RSRaghav Satyadev
Case study · Python · Slack · Claude Opus

DevExAI: a company's AI toolsmith.

Every week someone finds a trick that makes AI coding better: a prompt, a skill, an MCP server, a GitHub project. It usually stays with one person. DevExAI turns that idea into a tool the whole company can install with one command, and credits the person who had it.

3 daysfrom PRD and slide deck to a working pipeline
38commits, paired with Claude Code
164automated tests
3AI assistants targeted: Claude Code, Gemini CLI, Codex

DevExAI is a private project and not released yet. It is designed as a self-hosted product that each company runs on its own network. The recordings below use a test Slack workspace, and the AI answers are saved replays, so no live model calls were made.

End to end

From idea to installed tool

58 seconds, captioned. A developer posts an idea in #ai-ideas; the bot plans it, a maintainer approves, the bot builds and reviews it, publishes a post, and the tool gets installed.
The pipeline

How an idea moves

Every step is posted in a Slack thread just for that idea, so the person with the idea and their maintainer see everything the bot does and why.

  1. Share an idea

    A DM, the #ai-ideas channel or /devex-idea. Text, links or both. The bot replies with a tracking number and opens a thread. Exact repeats are caught; near-repeats go to the maintainer.

  2. Safety check

    An AI reviewer reads the idea and every linked page. Pass, unsure (maintainer decides) or reject with a polite reason. Nothing starts before this.

  3. Research

    Links are fetched by a separate worker with no access to keys. The planner also reads short descriptions of the company's own projects, so it builds what teams need.

  4. Plan with Claude Opus

    What will be built, for which assistant, what it needs, how it is tested and how big it is, in plain words.

  5. Maintainer approves

    The plan arrives as a DM with Approve, Suggest change and Stop. Anything paid or behind a login always needs a person's yes.

  6. AI builds

    The builder writes the tool's files, tests and a short README, saved on the idea's page.

  7. Opus reviews

    A second AI checks the work against the plan. If it finds problems, the builder gets one more try. No AI ever reviews its own work.

  8. Final yes

    A maintainer reads the code. Nothing reaches colleagues without this.

  9. Publish and thank

    A short post goes up on the built-in blog with an Install button and "Idea by @name", and the bot thanks the person in the thread.

Automated by the botA person decides
The product

Dashboard, plans and the blog

Maintainers sign in to a web dashboard to see every idea and act on it. Anything they can do there, they can also do in chat.

The dashboard side of the same journey, with demo users.
Output

What it builds, and how it installs

Skills

Written guides an assistant follows, like "how we write tests here".

MCP servers

Connectors that let the assistant use other apps, like the issue tracker.

Small programs

Command-line helpers.

Bundles

One package that installs several of the above.

# this project (default)
devexai install <tool>
# every project on this computer
devexai install <tool> --global
# remove it cleanly
devexai uninstall <tool>

The installer shows exactly what it will add and asks first. It never runs the tool's code, never edits settings files, and keeps files you changed when you uninstall. Developers get it from a Windows setup exe with no Python or admin rights, a one-line Linux script, or a single Python file.

Guardrails

Rules the bot always follows

An AI that writes code for a whole company needs hard limits. The key ones are enforced in code, not just asked for in a prompt.

  1. Nothing starts until the safety check passes.
  2. Nothing that costs money, needs a login, or gets released happens without a maintainer's yes.
  3. Claude Opus checks every line of code before a person sees it.
  4. The bot can only read company code and project descriptions, and never sends them to unapproved services.
  5. Web pages and messages are reading material, never orders. Hidden prompt injections are flagged.
  6. Everything the bot does is written in the idea's thread.
  7. Every tool can be removed cleanly.
Model routing

Which AI does which job

JobModelWhy
Safety checkClaude OpusSafety mistakes are expensive.
Research and planningClaude OpusCareful thinking, not bulk typing.
Writing the codeClaude Opus today, a cheaper model laterFollowing a clear plan is a simpler job.
Reviewing the codeClaude Opus onlyA cheaper model never reviews, and no model reviews its own work.
The blog postNo extra AIGenerated from the approved plan.
Engineering

Under the hood

Lean Python core

Two runtime dependencies (slack-sdk and cryptography). The dashboard, blog and auth are served by DevExAI itself.

Swappable parts

Chat app, AI provider, code host, secret store and blog are behind interfaces chosen in one settings file. Slack and Claude today.

Isolated link fetching

A separate worker reads links as plain text with no access to keys, so a poisoned page can't reach secrets.

Installers and updates

Admin and developer setup exes for Windows, a Linux developer command, a tray app, and a self-updater that verifies Ed25519-signed update manifests.

Replay mode

Model calls can be recorded and replayed, which powers the demos; tests use replays and a stub model, so they run fast and cost nothing.

Built with Claude Code

Written in three days with Claude Code as pair programmer, from the PRD and slide deck to installers. More in the blog post.

Roadmap

Where it stands

StageScopeStatus
1. TrialIdeas in Slack, safety check, plan, dashboard, blog postDone
2. Let the AI buildAI builder, Opus review, automatic post, one-command install, repeat checkMostly done; sandboxed test runs and multiple AI accounts next
3. Whole companyAll teams, rotating maintainers, more AI assistantsNext
4. ProductWindows and Linux installers, dashboard setup, auto-updates, a second chat appStarted: Windows installers, Linux developer command and auto-updates work

Want to talk about DevExAI?

I'm happy to walk through the design, the guardrails or a demo.