amaltash-trading

Amaltash trading via the MCP server

This plugin exposes the Amaltash account through the hosted amaltash MCP server (mcp.amaltash.com). Authentication is OAuth, handled by the MCP client: the user runs /mcp → amaltash → Authenticate and approves once in the browser; Claude stores and refreshes the access token. (Headless/CI setups instead send a personal amat_… token as a Bearer header — see the README.)

Always start here

1. Call connection_status to see what's set up (live/paper account, KYC, banks).
2. If a tool reports it isn't authenticated, the user hasn't connected (or their access was revoked/expired). Tell them to run /mcp, pick amaltash, and choose Authenticate, then approve in the browser. Do NOT attempt a headless login yourself.

Tool map (call these, don't hand-roll HTTP)

Goal	Tool
Confirm the connection works	authenticate (validates the current OAuth/Bearer credential)
Who am I / profile	get_account_info
Find strategies	search_marketplace
Create a strategy from an idea	create_strategy (streams token usage → strategy_id)
Improve a strategy	fine_tune_strategy (revise → draft → backtest → activate)
Backtest	backtest_strategy (streams; returns metrics + equity_curve)
Deploy with capital	deploy_strategy (respects the platform minimum)
Stop a deployment	undeploy_strategy
Get a $25K paper account	provision_paper_account (streams provision → fund)
Where am I in onboarding	connection_status
Links to finish KYC / connect bank / connect exchange	connect_links
Brokerage status	get_brokerage_status
Balance	get_account_balance
Balance + why cash isn't all withdrawable	balance_detail
Connected banks	list_banks
Add money	deposit_funds (ACH if a bank is linked, else a Plaid connect link, or method="wire")
Take money out	withdraw_funds (needs confirm=true)

Important behaviours

• Deep-links, not headless actions. KYC, bank connection (Plaid), and live exchange OAuth can't be completed headlessly. For those, call connect_links and give the user the relevant URL to finish in their signed-in browser.
• Money moves need confirmation. Never call withdraw_funds with confirm=true (or deploy_strategy live) until the user has explicitly approved the amount. Default to paper (paper_trade=true) when unsure.
• Streamed tools (create_strategy, fine_tune_strategy, backtest_strategy, provision_paper_account) emit progress as they run — relay it. backtest_strategy returns a downsampled equity_curve ([{t, v}]) you can render or summarize.
• No connected account? get_account_balance / deposit_funds will say so; guide the user to provision_paper_account (sandbox) or connect_links (live).

After creating a strategy — let the user choose

create_strategy makes one strategy and returns its strategy_id. When it succeeds, STOP and hand control back to the user. Do not automatically fine-tune, optimize, or chain into another create — that's the user's call. Briefly report what was created and offer the next steps as options, e.g.:

1. Backtest it — run backtest_strategy (suggest a window if they didn't give one).
2. Tune it — change entry/exit rules, assets, timeframe, or risk limits.
3. Deploy it — paper first unless they explicitly ask for live. Act only on the option the user picks.

Fine-tuning & error recovery

fine_tune_strategy revises a strategy's code from a plain-English instruction, then drafts → backtests → activates it. Never call it on your own initiative — only when the user has explicitly asked to change or fix the strategy.

• Code errors → offer to fix, don't auto-fix. If create_strategy, backtest_strategy, or deploy_strategy fails because the strategy code itself errored (a compile/runtime/execution bug in the generated code), tell the user what broke and offer to fix it via fine_tune_strategy (with the error text as the instruction, e.g. "Fix this execution error: "). Run it only once they say yes; if they accept and it still errors after two attempts, stop and show the error rather than looping.
  • This applies ONLY to code errors. For environment failures — not authenticated, below the capital minimum, KYC/bank not set up — never fine-tune; guide the user to fix the account instead.
• After a clean backtest, offer (don't auto-run) a next iteration. Present the choices and wait for the user to pick:
  1. Tune it your way — ask what they want to change (entry/exit rules, asset, timeframe, risk limits) and pass that to fine_tune_strategy.
  2. Optimize the returns — suggest 2–3 concrete improvements (e.g. add a stop-loss / take-profit, tune the lookback window, add volatility-based position sizing, filter trades by trend/regime, reduce overtrading) and, if they pick one, run fine_tune_strategy with it. Always re-backtest and compare metrics (CAGR, Sharpe, max drawdown) against the prior version so the user can see whether it actually improved.