Skip to content
pandudpnPublic

About

Your OMP coding agent on Telegram and WhatsApp. One engine for every chat platform: steer a running turn, answer its questions, approve its commands, on your own machine.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

ompwire

Your coding agent, wherever you already chat.
ompwire connects an OMP agent to the places you talk. You send a message, and the agent works on your own machine with your shell, your repos and your MCP servers.

Python 3.11+ MIT license

Install · Guide · Architecture · Contributing · Issues

chat platforms  ─>  ompwire  <->  omp --mode rpc-ui  ->  tools, skills, MCP servers

Ask from your phone, then approve the deploy when the agent asks. ompwire carries the conversation between a chat and OMP, and nothing else: the agent loop, the tools, the memory and compaction stay in OMP, so every platform gets the same agent.

Each platform plugs into one engine. Steering, questions, approvals, sessions and schedules are written once and work the same everywhere; a platform only decides how a message, a button or a file looks.

What you get

Every step, as it happens. The agent's text up to each tool call, then the call with its output, each as its own message and named after the tool it ran. Its todo list shows up as a checklist that is posted again each time an item moves. You read what it ran and what came back, not only the final answer.

Talk while it works. A message sent mid-turn is steered into the running turn by default (↪️ steered). /busy switches a chat to omp's follow-up queue, which runs it right after the turn, or to ompwire's own queue (📥 queued #2). /btw asks a side question that never enters the conversation.

Questions with buttons. When the agent asks with omp's ask tool, or ends a reply with choices, they arrive as buttons under the question. Tap one, type your own, or skip. Anything else you send meanwhile waits (⏸️ held #1) instead of derailing the question.

Approvals that matter. A shell command that can change the machine asks first: Allow, Allow for this session, Deny. git log, ls and rg run without a tap, so you are not trained to approve without reading.

Sessions that survive. Restart the gateway and the conversation is still there. A turn cut by the restart is resumed when the gateway comes back.

Context you can see. /usage shows the whole session's tokens, the context window, where OMP will compact it and when it last did. /compaction 70 changes the threshold from the chat.

Files both ways. Send a screenshot or a PDF to the agent. Ask for a chart and it arrives as a photo.

Scheduled turns. "Every morning at 7, summarise last night's commits." The agent proposes it, you confirm, and a systemd timer runs it and reports to the chat.

Install

You need Linux with systemd --user, Python 3.11+, and a way in to the platform you connect: a bot token for Telegram, a number of its own for WhatsApp. The installer brings omp if it is missing.

curl -fsSL https://raw.githubusercontent.com/pandudpn/ompwire/main/install.sh | bash

Then:

ompwire setup            # agent, model, thinking level, working directory
ompwire gateway start    # pick a platform, give it a credential, choose who may talk
ompwire doctor           # one line per check

gateway start asks only what it does not know yet. The first time it asks for the platform, a name for the gateway (it names the service and the logs), the credential (checked with the platform before anything is saved; on WhatsApp, a QR to scan) and the allow list, then runs the gateway under systemd. Run it again to restart a gateway, replace its credential, or add another profile on the same platform.

gateway update changes one later, on any platform: the groups it listens in (picked by name from the chats it has seen), the mode it answers in there, its model, working directory, approval mode, thinking level, and whether direct chats show reasoning and activity.

Each gateway is one connection with its own credential, allow list and systemd unit, so you can run several side by side. Only users on a gateway's allow list get an answer; everyone else gets silence. The guide has the per-platform steps.

Platforms

The goal is every place you already chat. Platforms are added one at a time behind the same engine, and ompwire gateway types lists what the installed version supports. The guide tracks which are ready and which are in progress.

A new platform is a surface plus a gateway. Adding a transport walks through it, and it is the most useful contribution right now.

Built for one person

ompwire runs your agent, with your credentials, on your machine. That shapes it:

  • one allow list per bot; strangers get no reply at all
  • groups and channels only when listed, each with its own mode (mention, all, read), and quiet there: the answer, not the tool trace
  • no inbound port: gateways reach out to their platform; one that needs a port has to say so in SECURITY.md first
  • the bot token never reaches the agent process, and logs mask it; a WhatsApp link stays in a 0600 file of its gateway
  • no telemetry: messages go to the platforms you connect and your model provider, nowhere else
  • ~/.omp stays OMP's; uninstall never touches it

The full threat model is in SECURITY.md.

Documentation

  • Guide: gateways and platforms, chat commands, approvals, schedules, configuration, troubleshooting
  • Architecture: how a turn flows, the RPC surface used, decisions and their costs
  • Rules: how to change this repo without breaking a running install
  • Security: threat model and private reporting

Contributing

Bug reports, ideas and pull requests are welcome.

  • Found a bug? Open an issue. The bug template asks for ompwire doctor output, logs and versions, which is usually all that is needed.
  • Have an idea? Open a feature request and say what you were trying to do. Ideas that need agent state belong in OMP itself, and that is fine to say in the issue.
  • Want to send code? Read CONTRIBUTING.md for setup and RULES.md for how changes are verified. Small fixes can go straight to a pull request; for anything larger, open an issue first so we agree on the shape.
  • Security problem? Do not open a public issue. Report it privately as SECURITY.md describes.
git clone https://github.com/pandudpn/ompwire.git && cd ompwire
make install && make test && make lint

License

MIT. See LICENSE.

About

Your OMP coding agent on Telegram and WhatsApp. One engine for every chat platform: steer a running turn, answer its questions, approve its commands, on your own machine.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

1 star

Watchers

0 watching

Forks

Releases

Used by

Contributors

Languages