a deep dive on BBRI" "do the morning scan" "draw the support and resistance on BBRI's chart" "which of my watchlist stocks had brokers accumulating yesterday" "size a BBRI position: entry 4100, stop 3900, risk 1% of Rp 50 million"

How Claude knows which tool to use

Tool reference

138 tools in 17 families. The full generated reference — every tool, its evidence, its arguments — is docs/TOOLS.md.

Verification status

Every tool carries one of three words, and they mean exactly this:

• Observed — a real response from a live account was seen, and the code was written against it.

• Read-back — a write whose effect is verified by re-reading the account afterwards. The request body may still be a guess, but a wrong guess shows up as not-visible, never as a false success.

• Projected — field names taken from Stockbit's web bundle, never seen on a live response. readFrom names the wire key each value came from, and an absent field means "not recognised", not zero.

Projected is not a warning that something is broken. It is a statement that nobody has checked it, and that the code is built so an unchecked guess fails loudly rather than quietly.

The trading and e-IPO families have never been observed live — reading them needs a securities session, which needs the account owner's PIN at their own terminal. Per-family detail, with dates and what was compared against what, is in docs/VERIFICATION.md; what is still open, and in what order it matters, is in docs/PENDING-VERIFICATION.md.

Safety model

The switches. Trading is off until you run stockbit-auth trading-enable --paper or --live yourself. A bare trading-enable is refused — the two differ by everything. STOCKBIT_TRADING can only move the mode down the ladder (live → paper → off); no value of it turns anything on. No module under src/tools/, src/trading/ or src/eipo/ may write the settings file, and a test asserts that.

The PIN. Typed at your terminal, used for one request, never stored. No MCP tool accepts one.

The ticket protocol. order_preview prices and checks the order and returns a summary you read. The write tools take that ticket id and an optional confirmation, and nothing else. A ticket expires in two minutes and carries a fingerprint that is rechecked before the request goes out.

Who confirms. Where your client supports MCP elicitation, you are asked directly, before confirm is even looked at, and your answer is the decisive one — declining refuses the order however the model set confirm. Elicitation is the only channel in MCP that reaches a person; confirm: true is a boolean the model sets, and treating the two as interchangeable was a real defect here, fixed in ADR-0010. On a client that cannot elicit, confirm: true is the only gate there is, the order proceeds, and both the result and the audit line say plainly that no human was asked. The model must never set confirm on your behalf.

You own three switches over that, all set at your own terminal and none reachable by any tool:

The confirmation dialog also carries a second box you may tick: don't ask again, for fifteen minutes, for orders worth no more than the one you just approved, under the trading policy that was in force when you ticked it. It lives in that server's memory and never on disk — a restart ends it, trading_forget ends it in that conversation, and stockbit-auth trading-forget ends it everywhere including servers already running. status says whether one is live.

Outcomes. After a write, outcome is one of seven classes. ok is the only clean success; landed-despite-error also means the read-back found the order, but the request itself errored. Never resend any non-ok result:

Audit. Every order attempt and every account edit appends a line to a log, redacted, whatever the outcome — and if that line could not be written, the result says so rather than implying an audit trail that does not exist.

Credential storage. macOS Keychain where available. Elsewhere an AES-256-GCM file whose key is derived from hostname and username: that is obfuscation, not a vault — anything running as you on your machine can derive the same key. See SECURITY.md.

Decision records: ADR-0004 (order entry), ADR-0007 (login as a tool), ADR-0008 (paper mode). Full guide: docs/trading.md.

Paper trading

stockbit-auth trading-enable --paper          # Rp 100,000,000 to practise with
stockbit-auth paper-reset --cash 250000000

A local ledger, no exchange, no session, no PIN — and the identical protocol, so nothing about the live path is a surprise later. Your portfolio, positions, cash, orders and history are served from the ledger while it is on, and every result says PAPER ACCOUNT — no real money.

Fills are approximate in three specific ways, stated on every result: close-only minutely data (so some real fills are missed), no queue position (so paper is optimistic), and no partial fills. Do not backtest against it and believe the number.

Tool profiles and context management

Every client pays for the whole tool list in the model's context on every turn — and that is a per-turn cost, not a startup one. The full surface is around 220,000 bytes of tools/list, roughly 55,000 tokens, on every single message; core is about a third of that.

core is the default. STOCKBIT_TOOLS changes it:

system (status, login, logout) is never filtered out — it is how you find out why everything else is missing. An unknown value stops the server with a message naming every family, rather than silently loading all 138.

Output sizes. analyze makes about 27 upstream requests and returns 4–8 KB. broker_summary takes a limit. financials is large. Chart tools return a base64 SVG and a file path. Prefer technicals when you only need the numbers.

Alerts and the daemon

An MCP server exists only while a client holds it open, so a rule that fires at 14:20 on a Tuesday needs a separate process:

npx -y -p stockbit-mcp stockbit-alerts watch      # every 60s during IDX hours
npx -y -p stockbit-mcp stockbit-alerts check      # one pass
npx -y -p stockbit-mcp stockbit-alerts test       # exercise every channel

Channels: an append-only log (always, and first — a channel that fails silently is worse than none), a desktop notification, a webhook (STOCKBIT_ALERT_WEBHOOK, https or localhost only), and Telegram.

For Telegram: get a token from @BotFather, message your bot once, then read the numeric chat id from https://api.telegram.org/bot<token>/getUpdates. Set STOCKBIT_TELEGRAM_BOT_TOKEN and STOCKBIT_TELEGRAM_CHAT_ID — environment only, because a token on the command line is visible to every user on the machine through ps.

Keep it running with launchd (macOS), Task Scheduler (Windows) or a systemd user unit (Linux).

CLI reference

Every command answers --help/-h without running anything, and an unknown flag or stray argument is an error naming what that command accepts — never silently ignored.

Files it writes

Everything lives in one directory — ~/.stockbit, or wherever STOCKBIT_STORE_DIR points. Nothing is written outside it.

Environment variables

Troubleshooting

Testing

npm ci
npm run typecheck
npm test          # the whole suite, entirely offline — fetch is stubbed, no network, no skips
npm run build
npm run smoke     # starts the built binary over stdio and asks it what it registered
npm run check:pack

CI runs all of that on Ubuntu, macOS and Windows against Node 22 and 24, plus a dependency audit, an offline link check, and npm run docs:tools && git diff --exit-code docs/TOOLS.md.

Architecture

bin/         three entry points: the MCP server, auth, the alert daemon
src/
  http/      the closed route table and the transport that enforces it
  auth/      login capture, the three token stores, refresh with a cross-process lock
  core/      one module per Stockbit domain — the readers everything else is built on
  analysis/  indicators, patterns, strategies, backtests, scans, position sizing
  render/    pure SVG: candles, flow diagrams. No browser.
  tools/     MCP registration, one module per family
  trading/   tickets, previews, order submission, the paper ledger
  eipo/      the IPO pipeline
  alerts/    rules, the daemon, delivery
  workflows/ the eight built-in recipes

Three invariants, each enforced by a test rather than a convention:

1. Nothing can reach a Stockbit host outside the route table (test/transport.test.ts).

2. A write tool is never reachable from a saved workflow recipe — define.write does not add to the handler map (test/tools.test.ts).

3. Nothing that serves a model can write the settings file (test/settings.test.ts).

Roadmap

A trade journal; rasterising charts to PNG for vision models; Windows DPAPI and Linux libsecret credential stores; CSV export; a stockbit run <tool> CLI. After that: converting the carina and e-IPO families from Projected to Observed, which takes one live session and is the account owner's call.

Contributing · Security

CONTRIBUTING.md · SECURITY.md · CODE_OF_CONDUCT.md · CONTEXT.md (the vocabulary) · docs/README.md (the index).

Longer than this page: the user guide walks through every feature and what to ask for; docs/TOOLS.md is the generated reference; docs/adr/ records every decision that changed what this server may do.

Please report vulnerabilities privately through GitHub Security Advisories, not a public issue.

Acknowledgements

Garit32 — collaborator, and the author of most of the commits on this project's history, including the Windows login capture and browser detection.

tradesdontlie/tradingview-mcp — the polish this README was measured against. Stockbit's chart page is built on TradingView's charting library, which is why Chartbit works the way it does.

Disclaimer

This software is for personal, educational and research use.

It is not affiliated with, endorsed by or supported by Stockbit, PT Stockbit Sekuritas Digital, or the Indonesia Stock Exchange. Stockbit's Terms of Use restrict automated access to their services, and using this software may conflict with them; account suspension is a possible consequence and it is yours to weigh. The API it uses is undocumented and can change or break at any time without notice.

Nothing this software produces is investment advice. The author is not a licensed investment adviser and is not registered with OJK. Indicators, backtests, pattern detections and broker-flow readings are computations over historical data, not predictions. Backtested results do not predict future returns.

If you enable live trading, this software can send real orders that spend real money. It is off by default. Orders require your explicit confirmation unless you separately opt into capped --auto-confirm; you remain solely responsible for every order placed through it, including orders within a cap you authorised in advance.

You are responsible for complying with Stockbit's terms, IDX rules, and Indonesian law.

Provided under the MIT licence, without warranty of any kind.

License

MIT © Marvel Harisson