plompy-engine.
EN/PL
GitHub

Small fish.
Big ideas.

Your coding agents, connected to Discord. Swimming through your tools, your context, your own machine.

Start your engineExplore the source
v0.12.0 · Self-hosted. Open source.
DiscordYou, from anywhere.let’s build something
plompy-engineOn your own Linux machine
Coding agentsCodex / Claude / OpenCode Your toolsGitHub / Browser / MCP Your contextMarkdown / Memory / Git and back to your thread
Follow the fish.
See how it works.A small engine. A whole ocean of possibility.

Close the laptop.
Keep the thread.

A message is just the beginning. Follow it from your phone to your own machine, and back.

An idea lands
in Discord.

Send a task to your private forum. Plompy checks who sent it, saves it to a durable queue, and starts the work on your Linux machine.

!work 30mA goal. A time budget. Room to work.
# ship-the-ideaIllustrative demo
You

!work 30m Add quantity validation to the API. Cover the edge cases with tests.

Plompy

Task accepted. I’ll inspect the project, make a focused change, and run the tests.

Saved to the queue.SQLite queued

Your agent.
A whole toolkit.

The saved runner session picks up your instructions and context. Inspect a repository, use a browser, read documentation. You choose the connections.

!goal 30m --repo "/srv/project"Add a fresh Codex supervisor to review the change.
Your chosen runnerCodex defaultClaude Code / OpenCode experimental
plompy-engineOn your Linux machine
GitHubYour repository
MarkdownYour instructions
BrowserYour connections

The result returns.
The context stays.

Read the result in the same thread. Ask a follow-up in the saved session. If work is interrupted, inspect its state and explicitly choose how to recover.

!resultRetrieve a saved answer without running the model again.
PlompyIllustrative demo

Added quantity validation and regression tests. Negative quantities are rejected; zero is still valid.

+ if (quantity < 0) {
+   throw new InvalidQuantityError();
+ }
Result saved. Session retained.
You

Nice. Now apply the same rule to the import.

Meet the tools

A whole ocean
of tools.

Your repositories. Your browser. Your sources. Connect the services you use to the runner you choose.

Explore integration setup
Preview setup for your runnerConfiguration preview
GitHub CLI01 / 10

GitHub

Inspect repositories, work with issues and prepare pull requests using a dedicated GitHub CLI configuration. You sign in with your own account.

Setup guide
Configuration command
npm run integrations -- add github --runner codex --yes

Add the integration, then configure the required account, key or local dependencies using the setup guide.

$ npm run integrations

One setup menu. Connections per runner. Your own accounts and API keys.

Make it
your kind of agent.

Give Plompy a voice, a working style and clear boundaries. Its context lives in plain Markdown files you can read, edit and keep in Git.

Browse the workspace templates
workspace/agent/Example configuration
MarkdownLocal files. Local Git history.

  

  
A little setup.
A lot of possibility.
Start with the offline demo. Then bring your own Discord bot, authenticated runner and a Linux machine.
Try the offline demo
git clone --branch v0.12.0 --depth 1 \
  https://github.com/dalkedev/plompy-engine.git
cd plompy-engine
npm --prefix apps/discord-agent ci --ignore-scripts
npm run build
npm run demo -- session
Synthetic data. No Discord token or model account needed. Requires Node.js 26.13.0+ (26.x LTS), Git and Linux.
Connect your first thread
Bot, runner, permissions. The complete setup.
Prefer a container?
The Docker path comes with Codex.
Give it a real goal
A worker, a reviewer, and a change to inspect.

    
Good to know before you go.
01
Is this a hosted AI service?
No. plompy-engine is MIT-licensed software you run yourself on Linux. You provide the machine, Discord bot and model access. A VPS is one way to host it; third-party services may have their own costs. Coding CLIs run on your host; model requests may go to an external provider.
02
Can I use any of the three coding agents?
Codex is the default. Claude Code and OpenCode adapters are experimental and have different execution boundaries. Supervised goals and still-image input currently require Codex. The published Docker image includes Codex only.
03
What happens if something stops?
Accepted work and results have durable state. Interrupted work remains visible for inspection and explicit recovery; uncertain external actions are not automatically replayed. Messages sent during a full outage may not have entered the queue. The public edition uses one shared worker.

  

  
Open source. Open invitation.MIT
Got an idea?
Make a splash.
Read the code. Try an integration. Open an issue. Build the agent you actually want to work with.
Find us on GitHub