Calculate with an AI agent
DECCF can calculate cashflows from an AI conversation and provide a compact form for adjusting assumptions. Your AI explains the results with tables and charts in the conversation. Choose Create application to save the setup and current values as a private app in Your Apps.
Install the skill or plugin
The skill explains the model format and includes a portable JavaScript runtime. The plugin includes the same skill and its MCP connection. An account is optional for calculating a submitted own model without domain modules. Protected account models, licensed apps and cases require an approved connection.
Both packages include the complete public calculation documentation, organised into searchable sections. Your AI can retrieve the relevant explanations and model syntax before building a calculation, including inputs, entities, accounts, cashflows, events, timing and results. Each search result links back to its documentation page. The documentation is also available through the DECCF connection, including for anonymous calculations.
For agents supported by the Skills CLI, install the public skill directly:
npx skills add https://deccf.app/integrations/deccf-skill.zip --skill deccf-calculateChoose the agent and installation scope in the installer. Add --agent codex for Codex, or --global to make the skill available across your projects. Installation does not require a DECCF account. The same package includes the model references, searchable manual and local calculation module.
Your host must support the corresponding installation format. Use the host's skill upload flow or plugin developer import; remote MCP clients need a publicly reachable HTTPS endpoint. Actual availability and publication depend on the host's rollout and approval process. OpenAI plugin packaging, skill packaging.
Calculate, compare and save
Ask your AI for a calculation and provide the important assumptions. Calculations are editable by default. The AI runs the model in the background before presenting it, reads the actual output, and explains it with tables and charts. Where several accounts have meaningful flows between them, a Sankey diagram shows their relative amounts.
The compact calculator contains input fields, Calculate and Create application. Scenarios and run options are collapsed until needed. Results and model source stay out of this form. Pressing Calculate sends the new output back to your AI for interpretation. If your chat cannot receive it, download the output and attach it to the conversation.
Opening or adjusting the calculator does not save an app. Create application opens the saving flow; after signing in, keep the setup in Your Apps for later use. This action has the same English label in every language.
Sign in to use domain packages
Without signing in, you can calculate with the basic framework, but domain packages cannot be used. Your AI explains this limit and offers the option to sign in or create an account. Connecting an account lets you choose the available domain models.
When domain packages are useful, the AI shows their names, versions, included dependencies and purpose, then asks whether you want to install them. After your explicit approval and sign-in, installation happens directly through your DECCF connection. Replacing a previously installed version also requires your approval. A package installation does not save the calculation.
Explore uncertain forecasts
When cashflows depend on uncertain forecasts or estimates, your AI asks whether you want a Monte Carlo simulation. If you agree, it starts with 100 simulations and explains the uncertainty assumptions. The main result is presented as a histogram with a cumulative probability S-curve, showing both the spread and the proportion of outcomes below each value.
Treat the first 100 simulations as an exploration of the setup. If the calculation looks useful, run about 1,000 simulations for a more stable picture. More simulations do not remove uncertainty in the underlying assumptions.
If a model cannot run
The AI should reread the documentation and correct its model after an execution error. If errors persist, it should make at least four attempts. If those corrections still fail, it should explain that it could not formulate the model correctly because it does not fully understand the documentation, and suggest trying a more capable AI model with stronger reasoning.