CLI & TUI
afk chat drops you into a streaming Claude session with a polished terminal UI —
keybindings, plan mode, slash skills, ANSI-clean output.
Private beta · invite only
Autonomous Claude, in your terminal.
A native Claude runtime with a beautiful TUI, a long-running daemon, and a Telegram bridge. Seven orchestration skills turn loose ideas into shipped code while you're away from keyboard.
$
What it is
Three surfaces. One brain.
Agent AFK exposes the same Claude runtime through three deliberately separate channels — interactive, background, and remote — all sharing one configuration root and one skill registry.
Background daemon
Long-running agent loop with bypass-permissions automation, plan persistence, and live streaming back to any attached surface. Runs while you sleep.
Telegram bridge
Trigger runs, review diffs, and approve plans from your phone. Tail your daemon's progress in a private chat — no extra infrastructure required.
Seven orchestration skills
/mint
/diagnose
/shadow-verify
/forge
/parallelize
/forge-gate-check
/forge-l2-eval
Slash commands that actually do the work.
Each skill is a deterministic pipeline of sub-agents — not a single prompt. They compose spec-writing, parallel research, code generation, and adversarial verification into one call.
End-to-end feature pipeline: spec → research → plan → parallelize → build → verify → heal → ship.
Takes a feature idea or refactor scope and delivers a ship-ready, verified implementation in one pass. Pauses after spec for approval; resume with /mint --continue approved.
Parallel hypothesis generation and validation for bugs and failing tests.
Forks parallel research subagents, synthesizes 2–4 hypotheses, then validates each in an isolated git worktree. Use when behavior is unexplained or a test is red.
Adversarial re-derivation of sub-agent claims before you act on them.
Independently re-checks claims from a prior sub-agent before they reach the build pipeline. Auto-invoked by /diagnose when confidence is low or coverage is thin.
Generate new skills autonomously, gated by L1 capability evals.
Discovers gaps, drafts skill code, and runs a qualify loop (≤3 iterations). Approved skills are written under ~/.afk/. Gated by /forge-gate-check.
Transform a linear plan into dependency-aware parallel waves.
Restructures an implementation plan into parallel waves, respecting file-level and import-level dependencies. Used by /mint when ≥3 files are touched.
Report whether /forge is thawed; rerun the L1 eval harness.
Returns OPEN or CLOSED for the forge capability gate. Run before /forge to confirm the gate, or to recheck after a previous failure.
Run L2 capability evals — live sub-agent verdict probes.
Probes actual dispatch behavior with live sub-agent calls rather than static harness checks. Supplements the L1 gate for a complete capability picture.
Architecture
Three surfaces, one isolated config.
All three surfaces converge on ~/.afk/ — a single configuration root that holds
plugins, credentials, plans, and session state. Nothing leaks into your global Claude config.
Surface 01
CLI REPL
Interactive TUI. Streaming chat, plan mode, slash skills.
Surface 02
Daemon
Background loop. Runs unattended with bypass-permissions.
Surface 03
Telegram Bridge
Mobile control plane. Approve, abort, tail from anywhere.
~/.afk/
Plugins · credentials · session state · plans · skill registry
Access
Currently in private beta.
Agent AFK is invite-only while we tighten the orchestration layer with a small group of engineers. Drop your email and we'll reach out as seats open up.
No spam. One email when your seat is ready.
- Native Claude runtime
- Private by default
- Local-first · your code never leaves your machine
- macOS · Linux