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.

Is Agent AFK a Cursor alternative?

Not for editing. Cursor's tab completion, inline diffs, and VS Code ecosystem are genuinely excellent, and AFK has none of those. The real comparison is between Cursor's agent mode and AFK's agent runtime. On that axis they diverge substantially: AFK's 23 skills, DAG orchestration, worktree isolation, and enforced terminal-state contract are different in kind from Cursor's Cloud Agents. Most developers who care about agent capabilities run AFK in Cursor's terminal and get both.

Can I run Agent AFK inside Cursor?

Yes. Open Cursor's integrated terminal and run AFK there. AFK is a CLI tool that stores its state in ~/.afk/ and runs independently of any editor. You get Cursor's editing experience and AFK's agent orchestration simultaneously. They write to the same files; the same repo-hygiene rules apply as with any two tools sharing a working directory.

How do AFK's agent features compare to Cursor's Cloud Agents?

Cursor's Cloud Agents run in isolated cloud VMs with Slack, GitHub, and Linear triggers and a web dashboard. AFK's agent runtime runs locally: 23 auto-routed skills, DAG fan-out to 8 concurrent subagents in isolated git worktrees, enforced terminal-state verdicts with evidence pointers, cross-session SQLite memory, and a cron daemon for scheduled runs. Cursor's cloud setup is easier to start with. AFK's local setup gives you full control, no per-agent compute cost, and a more structured verification contract.

Why is Agent AFK free when Cursor costs $20-$200/month?

AFK is Apache-2.0 open source. The runtime is free; you pay your model provider directly at their published API rates. Cursor bundles model compute into a subscription, which is a convenience with a recurring cost. With AFK, every dollar you spend on inference goes directly to the provider at transparent rates, and you can read every line of the code handling it.

Does Agent AFK have an interactive mode or is it just a daemon?

AFK has a full interactive TUI. Tool calls render in structured lanes, ghost-text suggestions autocomplete slash commands and @file references, Ctrl+B backgrounds a running agent so you can start another task, and terminal-state verdict cards summarize each session outcome. There are 52 built-in slash commands. The daemon and unattended modes are one part of the surface; the interactive REPL is another.

Is Agent AFK better for backend and infrastructure work?

For tasks that are largely invisible to the eye: migrations, large refactors, test generation, deployment scripts, log triage. AFK's terminal-state contract and Telegram push are a natural fit. You describe the goal, walk away, and come back to a Done verdict with an evidence pointer. Cursor's strength is the interactive editing loop. The right tool depends on whether you want to be present for the work or receive a receipt when it is finished.

What is the open-source option for teams that can't use Cursor?

Agent AFK is Apache-2.0: you can audit, fork, and self-host the full runtime. For teams with data-handling policies that require inspecting what their agent tooling does at the source level, or that prefer not to route inference through a third-party backend, AFK is the straightforward open-source path. Cursor is proprietary and closed source.

Comparison

Agent AFK vs. Cursor

Cursor's tab completion is the best inline autocomplete available. Agent AFK's orchestration engine runs 23 skills, DAG-parallel subagents, and unattended task loops that end in a verified receipt. These are not competing editors. AFK runs in any terminal, including the one inside Cursor.

At a glance

Pick AFK for: goal delegation — hand it a task and get a verified receipt
Pick AFK for: DAG orchestration, worktree isolation, and a terminal-state contract
Keep Cursor for: best-in-class tab completion and IDE-native inline diffs
Run both: AFK runs in any terminal, including the one inside Cursor

The editor vs. the agent. You can have both.

Cursor is one of the best editors ever built. Tab completion that predicts multi-line changes in under 200 ms, inline diffs you accept with a click, a full VS Code extension ecosystem. If you spend most of your day editing code at the keyboard, that feedback loop is hard to beat. Agent AFK is not an editor. It is a coding agent runtime: you state a goal, it orchestrates subagents across isolated worktrees, runs tests, and ends every session in a named terminal state with an evidence pointer. Done means a commit SHA or a green test run, not a best guess. The real comparison is between Cursor's agent mode and AFK's agent runtime. On that axis the features diverge substantially. And because AFK is a CLI tool, you can run it in Cursor's integrated terminal right now, no configuration required.

Feature comparison

Grounded in the actual implementations, not marketing copy.

Dimension	Cursor	Agent AFK
Tab / inline autocomplete	Best-in-class. Cursor Tab predicts multi-line edits and next-cursor jumps in under 200 ms. Powered by a proprietary Fusion model trained with reinforcement learning. A genuine daily-editing productivity multiplier.	Not an editor. No inline autocomplete. AFK operates at the goal level, not the keystroke level. Run Cursor for editing; run AFK for delegation.
Inline diffs and visual review	Diffs appear inline in the editor. You accept or reject changes with a click, see type errors from the language server in real time, and watch UI changes render in the browser. The fastest feedback loop for interactive editing.	No inline diffs. AFK writes files to disk and produces a terminal-state verdict with evidence. Review happens in your editor of choice after the agent finishes, including Cursor.
Agent task execution	Cloud Agents (formerly Background Agents) run in isolated cloud VMs. Triggered from the Cursor UI, Slack, GitHub, Linear, or cursor.com/agents. Parallel agent fleets, Projects for persistent context, /loop mode, custom modes.	23 auto-routed skills including /mint (spec, plan, parallel build, verify), /diagnose (parallel hypothesis worktrees), /review (security, correctness, coverage, API compat in parallel), /ship, /fix-pr, and more. Skill bodies load only when they fire; no idle token tax.
How a task ends	A diff or PR appears in the UI. You decide when the task is complete based on what you see. No enforced terminal-state contract.	Every session ends in exactly one of four named states: Done, Blocked, Asking, or Interrupted. Done must carry an evidence pointer: a file path, commit SHA, or test result. This is enforced in the runtime, not in a prompt.
DAG orchestration	Parallel agent fleets via Cloud Agents. Coordination happens in cloud VMs managed by Cursor's infrastructure.	Local DAG execution with up to 8 concurrent subagents per fan-out, queued beyond the cap. Each subagent runs in an isolated git worktree. No cloud VM dependency; orchestration runs on your machine or any SSH target.
Interactive TUI	Chat and composer panels inside the IDE. Inline diffs, live agent steering, and Slack/GitHub notification hooks. Tightly integrated with the editor.	Full TUI: tool calls render in structured lanes, ghost-text suggestions autocomplete slash commands, Ctrl+B backgrounds a running agent, @file injects context, verdict cards summarize each terminal state. 52 slash commands in the box.
Cross-session memory	Memories feature (beta). Conversation history stored per session. No structured queryable fact store across sessions.	SQLite with FTS5 full-text search. Facts persist across sessions and context compaction, delivered via system prompt for high model adherence.
Async notification	Email, Slack, and cursor.com/agents dashboard. Cloud Agents can be triggered and monitored from web and mobile without the IDE open.	Telegram push: the terminal-state verdict lands on your phone. Reply to answer an Asking state, approve a pending action, or interrupt the agent.
Unattended / scheduled runs	Cloud Agents run asynchronously. Triggers include Slack, GitHub, Linear, and cursor.com/agents. Not designed for local cron scheduling or programmatic invocation from external scripts.	Built-in cron daemon persists across reboots. Full programmatic access from any shell. Pipe AFK into whatever CI or automation you already have. Runs locally, over SSH, or on any machine with a terminal.
Model providers	Claude, GPT, Gemini, and Cursor's own models. Routing managed by Cursor; BYOK available on paid plans, but usage flows through Cursor's infrastructure.	Anthropic, any OpenAI-compatible endpoint, local models via Ollama or LM Studio. Requests go directly from your machine to your provider. Each pipeline step can target a different model or cost tier.
IDE / editor integration	Cursor IS the editor: a full VS Code fork. All extensions, themes, and keybindings transfer. If your team is already on VS Code, the migration cost is close to zero.	Editor-agnostic CLI. Works alongside Cursor, VS Code, Vim, or any other editor. Run it in Cursor's integrated terminal to combine Cursor's editing experience with AFK's agent capabilities.
License	Closed source. Proprietary Anysphere product.	Apache-2.0. Fully open source. No feature gates.
Pricing	Subscription: Hobby free (limited), Pro $20/mo, Pro Plus $60/mo, Ultra $200/mo, Teams $40-$120/user/mo. Credit pools deplete with heavy agent use; on-demand overage billed in arrears.	Free. Bring your own API key. Pay your model provider directly at published rates. No seat fees, no credit pools, no overage charges.

Install freenpm i -g agent-afk · free, Apache-2.0

Where Agent AFK pulls ahead

Genuine architectural differences, not marketing claims.

Agent runtime, not just agent mode

23 skills ship with AFK and auto-route by task shape. /mint runs a full spec-plan-build-verify loop, pausing for approval before it writes. /diagnose fans out parallel hypotheses into isolated git worktrees. /review dispatches subagents across security, correctness, API compatibility, and test coverage simultaneously. /fix-pr pulls review comments and pushes the fix. Skills load their full bodies only when they fire. Irrelevant skills cost zero tokens.

Every task ends in a verified receipt

Done, Blocked, Asking, or Interrupted: one of four, every time. Done requires an evidence pointer: a file path, a commit SHA, or a passing test run. This contract is enforced in the runtime, not in a prompt, so no model drift or configuration change can remove it. The verdict arrives on your phone via Telegram. You wake up to a receipt, not a question of whether anything happened.

DAG orchestration with worktree isolation

AFK's DAG engine fans work out to up to 8 concurrent subagents, each running in its own isolated git worktree. Subagents share findings via a structured workspace and merge results back to the parent session. Overflow queues instead of deadlocking. The entire orchestration runs on your machine: no cloud VM, no per-agent compute charge.

Free and fully open source

AFK is Apache-2.0. The runtime is free; you pay your model provider directly at their published rates. All 23 skills, the cron daemon, Telegram integration, cross-session SQLite memory, DAG orchestration: all free, no feature gates, every line auditable. Cursor is proprietary and subscription-priced from $20 to $200 per month.

Where Cursor pulls ahead

Honest. Cursor is a genuinely excellent tool with real strengths.

Best-in-class tab completion

Cursor Tab is the fastest and most context-aware inline autocomplete available today. It predicts multi-line edits, next-cursor jumps, and cross-file changes in under 200 ms, powered by a proprietary Fusion model trained with reinforcement learning. If you spend most of your day editing code at the keyboard, nothing comes close. AFK has no inline autocomplete at all; these are different surfaces.

IDE-native visual editing loop

Cursor is the editor. Diffs appear inline, you accept or reject with a click, the language server surfaces type errors in real time, and you can watch a UI change render in the browser the moment it happens. For frontend iteration on layouts and styles, the interactive feedback loop Cursor provides is fundamentally faster than any CLI agent.

VS Code ecosystem, zero migration cost

Cursor is a VS Code fork. Your extensions, themes, keybindings, and muscle memory transfer in one click. If your team is already on VS Code, adopting Cursor costs almost nothing. AFK is a CLI tool with its own learning curve: you need to think in goals and terminal states rather than keystrokes and inline diffs.

Cloud Agents and multi-platform triggers

Cursor's Cloud Agents run in isolated cloud VMs and integrate with Slack, GitHub, and Linear out of the box. For teams that want to trigger agent work from a Slack message or a GitHub event without managing their own infrastructure, Cursor's cloud-native setup is simpler to get started with than a self-hosted local daemon.

Frequently asked questions

Is Agent AFK a Cursor alternative?: Not for editing. Cursor's tab completion, inline diffs, and VS Code ecosystem are genuinely excellent, and AFK has none of those. The real comparison is between Cursor's agent mode and AFK's agent runtime. On that axis they diverge substantially: AFK's 23 skills, DAG orchestration, worktree isolation, and enforced terminal-state contract are different in kind from Cursor's Cloud Agents. Most developers who care about agent capabilities run AFK in Cursor's terminal and get both.
Can I run Agent AFK inside Cursor?: Yes. Open Cursor's integrated terminal and run AFK there. AFK is a CLI tool that stores its state in ~/.afk/ and runs independently of any editor. You get Cursor's editing experience and AFK's agent orchestration simultaneously. They write to the same files; the same repo-hygiene rules apply as with any two tools sharing a working directory.
How do AFK's agent features compare to Cursor's Cloud Agents?: Cursor's Cloud Agents run in isolated cloud VMs with Slack, GitHub, and Linear triggers and a web dashboard. AFK's agent runtime runs locally: 23 auto-routed skills, DAG fan-out to 8 concurrent subagents in isolated git worktrees, enforced terminal-state verdicts with evidence pointers, cross-session SQLite memory, and a cron daemon for scheduled runs. Cursor's cloud setup is easier to start with. AFK's local setup gives you full control, no per-agent compute cost, and a more structured verification contract.
Why is Agent AFK free when Cursor costs $20-$200/month?: AFK is Apache-2.0 open source. The runtime is free; you pay your model provider directly at their published API rates. Cursor bundles model compute into a subscription, which is a convenience with a recurring cost. With AFK, every dollar you spend on inference goes directly to the provider at transparent rates, and you can read every line of the code handling it.
Does Agent AFK have an interactive mode or is it just a daemon?: AFK has a full interactive TUI. Tool calls render in structured lanes, ghost-text suggestions autocomplete slash commands and @file references, Ctrl+B backgrounds a running agent so you can start another task, and terminal-state verdict cards summarize each session outcome. There are 52 built-in slash commands. The daemon and unattended modes are one part of the surface; the interactive REPL is another.
Is Agent AFK better for backend and infrastructure work?: For tasks that are largely invisible to the eye: migrations, large refactors, test generation, deployment scripts, log triage. AFK's terminal-state contract and Telegram push are a natural fit. You describe the goal, walk away, and come back to a Done verdict with an evidence pointer. Cursor's strength is the interactive editing loop. The right tool depends on whether you want to be present for the work or receive a receipt when it is finished.
What is the open-source option for teams that can't use Cursor?: Agent AFK is Apache-2.0: you can audit, fork, and self-host the full runtime. For teams with data-handling policies that require inspecting what their agent tooling does at the source level, or that prefer not to route inference through a third-party backend, AFK is the straightforward open-source path. Cursor is proprietary and closed source.

Compare other tools

See how Agent AFK compares across the landscape.

Try Agent AFK

Open source, Apache-2.0, bring your own model. Run it in any terminal, including Cursor's.

Install free Read the docs View on GitHub