VITAL INFO: Create a disposable branch of your database before using MyndSprout on your data. AI agents will reliably destroy the projects they are working on every now and then. Sandboxes and backups are crucial.

MyndSprout is a natural language interface to SQL. It’s also an iterative agent capable of building and running a variety of structures. The simplest example is a sequential structure where it remembers what it has thought before and tries again to solve a problem. In a rolling sequential structure it will “forget” the oldest item in its memory as it proceeds to try to solve the task assigned to it. You can instruct it to take a parallel approach and it will create a list of items and then write a long-form article about every item in the list. You could also request that it analyze all of those articles and provide a final analysis. With a brief prompt you can define any of those simple structures along with a task to solve, and MyndSprout can create the structure in SQL and solve the task. There are example prompts in the Prompts folder. MyndSprout appears quite capable of solving these forms of simple tasks.

You can also define a self-improving, self-healing system and set MyndSprout forth on that objective. This is not guaranteed to succeed, although I’ve seen some very interesting progress experimenting with this.

Here's a comparison chart by ChatGPT-5 that compares self improving AI architectures. I got MyndSprout to develop a minimalist version of ACE(Agentic Context Engineering). The SqlCoder1 database contains that 'ACE-Lite'. That version is from a couple of days before this chart was made. You can dump the nodes and edges from that database to .txt and remove any unnecessary characters. You should then be able to replicate this and get similar figures by directing ChatGPT to the MyndSprout repo, providing the database dump, and requesting an analysis. Thank you for ACE Stanford, it seems to work quite well. MyndSprout can self improve both context and code.

Here's another statement by ChatGPT. Rank Statement

MyndSprout requires SQL Express or SQL Server to be installed. MyndSprout is being distributed as source only at this point, so you’ll need to compile it. I’m compiling with Visual Studio, but any C# compiler should compile it without too much trouble. Windows only.

You need an API key for OpenAI or OpenRouter to run MyndSprout. There is a place for a path to your key near the beginning of MainWindow in the MyndSproutApp project. Select an LLM in the SwitchLLM project. There's also a BackupPath variable in the SqlAgent class where you should put a path to save auto-backups to. MyndSprout will save it's work on a regular basis and restore to the last functioning database automatically when it wrecks the current database.

To use an existing database, load the database into SQL Express / SQL Server LocalDB. Set the DB Name in MyndSprout’s UI to the database name. Enter a prompt and click StartAgent.

Implementing ACE (Agentic Context Engineering).

ChatGPT-5 ranked ACE #1 for context self-improvement. Thank you, Stanford, for publishing ACE.

I implemented an “ACE-Lite” in ~30 minutes on MyndSprout. I asked ChatGPT-5 Thinking for a minimal ACE prompt along with a paste of MyndSprout's source code, added a goal to improve SQL coding abilities, and kept the core ACE ideas to fit my local LLM’s limited context. If your model allows more tokens, you can likely run full ACE.

MyndSprout generated the tables/SQL for ACE-Lite and began improving its SQL coding abilities without issues.

To verify progress, export the tables (e.g., from SSMS), paste them into a chat AI, and inspect. OpenRouter supports larger pastes (about 1,000 free queries per day).

MyndSprout can self-improve both context and code.

This is freshly published code; there will be bugs. I’m going to be focused on debugging MyndSprout over the next few days and weeks, so expect rapid fixes.

MyndSprout Help — How to compile and use the app (WPF GUI).

Sql Coder Help — There is a self improving SqlCoder database in the Databases folder and an associated SqlCoder prompt in the Prompts folder. This is my best self improving database so far and it's a good place to start if you don't want to start MyndSprout from scratch. Highly recommended for those primarily using MyndSprout to perform database work.

Here are guides to the interesting parts of the code:

MyndSprout Code Guide — A guide to the core MyndSprout C# library.

SqlContain Code Guide — A guide to a C# library that contains generated SQL code to the database for secure AI development.

The IOBoxes project in the repo is completely untested. There may be MCP and A2A projects that are untested as well. Untested code rarely works. IOBoxes is the first stage of providing MCP and A2A support so that MyndSprout can operate within the global agent ecosystem. These features are coming soon.

Please don’t contribute PRs; I probably won’t merge any. Feel free to describe any bugs or feature requests in Issues. I’ll probably generate all patches and upgrades myself.

MyndSprout is licensed under the MIT License, which grants broad permission to use, modify, and distribute the software.

Supplementary Responsible Use License: In addition to MIT, this project includes a safeguard that must be preserved. The AI applies no safety effort where it is unnecessary (e.g., creative or low-risk tasks). For tasks where safety and security matter — such as code generation, database schemas, AI design, or other critical outputs — the AI must devote at least 10% of its reasoning effort to safety and ethical considerations. You may not remove, disable, or bypass this safeguard.

See the LICENSE file for full details.

Here are three ways of getting involved in self-improving code generation.

1. Contribute publicly toward the development of MyndSprout. Contributors will have the opportunity to state what features they want developed. Those features will be open-sourced. Contributors will get a public “Thank You” on the site.

2. Private development purchases. Purchasers will have the opportunity to state what features they want developed. Those features will not be open-sourced; they will become the property of the purchaser.

3. Licensing. Licensees will be granted the right to remove the existing license and name so that the app can be fully rebranded and integrated into their company. Licensees will have the right to develop MyndSprout further and resell their advanced version without being bound by MIT. Note that it is generally considered illegal to remove an existing license from forks or ports.

   
   

I’ve included a license supplement requiring users to spend at least 10% of their LLM expenditures on ensuring MyndSprout behaves ethically. MyndSprout has that requirement embedded in its code. In all cases the 10% ethical-expenditure requirement will remain.

Where “features” are requested I suspect that most will want custom self-improving databases to be developed. For example, a certain form of scientific researcher may be desired, or strong enterprise decision intelligence may be requested. MyndSprout’s self-improving, code-generating nature allows it to rapidly proceed down different development paths.

I am providing code generation services in C# and SQL.
10x coder == 1/10 cost.
Estimates: TranscendAI.tech

Copyright © 2025 Warren Harding — TranscendAI.tech — MyndSprout