Platform · AI copilot, agents & workflows

AI inside the structure.

Saltcorn gives generative AI a structured environment to work in. Build with AI, and build AI applications on top of your data — without giving up control.

Build Saltcorn with AI Build AI into your app

Try it now Read the docs →

Copilot · agents · workflows RAG · MCP · voice

01 / Build with AI

A copilot that builds Saltcorn.

The copilot is itself a Saltcorn agent, assembled from skills that build the application's own parts — tables, views, pages, actions and workflows — and write React and JavaScript snippets from plain-language prompts.

Type what you want — "generate a page for a company selling cheese to British expats in Spain" — watch it render, then refine it conversationally: "remove any mention of blue cheese." Good for getting an application off the ground, and for accelerating the everyday changes.

The same copilot agent drives two interfaces: this quick chat, and the App Constructor for planning and building larger applications end to end.

The Saltcorn copilot in chat mode: a prompt to generate a page, the rendered result, and a follow-up refinement instruction — The copilot in chat mode — prompt, live preview, and a follow-up refinement, with past sessions on the left.

02 / AI workflows

AI as one step, not the whole show.

Drop generative AI into a deterministic, durable workflow — generate text, an image or structured JSON at a chosen point, with everything around it (loops, conditionals, forms, database writes) under your control.

Because each run is persisted with its context, an AI step can pause for human input and resume later — surviving beyond a single request. And since Saltcorn 1.4 the server can stream LLM output to the user token by token, over the same websocket that powers real-time collaboration.

The Saltcorn visual workflow editor: a Start node, named steps, a branching next-step formula, and the step configuration panel — The visual workflow editor — named steps, branching `next_step` formulas, and a per-step configuration panel. The AI step lives in an inspectable flow, not a black box.

· The idea

Generative AI is powerful, but unbounded. Saltcorn gives it walls to push against — a schema, roles, actions and audit — so what it produces is something you can actually run.

Every table, view and page carries its own minimum role; every action is logged; every workflow run is durable and resumable. An agent operating inside that structure inherits its guardrails — reading only the rows a user may see, writing only through actions you defined, leaving a trail behind it.

Structured, not sandboxed away The database is the ground truth the model works against

03 / AI agents

When the flow itself is model-driven.

When the execution path should be decided by the model, build an agent. An agent's behaviour is defined by the set of skills you enable — most add a tool to the inference loop; others shape the chat.

Skills you compose an agent from

Query a tableRetrieve rows by matching against any field.

HTTP requestCall an external API as a tool.

Expose a functionHand the model a JavaScript function you wrote.

Generate & run codeWrite and execute JS — you decide if it can touch tables or the network.

Long-term memoryStore and recall memories in a database table.

MCPConnect the agent to an external MCP server.

Model pickerLet the user switch inference model from a dropdown.

PreloadDataLoad rows into the system prompt before the run.

Action / workflow as toolExpose any Saltcorn action or workflow to the agent.

SubagentHand over to another agent with a different tool set.

Web searchSearch the internet for relevant information.

Plan approvalPresent a plan with an approve button; inject a prompt once accepted.

Two ways to run an agent.

Headless · on events

Attached to an event

Fire an agent on a table insert, an inbound API call or a schedule. You supply an initial prompt built from the triggering row, and the agent runs on its own.

Interactive · chat

The Agent chat view

Build a view on the Agent chat pattern for a ChatGPT-style interface — past chats listed on the left, and conversations shareable with other users. Voice in and out via ElevenLabs and device voice input.

04 / RAG & MCP

Documents in, tools out.

Ground agents and workflows in your own content, and interoperate with the Model Context Protocol in both directions.

RAG

Bring your documents in

Pull documents into agents and workflows through vector or full-text search, so answers are grounded in your data instead of the model's guesswork.

MCP server

Expose Saltcorn as tools

Publish your tables and actions as an MCP server, so other AI tools can read and act on your application through a standard protocol.

MCP client

Consume external servers

Point a Saltcorn agent at an external MCP server with the MCP skill, and its tools become part of your agent's repertoire.

05 / App Constructor

Plan the build, then execute it.

For larger applications, the App Constructor drives the copilot agent through the way software is normally built — describe, clarify, research, then plan a build as phases of ordered tasks and work through them, with a human in the loop at every stage.

Requirements are scored 1–5 by importance and are yours to edit; tasks are ordered by dependency and by entity type, so tables get built before the views that depend on them. Run one task or a whole phase, test, and the next phase adapts to your feedback. It can even self-heal — responding to a runtime error and fixing the build itself.

The App Constructor Phases tab: numbered phases, each with a list of requirements scored one to five stars, and a specification / research / phases / schema / feedback / errors tab strip — The App Constructor — phases of scored requirements, across a Specification · Research · Phases · Schema · Feedback · Errors workspace.

The eight stages

1 · DescriptionYou describe the app in as much detail as you like.

2 · ClarificationThe agent asks about anything unclear.

3 · ResearchIt searches the web — e.g. regulatory guidance for a regulated industry.

4 · RequirementsA scored 1–5 list you can edit, add to and re-rank.

5 · PlanningPhases of tasks, ordered by dependency and entity type.

6 · ExecutionRun a task or a phase, test, and feed results back in.

7 · User feedbackEnd-user suggestions become tasks once an admin approves.

8 · Self-healingResponds to system errors and fixes build problems itself.

06 / Predictive analytics

Not every model is an LLM.

Run classical machine learning directly on table data, with results stored as fields the rest of the application can use — in views, workflows and actions.

Regression

Predict a number

Fit a continuous outcome from the fields you already hold — a price, a score, an estimate — and write the prediction back to the row.

Classification

Sort into categories

Assign each row to a class — a risk band, a segment, a triage bucket — as a stored field available everywhere else.

Time series

Project forward

Analyse trends over time to forecast what comes next, and surface the result on the same dashboards as your live data.

Next in the platform

Try it now

Give AI a structure to build in.

Spin up Saltcorn and let the copilot draft your first app — free, open-source, on top of a real database you control.

Try it now Deploy anywhere · open source →

Want a hand?

Build it with us.

The Saltcorn team can design the agents, copilot skills and AI workflows your application needs.

See our services