Natural-language Telegram entry
Plain text can route to ask, do, autopilot, note, schedule, digest, or project switch without command memorization.
Telegram control for local Codex execution
Local AI work,
without leaving your machine.
OpenFish lets one trusted owner supervise local Codex work from Telegram. Commands, approvals, runtime state, and logs stay local.
Single-user by design
Local-first runtime
Telegram-native control
Operator Monitor
One owner, one live surface
Human-in-the-loop
Telegram
Intent routed
“Switch to ops project and check service health every 30 minutes.”
Route
schedule wizard
Natural language maps to ask, do, autopilot, note, schedule, or project switch.
Controls
approve
pause
takeover
The owner stays in control while the execution loop keeps moving.
Runtime
Local machine
Control plane
Telegram
Scope
Trusted owner only
Where it fits
Made for owners who want local execution and remote control.
OpenFish is not trying to be a hosted agent platform. It is a pragmatic control layer for running Codex on your own machine while staying reachable from a phone.
Designed for
Single-user workflows, project continuity, conservative local execution, and mobile-friendly supervision.
Not designed for
Multi-user bot hosting, public remote shells, or cloud orchestration layers.
Local-first privacy
Code, approvals, logs, and runtime state remain on your machine so the owner keeps operational control.
Autopilot with explicit takeover
Supervisor-worker loops can keep tasks moving, while the owner can pause, resume, stop, or take over at any point.
Isolated runtime option
Use `openfish docker-init` when you want a tighter execution boundary around AI work.
Execution loop
A human-visible control loop, not a black box.
OpenFish keeps the owner in the loop. Telegram is the surface. The local machine is where execution happens. Autopilot is just a structured supervisor-worker cycle with explicit intervention points.
01
Tell Telegram what you want
Use plain language like “switch to ops project” or “check service health every 30 minutes”.
02
OpenFish routes and stages execution
Intent routing selects ask, do, autopilot, note, schedule, or project switching with the right context.
03
Approve, pause, or take over when needed
The owner keeps authority over long-running work instead of losing visibility into the runtime.
Autopilot
Supervisor-worker architecture with explicit human controls.
Controls:
/pause, /resume, /stop, /takeover
A
Supervisor
Inspects worker output, decides the next move, and escalates when human input is required.
Loop
B
Worker
Executes the current task, returns output, and waits for the next instruction from the supervisor.
Human role
The human is not expected to drive every round manually. The human stays observable, interruptible, and in charge.
Quick start
Install in minutes, operate from Telegram.
OpenFish requires Python 3.12+. The packaged CLI covers install, configuration, checks, and startup.
Install:
Use PyPI for the fastest path.
Configure:
Run the packaged setup flow and check your environment before starting.
Operate:
Open Telegram, send `/home`, and use the live dashboard as the main control surface.
bash
# Install from PyPI
$pip install openfish# Bootstrap and start
$openfish install$openfish configure$openfish check$openfish start
>
System ready. Open Telegram and send /home.