Start a task in your terminal and walk away. AFK verifies its own work, texts you when it’s done, and leaves a receipt of every step. Unsure? It stops and asks.
You kick off a task and check back every few minutes. “Autonomous” that needs watching isn’t.
The agent says “done” but you still have to verify it actually worked.
Watching every step isn’t trust, it’s supervision. AFK is built for the gap between “go” and “done.”
Built around the ways agents actually break.
I built this because finished kept not meaning correct: the agent's own check approved the work, but the next stage expected that approval in a slightly different format; good work, rejected three times, nothing crashed. That's what real agent failure looks like.
Griffin Long, creator of Agent AFK
Fake done
Claims completion without verification: confident, unchecked, sometimes just wrong.
Context rot
Long runs decay into unreadable soup. The longer it works, the less anyone can follow.
Babysitting tax
You still supervise every few minutes. “Autonomous” that needs watching isn’t.
Familiar surface. Deeper machinery.
01Subagent orchestration
Most agents can spawn subagents but rarely do. AFK makes them the default. The agent you talk to in-session is the orchestrator, tasked with delegating work to subagents and coordinating the final result.
02Verification workflows
A second adversarial agent re-derives claims before they reach your diff: finished has to prove it is correct.
03Remote supervision
Leave the desk. Your phone buzzes when the run finishes or stalls.
Compare
See how Agent AFK stacks up
Side-by-side breakdowns vs. the tools you already know — grounded in the actual implementations.
One command, a full parallel pipeline: each skill fans out subagents, gathers results, and re-checks. That's the orchestration you'd otherwise ask for every time. These five see the most use; /forge makes more.
afk ›/review 642
parallel reviewers · one diff
✓correctnesslogic holds
!testserror path uncovered
✓securityinputs sanitized
⚠︎ChangesAdd a test for the error path·2 reviewers clean
A demonstration of the /review skill. Reviewing pull request 642: three reviewers run in parallel over one diff. Correctness (logic holds) and security (inputs sanitized) pass; the tests reviewer flags an uncovered error path. The findings converge into a single merge recommendation: request changes to add a test for the error path, with two reviewers clean.
/mint
End-to-end feature pipeline: spec → research → plan → parallelize → build → verify → heal → ship.
Ship-ready in one pass; pauses for approval.
/diagnose
Parallel hypotheses for bugs and failing tests.
Each validated in an isolated git worktree.
/shadow-verify
Adversarial re-derivation of sub-agent claims before you act.
Re-checked before the build pipeline.
/review
Parallel reviewers across a diff or PR: security, correctness, api-compat, tests, perf.
Synthesized by severity into a merge recommendation.
/ship
Release pipeline for finished local work: pre-flight, tests, commit, push, PR.
The REPL is a full TUI: structured tool-call visualization, input assistance, background job management, and live reasoning feedback. Not a log you scroll through — a frame you watch work.
Tool-lane overlay
Every tool call rendered inline: arguments, elapsed time, parallel ×N badges, and diff previews. A structured frame, not a log.
Ghost-text suggestions
LLM-powered autocomplete on the input line, with history-prefix fallback. Start typing and the REPL finishes the thought.
Background subagent jobs
Ctrl+B backgrounds a running turn. /bgsub dispatches new work. Keep prompting while agents run in parallel.
Shell passthrough
!cmd runs a shell command without leaving the REPL. !&cmd backgrounds it. No context switch.
Loop-stage visualization
A live Observe → Model → Choose → Act → Update rail in the footer. See exactly where the agent is in its reasoning loop.
One-keystroke mode switching
Shift+Tab cycles between default, plan, and bypass permission modes. Change the trust level without interrupting the conversation.
Extensibility
Hooks, MCP, plugins. Wire your own logic in.
Every stage of the agent lifecycle is extensible. Block a tool call with a hook, connect an MCP server, or install a plugin that ships its own skills. Coming from Claude Code? afk migrate brings your config along.
Lifecycle hooks
PreToolUse, PostToolUse, SessionStart, SessionEnd, SubagentStart, SubagentStop. Each hook can block, inject context, or fire custom logic.
MCP client
4-layer config: plugin → user → project → flag. Supports stdio, HTTP, SSE, and OAuth transports. Wire any MCP server in minutes.
Plugin architecture
Git-sourced or npm plugins contribute skills, MCP servers, and config. Project-scope plugins in .afk/ load per-repo.
One-command migration
afk migrate imports CLAUDE.md rules, custom commands, and project context from Claude Code. Switch without starting over.
Architecture
Three attach surfaces. One isolated root.
All three converge on ~/.afk/: config, plugins, plans, skills, session traces, memory. Nothing leaks into your global config.
Surface 01
CLI REPL
Interactive TUI with tool-lane visualization, ghost-text suggestions, background subagent dispatch, and one-keystroke permission cycling.
Surface 02
Daemon
Background loop. Runs unattended with bypass-permissions and plan persistence.
Surface 03
Telegram Bridge
Mobile attach. Send messages, tail output, interrupt from anywhere.
Three attach surfaces (the CLI REPL, the background daemon, and the Telegram bridge) all route into a single isolated root directory, ~/.afk/, which holds config, plugins, plans, skills, session traces, and memory.
◆
The model isn’t the product. The loop is.
Agent AFK operating contract, published in the repo
07pricing
Pricing
Free and open source. Pay for ongoing expertise, not features.
The runtime is Apache-2.0 and free: every feature, no unlocks. Paid plans are monthly support from the maintainer, not feature gates.
Open SourceApache-2.0
$0free, forever
The whole runtime: yours to run, read, fork, and ship. Nothing held back.