The short answer
If your workflow is built around Pi, TelePi is the better fit. Pi gives you tighter context control, a more open command surface, and a session model that makes hand-off between CLI and Telegram especially clean.
If your workflow is built around Codex, TeleCodex is the better fit. It does more to turn Codex into a mobile-optimized supervision loop: launch-profile switching, live plan visibility, file and image exchange, built-in auth from Telegram, and clearer controls for sandboxed work from your phone.
"TelePi is closer to a Telegram face for a context-first harness. TeleCodex is closer to a best-in-class mobile control surface for Codex."
Feature-by-feature comparison
| Dimension | TelePi | TeleCodex |
|---|---|---|
| Base runtime | Pi AgentSession SDK with JSONL session files and a deliberately minimal coding harness. |
@openai/codex-sdk with Codex threads, SQLite-backed state, and Codex CLI subprocess execution. |
| Best use | Context-disciplined coding workflows where you want command discovery, session branching, and broad model freedom. | Phone-first Codex supervision with strong controls around launch mode, thread visibility, and mobile-friendly inputs. |
| Handoff model | Strongest CLI ↔ Telegram roundtrip because the session file is the source of truth and /handoff exists in Pi CLI. |
Strong Telegram → CLI handback with codex resume <id>; ideal once the Codex thread is already in flight. |
| Mobile controls | Excellent inline keyboards and command picker, plus Pi slash-command forwarding and extension dialog support. | Richer mobile supervision layer with launch profiles, reasoning-effort control, live todo updates, and attachment workflows. |
| Voice / ASR | Parakeet CoreML, Sherpa-ONNX Parakeet, or Whisper. Strongest on-device story across Apple Silicon and Intel Mac setups. | Parakeet CoreML or Whisper. Great for quick spoken corrections and mobile continuation. |
| Files and images | Mainly prompt and session oriented; stronger on command/session control than on attachment exchange. | Stronger attachment story: documents in, generated artifacts back out, plus direct photo input for screenshots. |
| Profile switching | Model switching is strong; launch profiles are not the center of the product. | One of the key differentiators. /launch makes sandbox and approval choices practical on mobile. |
| Workflow depth | Better if you want Telegram to expose more of the underlying harness surface, including discovered commands and branching. | Better if you want Telegram to be an efficient supervisor UI for real Codex work already happening in local threads. |
Choose TelePi if your bottleneck is context control
- You already prefer Pi as a harness. That is the main answer. TelePi makes the most sense when Pi’s context discipline and model-agnostic structure are already why you are there.
- You want Telegram to expose more of the harness surface. TelePi’s command picker, discovered slash commands, branching, and labeling make it feel like an extension of Pi rather than just a remote chat shell.
- You care about local/offline ASR options. TelePi has the strongest local transcription story because it supports both Parakeet CoreML and Sherpa-ONNX Parakeet, not just cloud fallback.
Choose TeleCodex if your bottleneck is mobile supervision
- You want the cleanest mobile UX for Codex. TeleCodex is where I pushed harder on profile selection, effort tuning, live plan display, and attachment handling.
- You review work from your phone. Seeing plan progress, file changes, tool calls, screenshots, and generated artifacts matters more in TeleCodex than it does in TelePi.
- You need safer mode switching from Telegram. Launch profiles turn sandbox and approval choices into a deliberate mobile UI instead of a flag-typing problem.
The bigger pattern
The deeper point is that both projects are arguing for the same thing: coding-agent workflows should not be trapped in a terminal or desktop app just because the runtime lives there. Once the architecture pass is done, a lot of the work becomes dispatch, review, and correction. Telegram is surprisingly good at that layer.
If you want the broader essay version of that idea, read AI, the UI Anywhere. If you want the individual build stories, read the TelePi deep dive and the TeleCodex deep dive.