Agentic systems for trading and risk, built so that every number arrives with its record.

AvidQuant builds agents, MCP servers and risk software for trading and risk desks. The figure below is a real result from one of them.

97.5% one-day Value at Risk Synthetic example

511,773 USD

Expected Shortfall 617,885.16 USD, on a book of 12,000,000 USD.

Distribution of the 1,000 synthetic one-day P&L observations A histogram of profit and loss from minus one million to plus one million USD, centred near zero with heavy tails. A solid line marks Value at Risk at a loss of 511,773 USD and a dashed line marks Expected Shortfall at a loss of 617,885 USD. The 25 observations at or beyond Value at Risk are drawn in red. Three observations lie beyond plus or minus one million and are drawn in the bars at the two ends of the axis. ES 617,885 VaR 511,773 −1.0m −0.5m 0 +0.5m +1.0m Loss side: P&L below zero 1,000 observations, USD
One-day profit and loss of the example book. The 25 observations at or beyond Value at Risk are red. Three observations beyond ±1.0m are drawn in the end bars.
Method
Historical Value at Risk and Expected Shortfall, empirical order statistic
Confidence
0.975
Horizon
1 day. Each observation already spans it, so nothing is rescaled
Currency
USD
Expected Shortfall
617,885.16 USD, the mean of the 25 largest losses
Observations
1,000, as of 12 September 2026
Sign convention
Loss is positive. A negative value is a gain and is not floored at zero
Data mode
Synthetic. A drawn sample, not market data and not a real book
Source
Reproduce with numpy.random.default_rng(7).standard_t(4, 1000) * 180000, rounded to whole USD
Input digest
sha256 1cf2b44607c3ddbdd5b64f3218f7622fa9f68dc9fa8d8da3439607186c460f7c
Assumptions
  • Equally weighted empirical distribution.
  • Sample dependence and forecast validity are not tested.
  • Expected Shortfall integrates fractional tail mass.
Degraded
false. Nothing was missing or substituted

Returned by the risk_tail tool of the public Risk Desk endpoint, called with no arguments, on 10 November 2026. Risk Desk 0.2.0. Research software, not investment advice.

Limits, read like a meter bridge

Risk Desk’s limit check over its own synthetic example. Each needle shows how much of its limit a measure uses; the red arc is past the limit, and a short bar marks a warning level where one is set.

Synthetic example Degraded The tail result behind Value at Risk and Expected Shortfall has fewer than 20 observations of tail mass, so the check reports itself as degraded.

  • Gross exposure

    800 of 1,000 USD

    Warning, 80% of limit

  • Net exposure

    400 of 500 USD

    Within, 80% of limit

  • Asset A share of gross

    75% of 70%

    Breach, 107% of limit

  • Value at Risk

    50 of 60 USD

    Warning, 83% of limit

  • Expected Shortfall

    83.33 of 100 USD

    Within, 83% of limit

  • Stress loss, combined down

    20 of 50 USD

    Within, 40% of limit

Returned by the risk_limits tool of the public Risk Desk endpoint, called with no arguments, on 10 November 2026. Overall status: breach, with three limits within, two at warning and one breached. A separate small example from the 12,000,000 USD book above.

Two desks, public and running

Both are source-available, have a hosted sample that needs no account, and give an agent and a person the same figures from the same functions.

Option Desk

Option analytics from a chain that you supply, or from the built-in synthetic sample.

Computes
Greeks, dealer positioning, strategy payoffs, GARCH simulation and backtests
Surfaces
Command line, local dashboard, eight agent skills, a LangGraph agent
MCP
14 tools on the local server, 13 on the hosted endpoint https://optiondesk.avidquant.com/mcp
Licence
PolyForm Noncommercial 1.0.0

Risk Desk

Portfolio risk from data that you supply, or from a shipped synthetic example.

Computes
Value at Risk and Expected Shortfall with backtests, stress tests, SA-CCR, credit VaR and FRTB capital components
Surfaces
Command line, local dashboard, 14 agent skills, a LangGraph agent
MCP
30 tools on the local server, 29 on the hosted endpoint https://riskdesk.avidquant.com/mcp
Licence
PolyForm Noncommercial 1.0.0

The work AvidQuant takes on

Four areas. Each one is already on show in the desks, so you can read the code before you write.

Agentic development

Agents that call typed tools, write each result to a file and say what they could not do. No figure comes from the model's memory.

On show in the LangGraph agent packages of both desks.

MCP servers and tools

Local and hosted Model Context Protocol servers, with titled and annotated tools, consent on supplied data, and work limits measured on the server.

On show in four servers, two local and two hosted, listed on the official MCP Registry.

Financial engineering

Option pricing and Greeks, volatility models, Value at Risk and Expected Shortfall, stress testing, counterparty and credit risk, and FRTB capital components.

On show in the methods tables of both repositories, formula by formula.

AI governance and model risk

Model inventories, assumptions registers, independent numerical review, and gates that fail a build when a document and the code disagree.

On show in Risk Desk: 29 models inventoried and 186 review checks recorded.

Questions

What is AvidQuant?
AvidQuant builds agentic systems, MCP servers and quantitative risk software for trading and risk desks, and publishes two of them as source-available projects: Option Desk and Risk Desk.
What do Option Desk and Risk Desk do?
Option Desk computes Greeks, dealer positioning, strategy payoffs, simulations and backtests from an option chain that you supply. Risk Desk computes Value at Risk, Expected Shortfall, backtests, stress tests and counterparty, credit and FRTB figures from data that you supply. Each result carries its source, units, assumptions and a degraded flag.
How do I connect an AI assistant to the desks?
Both desks speak the Model Context Protocol. Add https://optiondesk.avidquant.com/mcp or https://riskdesk.avidquant.com/mcp as a remote MCP server in a client such as Claude, ChatGPT or Codex. Neither endpoint needs an account. Each repository also installs a local MCP server that computes on your own machine.
Are the desks free to use?
The source of both desks is public under the PolyForm Noncommercial License 1.0.0, so noncommercial use is free. Commercial use needs a separate agreement. Write to [email protected].
Do the desks give investment advice or place orders?
No. Both desks are research software. They connect to no broker, place no orders and give no personal recommendation. The hosted services fetch no market data.
How do I contact AvidQuant?
Write to [email protected] about agentic development, MCP solutions, financial engineering, AI governance or the desks.

For any enquiry, please write.

[email protected]

Agentic development, MCP solutions, financial engineering, AI governance, a commercial licence for either desk, or a question about a result. A few lines on what you are working on is enough to start.