DeepVest MCP

Server URL

https://api.deepvest.ai/mcp

Transport

Streamable HTTP

Auth

OAuth sign-in with your DeepVest account (Claude, ChatGPT, Cursor, VS Code). No key needed.

Also supported

An API key, sent as Authorization: Bearer YOUR_API_KEY or X-API-Key: YOUR_API_KEY.

Results

Text plus structured data your assistant can read (and charts/CSVs where the tool produces them). Tools that draw charts return them as data, and hosts that support MCP Apps (Claude on the web and desktop, ChatGPT, VS Code with GitHub Copilot, Cursor and others) render them in the DeepVest chart viewer. CSV exports arrive as attached text/csv files.

Rate limits are enforced per API key in two classes. Tool calls (tools/call) have a concurrency cap and a per-minute cap. Everything else (initialize, tools/list, ping, SSE streams) only has a looser per-minute cap and never holds a tool slot.

Need more for a benchmark or a high-fan-out agent? Contact support: limits are set per key and take effect immediately.

401

Missing or invalid credentials. OAuth clients use this response to start sign-in.

Limits on tool calls

Out of credits or rate limited on a tool call: the call returns a readable tool error (isError: true) that your assistant can relay, for example a prompt to top up on the Billing page or to retry in a number of seconds. Rate-limited tool calls keep the Retry-After and X-RateLimit-* headers.

Credit checks

Credits are only checked on tool calls: with an empty balance your client can still connect and list tools. A tool call sent without a JSON-RPC id (a notification) has no result to carry a message, so it gets a plain HTTP 402 or 429.

429 on other requests

Requests other than tool calls (initialize, tools/list, streams) that exceed their per-minute cap get HTTP 429. A 429 carries a Retry-After header (seconds) and a JSON body with reason (per_minute_limit; these requests have no concurrency cap), scope, limit and retry_after. Back off for that many seconds, then retry.

Rate-limit headers

Every successful response includes X-RateLimit-Limit, X-RateLimit-Remaining and X-RateLimit-Reset (seconds) for the budget that applied.

Tool errors

A failed tool call returns isError: true with a message that says what to change. Unexpected failures read <tool> could not complete this request (reference <id>).; quote the reference id in your support request. A failure the analysis engine reports as text is handled the same way: an unexpected one gets the generic message with a reference id, while a validation, limit or no-data message is shown as written.

Typed choices

The math and tax tools accept their fixed choices in any letter case (CALL or call, MFJ or mfj) and common filing-status spellings: married, joint or married filing jointly for mfj; separate or married filing separately for mfs; head of household for hoh. A spelling only works where the tool lists its target value (roth_conversion_tax takes single and mfj only).

Default size

Results are capped at 40,000 characters (about 10k tokens) by default. Long output is cut with a note saying how to narrow the request.

Up to 100,000

Tools that return long tables or reports: asset_analysis, stock_screener, quick_analysis, web_search, strategy_backtest, event_backtest, macro_conditional_backtest, options_hedge_backtest, portfolio_optimization, portfolio_comparison, options_chain, options_screener, merged_data_analysis, dividend_history, fund_filings, sec_filing_text_changes, document_processing, holdings_overlap_tool, lookthrough_exposure_tool and asset_class_breakdown_tool.

Up to 80,000

options_analysis, earnings_analysis and earnings_event_backtest.

Structured data

The data in structuredContent is bounded to the same budget (trailing list and series items first; a record never loses a field) and structuredContent.omitted.data_keys records what was trimmed.

Whole result

Stays within 140,000 characters, spent on text and data, then charts, then CSV exports, so a CSV may be cut (keeping its header) or left out; omitted.csv_rows, omitted.csv_files and omitted.charts count what was cut.

Response times

Most tools respond in 5 to 30 seconds. The backtests, portfolio optimization and comparison, and merged data analysis can take up to 2 minutes; while they run they send MCP progress notifications every 5 seconds to clients that ask for progress.

Your usage

Monitor your consumption on the Usage page.

Current fundamentals

latest_fundamentals takes tickers: list[str] and optional forward_pe: bool. Pass all tickers in one call: up to 25 distinct names use one backend batch; larger lists are automatically chunked inside the call. Missing names are explicitly reported. Set forward_pe=true only for a requested forward P/E; this adds dated next-fiscal-year consensus estimates.

Current versus historical

Current P/E uses this dated snapshot. For P/E since January 2020, use quick_analysis with the same tickers and that start date.

Measurement basis

Unqualified P/E is trailing TTM. Keep each returned date and basis; PEG's growth horizon is vendor-defined. Report coverage gaps rather than reconstructing missing ratios from SEC filings.

Statement cadence

Annual and quarterly statements are independent of price-bar frequency. The statement path returns the last four quarters or five fiscal years; older filing archives require sec_financial_analysis.

Rolling windows

Unless you specify a size, rolling metrics use 252 daily bars, 52 weekly bars, or 24 monthly bars.

Dividends

dividend_history takes symbols, optional start_date, end_date; optional drip_shares, drip_purchase_price, drip_spinoffs. A single call returns at most the 60 most recent events per symbol, so pass a date range for older periods. With drip_shares (one symbol) it also simulates DRIP cost basis.

Documents

document_processing takes file_path (one File Library key or a list of up to 10, starting with files/ or images/) and a query. Local paths, URLs and other accounts' files are refused before anything is read.

Valuation workflow

Use wacc_calculator_tool to compute the discount rate, then pass it to dcf_valuation_tool. Use justified_fcf_multiple when the question is about a fair multiple before a full DCF. Use dcf_sensitivity_table_tool to stress-test assumptions, dupont_analysis_tool to see what drives a company's ROE, dividend_discount_model_tool for dividend-paying equities and the swap and Z-spread tools for rate and bond valuation questions.

Swaps

spot_rates (or a single flat_rate) are annual-effective zero rates, exactly one per payment period; the i-th discounts the payment at t = i / payment_frequency years with DF = (1 + s)^-t. The fixed leg pays notional x fixed_rate / payment_frequency each period, and floating coupons are the single-curve forwards implied by the same discount factors. Whole periods only: swap_tenor x payment_frequency must be a whole number (5.5 years semi-annual is 11) and swap_tenor is at most 50 years.

Options

time_to_expiry is at most 30 years and binomial_option_price_tool takes at most 2,000 steps. monte_carlo_option_price_tool takes one step per trading day (252 a year) within a budget of 2,000,000 path-steps and at most 200,000 paths, so about 7,900 paths fit a 1-year option; omit num_simulations and it runs the most paths that fit, up to 100,000. For a vanilla European option use black_scholes_option_price_tool (exact); Monte Carlo is for barrier and path-dependent payoffs.

Bonds

bond_calculator_tool needs years_to_maturity (at most 100). price needs ytm; ytm, current_yield and zero_coupon_yield need price in dollars, not a percent of par; duration, convexity and full_analysis need ytm or price, not both; zero_coupon_price needs ytm above 0. Yields from -5% to 200% are supported, negative yields included; a price implying a yield outside that range is refused with the price range that fits.

Tax years

capital_gains_tax, roth_conversion_tax and rmd_schedule_tool (tax_year), amt_estimator_tool and irmaa_surcharge_tool (year) take 2025 or 2026, the years with published federal tables; another year is refused rather than priced on a different year's table.

Social Security

social_security_rules_tool takes claiming ages from 62 to 70 (fractions allowed, counted to the nearest month).