Power your agent
คำสั่งของผู้เผยแพร่เอง ไม่ใช่เอนด์พอยต์ของ PlugMCP
Host a copy with our keys and logs.
npx -y [email protected]Claude Code ไม่มีลิงก์ติดตั้งสำหรับคำสั่งแบบโลคัล ให้คัดลอกคำสั่งด้านล่างแล้วรันในเทอร์มินัล
claude mcp add agent-afk -- npx -y agent-afk@5.354.0รันในเทอร์มินัล
{ "mcpServers": { "agent-afk": { "command": "npx", "args": [ "-y", "[email protected]" ] } }}claude_desktop_config.json
URL แบบรีโมตหรือคำสั่งบนเครื่องของคุณ? การติดตั้งสองแบบต่างกันอย่างไร
รันบน PlugMCP
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
เปิดจากผู้เผยแพร่
Start a task and walk away. AFK builds the feature, verifies its own work, and texts you when it's done. You come back to a receipt of every step, every decision, and what's waiting on you.
Open source. Runs locally. Works with any model provider, including local models. Your keys, your machine, your rules.
Open-source coding-agent harness you can actually change — own the loop (prompts, gates, routing, skills, terminal states), use any model, run long tasks while you're away.
⭐ Like the idea of an agent you can walk away from? Star the repo -- it's the fastest way to help other people find it.
npm install -g agent-afk # Node 22.22.2–22.x, 24.15–24.x, or 26+ (23.x and 25.x excluded)
afk login # authenticate (auto-detects Claude Code / Codex creds)
afk doctor # verify everything works
afk chat "hello" # first real conversation
# Preview a behaviour change before applying it
afk whatif --append "Always ask a clarifying question before using tools." --verify
That's it. You're in.
Using pnpm? pnpm add -g agent-afk works, but pnpm 10 and later block build scripts by default and skips postinstall (with an easy-to-miss warning — see #2199). Run pnpm approve-builds -g after install to approve the hook, or use npm install -g agent-afk instead. If an afk daemon is already running and pnpm skipped the hook, restart it manually with afk service restart daemon — on macOS the hook restarts the daemon automatically when it runs, but on Linux you must always run this command after an upgrade even when the hook ran. If you have a manually-started Telegram bot, also run afk telegram restart.
Try without installing: npx agent-afk chat "hello" runs a one-shot turn with zero global install.
Already using Claude Code or Codex? afk login will detect your existing credentials automatically. Run afk migrate to import your plugins, skills, and MCP servers too -- it live-reads the source tool's dirs, so anything you install there keeps showing up in AFK with no re-run.
Codex imports use CODEX_HOME when set (otherwise ~/.codex) and also discover shared skills in ~/.agents/skills. Native .codex-plugin/plugin.json manifests are recognized alongside Claude-format manifests; explicit [plugins."name@marketplace"] enable/disable settings apply at runtime. Without an installed-plugin registry, discovery selects one cached directory per marketplace/plugin using descending numeric-aware directory order. This is a filesystem fallback, not a guarantee of the version selected by Codex. Codex-hosted app connectors and TOML MCP server loading are not imported.
📖 Full documentation at docs.agentafk.com -- quickstart, configuration, model setup, surfaces, skills, and SDK reference.
afk chat "..." for one-shot, afk for a REPL with full tool access (Bash, file ops, web fetch, grep/glob, subagents).afk daemon runs headless. Pair it with send_telegram and you get pings on your phone when work lands in a terminal state.afk trace show prints it back as a human-readable receipt — tool calls, gate decisions, subagent lifecycles, cost — so you can audit a run without reaching for jq.afk telegram setup walks you through bot token + allowlist. After that you have a private chat surface backed by the same session manager as the REPL./mint, /diagnose, /spec, /research, /ship, /review dispatch subagent waves. /mint takes a feature idea and runs spec → research → plan → parallelize → build → verify → ship. /diagnose forks parallel root-cause hypotheses for failing tests and bugs.mode:'background'; results auto-deliver into the model's context when they finish. /bgsub lists running and completed jobs, /bgsub:join <id> replays a result manually.afk whatif behavioural impact predictor. Before changing a prompt, model, skill, or memory rule, predict how the agent's behaviour will shift without touching your real config. Run afk whatif --append "Always ask a clarifying question before using tools." --verify and the engine diffs the system prompts, predicts labelled behaviour shifts, then optionally verifies empirically by running sandboxed episodes and measuring rates. No real writes happen during the run. Source: src/whatif/, docs/whatif.md.
Jev calibrated judgment. Jev is an external cross-family judge you configure as an MCP server in ~/.afk/config/mcp.json. When present, afk whatif --judge auto selects Jev to grade verification episodes; it returns calibrated probabilities and removes Anthropic self-preference bias in verdicts. Falls back to Claude when Jev is not configured. Use --judge claude to keep all episode data within Anthropic. Source: src/whatif/judge/jev-connect.ts, docs/whatif.md.
Constructed SW(29,16), the last open symmetric weighing matrix case for weight 16 listed by Rosin (arXiv:2505.23881). Given a research campaign goal, agent-afk agents picked the target, designed a prescribed-automorphism SAT search, found a symmetric 29x29 weighing matrix in 5620 seconds of solver time, checked it with independent exact checkers on two machines, and ran a kill lane that could not refute it. The standard databases list no symmetric W(29,16). Check it yourself in under a second: docs/results/sw-29-16.
Constructed SW(27,16), a symmetric weighing matrix the standard archives do not list. Given a research campaign goal, agent-afk agents picked the target, designed a prescribed-automorphism SAT search, found a symmetric 27x27 weighing matrix in 1903 seconds of solver time, checked it with independent exact checkers on two machines, and ran a kill lane that could not refute it. The standard databases list no symmetric W(27,16). Check it yourself in under a second: docs/results/sw-27-16.
Constructed SW(25,16), a math case a 2025 paper listed as open. Given a one-sentence goal ("settle one published open question"), agent-afk agents picked the target, designed a structured SAT search, found a symmetric 25x25 weighing matrix in 760 seconds of solver time, checked it with independent exact checkers and a prior-work search, and wrote up a short proof. The author of the paper that listed the case as open confirmed it passes his verifier. Check it yourself in under a second: docs/results/sw-25-16.
The model isn't the product -- the loop is. Agent AFK hands you the loop as code: prompts, gates, routing, skills, traces, providers, terminal states. Edit any of them.
agent-afk keeps all of its state under ~/.afk/ — sessions, plugins, memory, logs, settings. Nothing is shared with ~/.claude/. You can delete ~/.claude entirely and afk still runs.
Local-first, no phone-home. There is no analytics or remote telemetry — Agent AFK never sends your prompts, code, or usage anywhere except directly to the model provider you configure. What telemetry exists is local JSONL under ~/.afk/ that you can read or delete.
Optional, in order of usefulness:
# Pick a model — opus | sonnet | haiku | fable (Anthropic) or codex (OpenAI)
AFK_MODEL=sonnet
# Enable the Telegram bot + send_telegram tool
TELEGRAM_BOT_TOKEN=1234567890:ABC...
AFK_TELEGRAM_ALLOWED_CHAT_IDS=12345678
# Optional per-session cost cap (unset = no limit)
# AFK_MAX_BUDGET_USD=5.00
Project-scoped system prompt. Drop an AFK.md at your project root and afk appends it to its built-in framework prompt whenever you run from that directory — your instructions layer on top of the base, they don't replace it. No frontmatter needed.
Check what resolved. afk config dumps the live configuration. afk doctor validates keys, paths, and provider connectivity.
Claude remembers things across sessions — preferences, decisions, conventions, and reusable procedures — without any manual setup. Memory is local-only and stored under ~/.afk/state/memory/.
Two storage tiers:
~/.afk/state/memory/HOT.md) — a small markdown file (≤ ~1,500 tokens) injected into every future session's system prompt automatically. Use it for the facts you want Claude to always carry: your name, working style, standing instructions. Capped at 5,250 characters; overflow is truncated with a sentinel comment.~/.afk/state/memory/memory.db) — an unbounded SQLite store, full-text-searchable via FTS5. Facts are queried on demand with the memory_search tool; they don't bloat every prompt.Fact categories (set when writing to the archive):
Tools available inside a session:
memory_search — full-text search across facts and procedures (supports FTS5: AND, OR, NOT, "exact phrase", prefix*).memory_update — write or supersede a fact in the archive (target: "fact") or overwrite hot memory (target: "hot").procedure_write — save a reusable step-by-step workflow as a named markdown file under ~/.afk/state/memory/procedures/. Procedures are searchable via memory_search.Hot memory is injected at session start; memory_search is called explicitly during a run. The session-end hook logs each completed top-level session to the archive automatically — subagent sessions are excluded.
All four surfaces (REPL, chat, daemon, Telegram) share the same store — memory written in one surface is available everywhere.
Default is sonnet. Override per-call with --model:
afk chat "explain this stack trace" --model opus_1m
afk --model sonnet_1m
afk chat "refactor this" --model gpt-5.5
Model slots — rebind any capability tier to any model:
The haiku/sonnet/opus/fable handles are fixed identities, not tier
aliases: they always resolve to their Claude model even after you rebind a tier
(so rebinding medium to an OpenAI model won't hijack sonnet). See
docs/model-slots.md for the full configuration reference.
MCP servers (tool-providing plugins over stdio/HTTP) — see docs/mcp.md for config, transports, OAuth, and security notes.
afk status # connection, model, bypass-mode state
afk doctor # environment self-check
afk config # dump resolved config
afk plugin list # installed plugins under ~/.afk/plugins/
afk completion zsh # shell completion (also: bash, fish)
afk --help # full command tree
Aliases: afk c → chat, afk i → interactive, afk s → status.
afk queue manages the pull-trigger daemon's task queue — tasks are persisted as JSON files and consumed one-by-one by afk daemon --trigger pull.
afk queue add "/forge-friction --auto" --notify-on failure
# enqueue a command; daemon picks it up on next poll
afk queue list # print all pending tasks (id, enqueued time, command)
afk queue remove <id> # drop a single pending task by id
afk queue clear # remove all pending tasks (prompts for confirmation)
afk queue clear --yes # clear without prompting (CI / non-interactive)
afk improve is a zero-LLM, deterministic pipeline that mines ~/.afk/state/witness/ session traces for ranked failure patterns, without making any model calls.
# Scan traces for failure patterns (dry-run; add --write to persist cards)
afk improve scan [--since 7d] [--write] [--only <detector,...>]
# Inspect failure cards
afk improve cards list [--pattern <name>] [--severity <level>] [--status open|deferred|resolved]
afk improve cards show <slug>
afk improve cards triage <slug> --note "..." [--status resolved]
# Draft a template-mode improvement proposal for a card (no LLM)
afk improve propose <slug> [--no-write]
# Generate a replay-mode eval-case from a failure card
afk improve eval-gen <cardSlug> [--evidence-row <i>] [--no-write]
# Run the deterministic guardrail contract for an eval-case
afk improve eval-run <evalCaseIdOrCardSlug> [--no-write]
Available detectors (run afk improve scan --help for thresholds): repeated-tool-use, subagent-block, closure-anomaly, tool-failure-density.
afk farm spawns N isolated git worktrees, runs an agent on each in parallel, scores the results (tests + lint + LoC delta), and prints a ranked summary.
afk farm "add retry logic to the queue consumer" --branches 3
afk farm "<task>" -n 5 --model sonnet --fail-fast
afk farm "<task>" -n 3 --no-score # skip post-run scoring
afk farm "<task>" -n 3 --labels "approach-a,approach-b,approach-c"
afk farm "<task>" -n 3 --no-memory --no-digest # skip memory write + Telegram
Each branch gets a dedicated worktree under ~/.afk/state/farm/. A commit-count escape check confirms each agent did real work. On completion, the winner (branch that passes tests and makes the smallest net change) is surfaced; the rest are left for manual cherry-pick or deletion.
afk does not prompt before each tool call — there is no per-tool approval flow. Claude runs bash, reads and writes files, fetches URLs, and calls MCP servers without asking each time. This is intentional — afk is built for unattended work, where a permission prompt with no human in front of it is just a wedged session.
New installs default to bypass mode. afk chat and afk interactive start in permissionMode: "bypassPermissions" when afk.config.json sets none — the agent reads and writes anywhere with no path-approval prompt. This is the equivalent of Claude Code's --dangerously-skip-permissions, on by default; use afk only on a machine and account you trust. Bypass does not change ask_question — that is the model choosing to ask you something, a separate axis.
To re-enable path containment, set the mode back any of these ways:
afk config set permissionMode default (or plan), or "permissionMode": "default" in afk.config.json. It stays that way until you change it again./bypass off in the REPL (the status line clears ⚡ bypass).With containment on (default), a file tool (read/write/edit/list/glob/grep) targeting a path outside the session's working directory triggers a path-approval prompt; pre-authorize paths with /allow-dir <path> (or answer "persist" at the prompt to remember them across sessions). Toggle bypass live anytime with /bypass. afk daemon always runs in bypass (no human to prompt); Telegram sessions stay contained (default) and rely on hook-based enforcement.
invalid x-api-key / ANTHROPIC_API_KEY not found — run afk login to authenticate interactively, or set the key manually: export ANTHROPIC_API_KEY=sk-ant-... (or add it to ~/.afk/config/afk.env). Run afk doctor to confirm.
Cannot send message: session is closed — the session timed out or was closed. Start a new one (afk i or a fresh afk chat).
Maximum turns exceeded — safety rail tripped. Bump it with --max-turns 50 or higher.
Hit the budget cap — raise AFK_MAX_BUDGET_USD or unset it for the session.
Telegram bot won't start — afk telegram status then afk telegram logs. Most common cause: missing AFK_TELEGRAM_ALLOWED_CHAT_IDS after token setup.
Recent releases at CHANGELOG.md, also viewable in-REPL via /changelog.
Building this in public at The Goblin Files -- field notes from running an autonomous coding agent at full tilt.
Agent AFK is open source, licensed under Apache-2.0 (SPDX: Apache-2.0).
Use it, modify it, fork it, embed it in commercial or closed-source products -- the
standard Apache conditions apply (keep the license and NOTICE, note
significant changes, and don't use the "Agent AFK" name or marks to imply
endorsement). No copyleft, no CLA. Every feature ships in the open-source repo --
there are no paid unlocks.
Paid plans are monthly support from the maintainer (setup, custom skills, tuning),
not feature gates. See LICENSING.md.
Contributions are accepted under the DCO
(git commit -s). "Agent AFK" is a trademark of Griffin Long. (c) Griffin Long.
Open-source coding-agent harness you can actually change — own the loop (prompts, gates, routing, skills, terminal states), use any model, run long tasks while you're away.
{ "mcpServers": { "agent-afk": { "command": "npx", "args": [ "-y", "[email protected]" ] } }}เครื่องมือ
ยังไม่มีรายการเครื่องมือ เซิร์ฟเวอร์แบบรีโมตจะได้รายการหลังการตรวจสอบสถานะครั้งแรก
FAQ
คัดลอกคอนฟิก stdio สำหรับ Cursor, Claude Desktop, Claude Code หรือ VS Code ได้ในหน้านี้ คอนฟิกนี้ใช้การติดตั้งของผู้เผยแพร่ ไม่ใช่พร็อกซีของ PlugMCP
ระเบียนนี้ไม่ได้ระบุตัวแปรสภาพแวดล้อมที่จำเป็น
Agent Afk ทำงานแบบโลคัลผ่าน stdio
ตรวจสอบล่าสุดเมื่อ 2026-11-14 Open-source coding-agent harness you can actually change — own the loop (prompts, gates, routing, skills, terminal states), use any model, run long tasks while you're away.