senti-mcp-server

Tool

Input

What it does

list_accounts

none

Lists the MT5 accounts linked to the configured API key: id, login, broker, last known balance and equity, sync state, running strategies.

list_brokers

none

Lists the platform-wide catalog of brokers Senti Quant supports — not the accounts this API key already has — with each broker's MT5 server names and account types.

list_strategies

none

Lists the platform-wide catalog of strategies (expert advisors) available to deploy — not the strategies currently running on any account — with each strategy's supported symbols, timeframes, rating and presets.

list_account_strategies

accountId (the id from list_accounts, not login)

Lists the strategies currently deployed on one MT5 account, with each deployment's symbol, timeframe and status.

list_positions

accountId (the id from list_accounts, not login)

Lists the positions currently open on one MT5 account, read live from the terminal: symbol, direction, volume, open/current price, stop loss, take profit, swap and floating profit. A 409 means the account's terminal is offline — not that the account holds no positions.

list_pending_orders

accountId (the id from list_accounts, not login)

Lists the pending limit and stop orders resting on one MT5 account, read live from the terminal: symbol, order type, volume, trigger price, stop loss, take profit and stop-limit price. These are orders that have NOT been filled — for open positions, use list_positions. A 409 means the account's terminal is offline — not that the account has no pending orders.

list_deals

accountId, plus optional limit (1–500, default 50), cursor, entry (in or out), from and to (ISO-8601)

Lists one page of an MT5 account's closed deal history — the fills that already happened: symbol, direction, entry kind, volume, price, realized profit, costs, the linked position and order. Paginated, and it never pages on its own: one call is exactly one request, and when more deals exist the answer reports a cursor you must pass back to read the next page. For totals over a period use get_account_performance rather than adding these rows up.

get_account_performance

accountId, plus optional from, to (YYYY-MM-DD, UTC) and reporting (an ISO-4217 currency code, default USD)

Summarizes how one MT5 account performed over a date window: net P&L, win rate, profit factor, gross profit and loss, deal counts, costs, cash flow, period ROI and IRR, lifetime IRR, and the live terminal state. Omit from/to for the last 30 days. Unlike list_positions and list_pending_orders there is no 409 — an unreachable terminal arrives as a null live block inside a success, and is reported as unreachable rather than as zeroes.

get_performance_breakdowns

accountId, plus the same optional from, to and reporting as get_account_performance

Breaks one MT5 account down three ways over a date window: a day-by-day P&L, volume and notional series; a per-symbol P&L and deal-count series; and P&L by hour of the day. Answers "which symbol is losing me money" and "what hour do I trade worst". This response is shaped. The endpoint is the largest the API serves — 87 KB for a 63-day window on a single-symbol account — so per-account rows and running totals are dropped, at most 10 symbols are kept (those with the largest absolute net P&L), and the hourly grid is totalled across the window. Whatever that costs is listed in notes and repeated in the text; notes is empty when nothing was cut. For a single whole-account figure use get_account_performance — it is smaller and it is the default for a performance question.

get_equity_timeseries

accountId, plus the same optional from, to and reporting as get_account_performance

Returns the reconstructed equity curve and floating drawdown for one MT5 account over a date window, as a series of points — answers "how has my equity moved" and "what was my worst drawdown". This response is shaped. A wide window returns a point per interval and grows without bound, so the series is downsampled to at most 200 points — but the first point, the last point and the point of deepest drawdown are always retained, so the start, the end and the worst of the curve are exact rather than sampled near. Measured live on 2026-08-12: 499 points over 63 days → 200. Every downsample is recorded in notes, which is empty when the series was short enough to return whole; narrow from/to for finer resolution. caveats and portfolioCaveats — the API's own statements about figures it could not fully reconstruct — are always returned in full, never shortened.

get_authoring_conventions

none

Reads the Senti Quant MQL5 authoring contract as data: hard-safety constraints, trading-safety requirements, the static analyzer's forbidden-construct list, and the platform limits on draft count and source size. Call this before generating any MQL5 source — code that breaks these rules is rejected by a static scan before it reaches the compiler, and compile slots are globally serial, so discovering a rule by failing a compile is expensive and still fails. Limits are reported exactly — a ceiling that is not a whole multiple of 1024 stays in bytes rather than being rounded into a KiB figure the API does not honour. The response is small (~2 KB) and static per deploy; forbiddenConstructs[].pattern values are regular expressions reported verbatim, never evaluated.

get_draft

draftId (the id field from list_drafts)

Reads one MQL5 draft the API key owns: its full source code, its compiler log, its diagnostics, and whether the last compile still matches the current source. Answers "why did this fail to compile" or "show me the code". The response can be large — a draft may hold up to 192 KiB of source plus 16 KiB of compiler log, and this server returns that content twice, once as text and once as structured data — roughly 105,000 tokens worst case. Attachment source is NOT included; attachments are listed with their size, and list_draft_attachments returns their code — notes points there only when an attachment actually carried source to lose.

list_drafts

none

Lists the MQL5 drafts this API key owns, most recently updated first, with each draft's compile status, size, attachment count and registered-EA id. Use it to find a draftId, or to answer "what am I working on" and "which of my drafts are broken". This response is shaped. GET /api/v1/drafts is the largest payload the API can produce — up to 10.3 MiB across 20 drafts — so source code, compiler logs and diagnostics are ALL dropped; what was cut is listed in notes, and the note only ever names a category that actually lost something — with a byte figure only where bytes were measured. Measured live on 2026-08-20: 19,853 B → 1,898 B, 90.4% removed; worst case at maxDrafts with 5 attachments per draft is roughly 5,000–7,000 tokens, counting both response channels. Call get_draft for one draft's source and compiler output, or list_draft_attachments for its indicator sources. There is no option to request the unshaped response.

list_draft_attachments

draftId (the id field from list_drafts), plus optional filename

Reads the indicator source files a draft's EA embeds via #resource — the source get_draft deliberately leaves out. Pass filename to read at most one attachment whole, by exact name; that is also how to read one a default call had to leave out — if more than one attachment shares that filename, only the first is returned and notes says how many were skipped. A filtered read says so in its text and names the draft's real attachment count, so content alone is never read as the whole set. This response is budgeted, not truncated. With filename omitted, attachments are returned whole while the running total stays within a 64 KiB budget — the first attachment is always returned whole regardless of size, and once one is cut every later one is cut too, and notes says exactly that rather than claiming each cut file exceeded the budget; a cut attachment keeps its metadata and reports sourceCode: null, never a partial source. notes says whether a cut happened. Worst case, counting both response channels, is roughly 33,000 tokens.

create_draft

name (1–120 chars, unique per user), sourceCode (the complete EA)

Write tool — registered only when SENTI_ENABLE_AUTHORING_WRITE is set (see Enabling the write path). Creates a new MQL5 draft from source you have written, and returns its id. Call get_authoring_conventions first: code that breaks the platform rules is rejected by a static scan before it reaches the compiler, and this tool does not check them for you. The response does not echo your source back — you just sent it — so it returns the new id, the byte count written and the compile state, and notes points at get_draft for a read-back. Nothing is compiled until you call compile_draft. A 409 means the name is taken; a 403 means either the key lacks authoring:write or your draft cap is full, and the message says both because the API does not distinguish them.

update_draft

draftId, name, sourceCode

Write tool, behind the opt-in. Replaces an existing draft. THIS IS A FULL REPLACE, NOT A PATCH — both fields are always written, so send the complete draft every time; sending only what you changed deletes the rest of the file, because the API has no partial-update verb. Call get_draft first if you do not have the current source. Annotated destructiveHint for exactly that reason, despite the name. Reports the bytes written, not a before/after delta — the PUT response carries only the new draft, and this server does not make a hidden second request to invent the missing figure. Compiles nothing; if a previous compile no longer matches, the text says so and points at compile_draft.

delete_draft

draftId

Write tool, behind the opt-in — and it asks first. Deletes one draft and every indicator attached to it. Cannot be undone, and no tool here restores one, so it pauses for an explicit human confirmation before anything is sent; declining returns a success saying nothing was deleted, not an error. An EA already registered from the draft is unaffected — a separate resource. Use it to free a slot when create_draft reports the draft cap is full. Needs a host that supports MCP elicitation; on one that does not, this tool cannot be used, and that is deliberate rather than degraded to a silent delete.

add_draft_attachment

draftId, filename (a bare .mq5 basename), sourceCode

Write tool, behind the opt-in. Attaches one MQL5 indicator source to a draft so the EA can embed it. Filenames are unique within a draft case-insensitively — MyInd.mq5 collides with myind.mq5, because the compile host writes them into one flat Windows directory. Attaching does not wire it up: the text names the exact #resource "<stem>.ex5" and iCustom(…) lines the EA still needs, which means an update_draft afterwards, or the file is compiled and never used. The response does not echo your source back.

update_draft_attachment

draftId, attachmentId, sourceCode

Write tool, behind the opt-in. Replaces one indicator's source. The filename cannot be changed and this tool takes no filename — an EA embeds an indicator by name, so a rename would orphan every reference; to rename, delete, re-add and update the EA source. A full replace of that file's contents, so send the complete indicator. A 404 here may also mean the attachment belongs to a different draft.

delete_draft_attachment

draftId, attachmentId

Write tool, behind the opt-in — and it asks first. Removes one indicator from a draft. Cannot be undone. Afterwards the EA still references it: remove its #resource and iCustom lines with update_draft, or the next compile_draft fails on a file that is no longer there — the text says so. Also how to free a slot when the attachment cap is full, and the only way to rename a file. Needs a host that supports MCP elicitation.

compile_draft

draftId

Write tool, behind the opt-in. Runs the static-safety scan and the MQL5 compiler over a draft and every indicator attached to it, and returns the verdict, the diagnostics and the compiler log. A check only — it registers and deploys nothing. A failed build is not an error: the tool succeeds and reports ok: false with diagnostics, so read the result rather than retrying. The compile slot is one per account and the compile server is globally serial, so a second concurrent call is a 409 and contention is a 503 with a wait — this server reports both and retries neither. If the 15s client timeout fires, the compile keeps running on the server: the message says so and sends you to get_draft for lastCompileStatus.

Variable

Required

Default

Purpose

SENTI_API_KEY

✅

—

First-party key. The server exits at startup without it.

SENTI_API_BASE_URL

https://api.sentitrade.xyz

Set to https://be-dev.sentitrade.xyz for development.

SENTI_ENABLE_AUTHORING_WRITE

unset (off)

1 or true registers the authoring write tools. See below.