Agent AFK is an open-source AFK coding agent designed to run tasks while you are away. It plans, acts, verifies its own work, and notifies you when it finishes or needs input. Free, Apache-2.0, works with any model provider.

What does AFK mean for a coding agent?

AFK stands for Away From Keyboard. An AFK coding agent is designed to keep working after you walk away; it runs a task end to end, verifies the result, reaches a clear stopping point, and texts you when done.

How does Agent AFK notify me when it finishes?

Agent AFK sends a Telegram message when a run finishes, stalls, or needs your input. You can answer, redirect, or interrupt from your phone. Every run also writes a full trace to disk so you can see exactly what happened.

Can Agent AFK replace Codex?

For terminal-first developers who want model freedom: yes. AFK covers the same terminal agent workflow, adds a verdict contract, 23 skills, SQLite memory, and Telegram push. The gaps: no desktop app, no IDE extension, no cloud agents, no mobile. If you need those surfaces or are deep in the OpenAI ecosystem, Codex is the stronger fit.

No. It requires a ChatGPT subscription ($20+/mo for meaningful use). If you already pay for ChatGPT, agent access is bundled. That quota is shared across your entire ChatGPT account. AFK is Apache-2.0 and free; bring your own API key.

Which has better benchmarks?

Codex leads Terminal-Bench 2.1 at 88.8% (GPT-5.6 Sol) and scores 64.6% on SWE-Bench Pro. AFK has no published Terminal-Bench score; results depend on the configured model. If benchmark leadership on terminal tasks matters, Codex wins today.

What about the shared quota problem?

Codex quota is shared across your ChatGPT account: chat, images, Sora, everything. A heavy usage day competes with your agent runs. AFK uses a separate API key. Agent budget is fully isolated from any other product.

Can I use Claude or local models with Codex?

No. Codex routes exclusively through OpenAI: GPT-5.6 Sol default, GPT-6 Sol/Luna as of Sep 2026. AFK supports any Anthropic or OpenAI-compatible provider, including local Ollama/LM Studio.

How does unattended work compare?

Codex Web runs cloud agents on OpenAI infra with GitHub/GitLab/Slack triggers. AFK runs a local cron daemon with Telegram push. Codex survives a closed laptop; AFK requires your machine to be on. AFK sends structured verdicts to your phone; you can reply from Telegram to unblock a stuck task.

Which should I choose?

Choose Codex if: you already pay for ChatGPT, need multi-surface coverage, want cloud agents, or need enterprise SLA. Choose AFK if: you want any model provider including local, a verdict contract, 23 auto-routed skills, cross-session memory, Telegram push, and no subscription.

Comparison

Agent AFK vs. Codex

Both are terminal-native coding agents under Apache-2.0. Codex brings GPT-6 models, 126K GitHub stars, a desktop app, IDE extensions, and a ChatGPT bundle. AFK brings model freedom, a verdict contract, and no subscription. The choice is ecosystem lock-in versus freedom.

Same category, different bets

Codex and AFK occupy the same category: autonomous terminal coding agents that take a goal, execute tool calls, and return a result. The bets diverge. Codex bets on the OpenAI stack: GPT-6 Sol and Luna (September 2026), a Rust core, OS-native sandboxing (Seatbelt on macOS, Landlock + seccomp on Linux), and ChatGPT integration across CLI, desktop, IDE, and web. That integration is a genuine advantage if you already pay for ChatGPT. It is a hard constraint if you want to route tasks to Claude, Gemini, or a local model. AFK bets on freedom: any provider that speaks Anthropic or OpenAI format, including local inference. A terminal-state contract (Done / Blocked / Asking / Interrupted) is enforced in runtime code, not a prompt instruction. 23 skills auto-route by task shape. A cron daemon pushes Telegram verdicts to your phone. Codex quota is shared across your entire ChatGPT account. AFK quota is whatever you purchase directly from the provider.

Feature comparison

Grounded in the actual implementations, not marketing copy.

Dimension	Codex	Agent AFK
Model access	OpenAI only. GPT-6 Sol/Luna (Sep 2026), GPT-5.6 Sol default. No Claude, Gemini, or local.	Any Anthropic or OpenAI-compatible provider, including local via Ollama/LM Studio. Each step can target a different tier.
Pricing	ChatGPT Plus $20/mo, Pro $200/mo, Business $25/user/mo. Quota shared with chat, images, Sora.	Free. Apache-2.0, BYOK. Pay your provider at published rates. No seat fees, no shared quota.
How a turn ends	Free-form output. You determine whether the task succeeded.	Done (with evidence pointer), Blocked, Asking, or Interrupted. Enforced in runtime code.
Benchmarks	Terminal-Bench 2.1: 88.8% (GPT-5.6 Sol). SWE-Bench Pro: 64.6%. 2-4x token efficiency vs Claude Code.	No published Terminal-Bench score. Results depend on configured model. AFK provides orchestration; benchmarks vary by provider.
Skills / automation	Plugin marketplace with community extensions. No published first-party skill count.	23 auto-routed skills: /mint, /diagnose, /review, /ship, /fix-pr, /refactor, /shadow-verify, and more. All open source.
Interactive TUI	/tui fullscreen (v0.156), async inline questions, voice input GA. No tool-lane overlay.	52 slash commands, ghost-text, tool-lane overlay with timing/parallel badges, loop-stage rail, verdict cards, shell passthrough, @file injection, Ctrl+B backgrounding.
DAG orchestration	Worktrees GA in v0.156. Parallel agent workflows supported.	Up to 8 concurrent subagents in isolated worktrees. Structured workspace for findings. Ctrl+B and /bgsub manage live subagents.
Sandboxing	OS-native: macOS Seatbelt, Linux Landlock + seccomp. Rust core. Among the most robust in the field.	Path-containment by default. Permission modes via Shift+Tab (default/plan/bypass). No OS-level sandbox.
Unattended work	Codex Web: cloud agents on OpenAI infra. GitHub/GitLab/Linear/Slack triggers. Survives closed laptop.	Local cron daemon with Telegram push. Verdicts land on your phone. Reply to unblock. Requires machine to be on.
Cross-session memory	AGENTS.md for per-repo instructions. /usage dashboard for cost tracking.	SQLite with FTS5 full-text search. Facts queryable across sessions. HOT.md for hot-path context.
Multi-surface ecosystem	CLI + desktop app + IDE extension (14.7M installs) + Codex Web + ChatGPT mobile. One account.	Terminal-only CLI. Non-terminal surface is Telegram for notifications and reply loop.
Release cadence	1,125+ releases, near-daily. v0.156 shipped Sep 22, 2026. 126K GitHub stars.	Active single-maintainer development. Apache-2.0. Fully auditable, no feature gates.
Enterprise	Gartner MQ Leader 2026. 20M active users (Aug 2026). Enterprise procurement and SLA.	One maintainer. No SLA. No enterprise procurement path.
Migration	No published migration from Claude Code or other agents.	afk migrate imports CLAUDE.md rules and project context from Claude Code in one command.

Where Agent AFK pulls ahead

Genuine architectural differences, not marketing claims.

Model freedom, not OpenAI lock-in

Codex routes every request through OpenAI. AFK supports any provider: Claude, GPT, Gemini, Grok, Mistral, or local inference. Each step in a multi-agent workflow can target a different model or cost tier. No feature is gated behind a specific provider.

Free with no shared quota

Codex requires a ChatGPT subscription ($20+/mo), and that quota is shared with chat, images, and everything else in your account. AFK is Apache-2.0, BYOK. Agent budget is fully separate from any other product.

Verdict contract in the runtime

Every turn ends in Done (with evidence), Blocked, Asking, or Interrupted. Enforced in runtime code, not a prompt instruction. When you review an overnight run, you see verdicts with evidence pointers, not prose to reconstruct from.

23 auto-routed skills in the box

/mint runs spec-plan-build-verify. /diagnose fans parallel hypotheses into isolated worktrees. /review dispatches across security, correctness, API compat, and coverage. Auto-routing fires by task shape. All 23 are open source and extensible.

Telegram push with reply loop

Cron-scheduled tasks push the terminal-state verdict to your phone via Telegram. Reply to answer an Asking state or interrupt a running task. Delegate before bed, wake up to a receipt with evidence.

Cross-session SQLite memory

SQLite with FTS5 full-text search. Facts survive compaction and are queryable across sessions. HOT.md for hot-path context, AFK.md for project conventions. The agent recalls decisions made weeks ago without being told.

Where Codex pulls ahead

Honest. Codex has real strengths that Agent AFK does not match.

Multi-surface ecosystem

CLI, desktop app (macOS/Windows), IDE extension (14.7M installs), Codex Web, ChatGPT mobile. One account across all surfaces. AFK is terminal-only.

ChatGPT bundle value

If you pay for ChatGPT Plus ($20/mo), Codex is bundled at no extra cost. The marginal cost to add an agent is zero for existing subscribers.

Terminal-Bench leadership and token efficiency

88.8% on Terminal-Bench 2.1 (GPT-5.6 Sol), the highest published score. 2-4x fewer tokens per task vs Claude Code. For cost-sensitive, repetitive workloads that gap is real.

Cloud agents that survive a closed laptop

Codex Web runs agents on OpenAI infrastructure with GitHub, GitLab, Linear, and Slack triggers. AFK requires your machine to be on. For cloud-resident scheduled work, Codex is more reliable.

GPT-6 models and near-daily releases

GPT-6 Sol and Luna available Sep 22, 2026. 126K stars, 1,125+ releases. v0.155-0.156 shipped voice, /tui fullscreen, /usage, and worktrees GA within weeks.

OS-native sandboxing

macOS Seatbelt and Linux Landlock + seccomp for OS-level process isolation. Rust core. One of the most hardened sandbox implementations among terminal coding agents.

Enterprise infrastructure

Gartner MQ Leader 2026. 20M active users (Aug 2026). Enterprise procurement, SLA, compliance documentation. AFK has one maintainer and no SLA.

Frequently asked questions

Can Agent AFK replace Codex?: For terminal-first developers who want model freedom: yes. AFK covers the same terminal agent workflow, adds a verdict contract, 23 skills, SQLite memory, and Telegram push. The gaps: no desktop app, no IDE extension, no cloud agents, no mobile. If you need those surfaces or are deep in the OpenAI ecosystem, Codex is the stronger fit.
Is Codex free?: No. It requires a ChatGPT subscription ($20+/mo for meaningful use). If you already pay for ChatGPT, agent access is bundled. That quota is shared across your entire ChatGPT account. AFK is Apache-2.0 and free; bring your own API key.
Which has better benchmarks?: Codex leads Terminal-Bench 2.1 at 88.8% (GPT-5.6 Sol) and scores 64.6% on SWE-Bench Pro. AFK has no published Terminal-Bench score; results depend on the configured model. If benchmark leadership on terminal tasks matters, Codex wins today.
What about the shared quota problem?: Codex quota is shared across your ChatGPT account: chat, images, Sora, everything. A heavy usage day competes with your agent runs. AFK uses a separate API key. Agent budget is fully isolated from any other product.
Can I use Claude or local models with Codex?: No. Codex routes exclusively through OpenAI: GPT-5.6 Sol default, GPT-6 Sol/Luna as of Sep 2026. AFK supports any Anthropic or OpenAI-compatible provider, including local Ollama/LM Studio.
How does unattended work compare?: Codex Web runs cloud agents on OpenAI infra with GitHub/GitLab/Slack triggers. AFK runs a local cron daemon with Telegram push. Codex survives a closed laptop; AFK requires your machine to be on. AFK sends structured verdicts to your phone; you can reply from Telegram to unblock a stuck task.
What is the verdict contract?: Every AFK turn ends in Done (with evidence pointer), Blocked, Asking, or Interrupted. Enforced in runtime code, not a prompt instruction. When reviewing scheduled runs, you see verdicts with evidence, not prose to reconstruct from.
Which should I choose?: Choose Codex if: you already pay for ChatGPT, need multi-surface coverage, want cloud agents, or need enterprise SLA. Choose AFK if: you want any model provider including local, a verdict contract, 23 auto-routed skills, cross-session memory, Telegram push, and no subscription.

Compare other tools

See how Agent AFK compares across the landscape.

Try Agent AFK

Apache-2.0, bring your own model. No subscription, no shared quota, no lock-in.

Install free Read the docs View on GitHub