seek

<p align="center"> <img src="examples/hero.svg" alt="seek — a local agent platform that plans, spawns a team of sub-agents, and ships while you sleep" width="820"> </p> <p align="center"> <strong>Not an agent demo — a local agent platform that plans, spawns a team of sub-agents, and ships while you sleep.</strong> </p> <p align="center"> <img src="https://img.shields.io/badge/license-MIT-f5a623?style=flat-square" alt="MIT"> <img src="https://img.shields.io/badge/CI-race--tested-2ea043?style=flat-square&logo=github&logoColor=white" alt="CI"> <img src="https://img.shields.io/badge/release-v0.10.3-f5a623?style=flat-square" alt="release"> <img src="https://img.shields.io/badge/platforms-macOS%20%C2%B7%20Linux%20%C2%B7%20Windows-2b2b32?style=flat-square" alt="Platforms"> &nbsp;·&nbsp; <a href="README.zh.md">中文</a> </p>

A coding agent that runs as a single static Go binary in your terminal — a ~6 MB download, no daemon, no telemetry, no Python/Node runtime, no seek-operated backend. Bring your own model key: DeepSeek-first, and it also speaks OpenAI / Anthropic / Gemini and any OpenAI-compatible endpoint (KIMI, …).

30-second version — It's Claude Code / aider territory, but DeepSeek-first: the autonomous run that audited and fixed these very docs (33 files, 212 tool calls) cost $0.17 at a 97.7% prefix-cache hit. One binary, your own key, nothing phones home. Scroll down to watch that exact run ↓

---

Why seek is different

The three claims in the tagline, each backed by shipping code:

• 🧭 Plans before it acts — /plan gates work behind analyze → propose → you approve → execute. Read-only until you approve a step-by-step task list; push back mid-run and it re-plans without redoing finished work. State rebuilds from the transcript, so seek -resume resumes the plan exactly where it left off.
• 👥 Spawns a team of sub-agents — fans a task out to sub-agents that run in parallel, each isolated in its own git worktree so they never collide. Permissions narrow monotonically; cost aggregates to the parent's status bar. Watch live with /agents · /worktrees. → guide
• 🌙 Ships while you sleep — zero-daemon scheduling on your OS (launchd / systemd / Task Scheduler). seek cron, a model-scheduled schedule_wakeup, or a CI trigger file runs unattended, commits, and pushes the result to your phone (ntfy / Slack / Discord / any URL). → guide

---

See it run — these docs were audited by seek itself

We handed seek one goal — "audit the codebase docs for accuracy, fix them, and weave our pitfalls log into the book" — and walked away. It planned the work, spawned a team of sub-agents, edited 33 files across all 24 book chapters, and stopped only when a cheap judge confirmed the goal was met: 120 turns · 212 tool calls · clean build · 97.7% cache hit · $0.17. Real run, unedited (just sped up):

1 · Plans, then spawns a team of sub-agents

<p align="center"><img src="examples/demo-plan-agents.gif" alt="seek plans the task, then fans out explore sub-agents" width="820"></p>

2 · Works autonomously across the codebase

<p align="center"><img src="examples/demo-edit.gif" alt="seek edits doc comments and weaves pitfalls into book chapters" width="820"></p>

3 · …until a cheap judge says the goal is met

<p align="center"><img src="examples/demo-goalmet.gif" alt="goal met after 3 turns, clean build, $0.17" width="820"></p>

Every doc change in this PR came from that single $0.17 run. The recording recipe is examples/demo-fullrun.tape.

---

Install

OS=$(uname -s | tr '[:upper:]' '[:lower:]')
ARCH=$(uname -m | sed 's/x86_64/amd64/;s/aarch64/arm64/')
VER=$(curl -fsSL https://api.github.com/repos/whyiyhw/seek/releases/latest | sed -nE 's/.*"tag_name":[[:space:]]*"v([^"]+)".*/\1/p')
curl -fsSL "https://github.com/whyiyhw/seek/releases/download/v${VER}/seek_${VER}_${OS}_${ARCH}.tar.gz" \
  | sudo tar -xz -C /usr/local/bin seek
seek

First launch walks you through picking a provider + saving a key — that's it. Windows: docs/guide-windows.md. Upgrade: seek -upgrade (sha256-verified, atomic) or /upgrade in the TUI.

seek                                    # interactive TUI
seek -p "what does internal/agent do?"  # one-shot, pipeline-friendly
seek goal run "all tests in ./auth pass" # autonomous: work until the goal is met

---

An order of magnitude cheaper

DeepSeek pricing (from internal/pricing/pricing.go):

| Metric | DeepSeek V4-Flash | DeepSeek V4-Pro | Claude Sonnet 4 | |---|---|---|---| | Input (cache miss) | $0.44 / $0.22¹ | $1.32 / $0.66¹ | $3 / 1M | | Input (cache hit) | $0.014 / $0.007¹ | $0.044 / $0.022¹ | $0.30 / 1M | | Output | $1.32 / $0.66¹ | $3.96 / $1.98¹ | $15 / 1M | | Off-peak² | −50% | −50% | — |

_{¹ Peak / off-peak rates (off-peak = half price), effective 2026-08-16 16:00 UTC — see internal/pricing/pricing.go. ² Peak hours are 01:00–04:00 and 06:00–10:00 UTC (09:00–12:00 / 14:00–18:00 Beijing); everything else is off-peak.}

Measured 95–97% prefix-cache hit — discipline, not luck: tool schemas are byte-stable []byte constants, tool output is capped at write-time, and history is never rewritten before send. The status bar shows the live ratio + dollars saved.

---

Features

| | | |---|---| | Autopilot | unattended end-to-end: decompose → parallel worktree fleet → commit → push. 📘 | | /goal | loop across turns until a cheap model judges a condition met — TUI / headless / cron. 📘 | | OS sandbox | seatbelt (macOS) / landlock (Linux) — kernel-level confinement, zero runtime deps. 📘 | | ACP / Zed | run seek as an editor agent via the Agent Client Protocol. 📘 | | Portable skills | Anthropic Agent Skills format — any Claude Code skill installs unmodified. 📘 | | Three-tier memory | session · project (decay-score GC) · cross-project "soul" (seek -dream). 📘 | | Checkpoints | per-turn git snapshot + file-level /undo /redo /restore. 📘 | | Background jobs | detach long builds/servers → bg-N; monitor polls / waits / kills. 📘 | | Semantic references | LSP "who calls this?" (gopls / pyright / tsserver), grep fallback. 📘 | | Native vision input | paste or @img.png → image bytes sent natively to the multimodal default model (deepseek-flash, V4.1). 📘 | | Code review | effort-graded diff review with --fix and --comment. 📘 | | MCP client | pass through any MCP server's tools. 📘 | | Push to phone | cron / autopilot / long-turn completions → webhook. 📘 | | Dual-axis permissions | Preference (Deny / Ask / Yolo) × Workflow (None / Plan-analyze / Plan-execute). |

Plus shell hooks, a JSON-RPC 2.0 server mode, and DeepSeek extras (V4 reasoning via think, FIM for 5–10× cheaper small edits, off-peak countdown).

---

Commands

seek                       # interactive TUI
seek -p '<prompt>'         # one-shot print mode (pipeline-friendly)
seek -resume <sid>         # resume a session (-continue for the most recent)
seek -rpc                  # JSON-RPC 2.0 server (IDE integration)
seek acp                   # Agent Client Protocol server (Zed, …)

seek goal       run "<condition>"          # autonomous loop until met
seek skill      install / list / stats / uninstall / update
seek memory     list / show / search / archive
seek cron       create / list / run / delete / tick   # --autopilot / --goal
seek worktree   list / gc
seek checkpoint list / clean               # + seek undo / seek redo
seek hooks      list / check / trust / audit

Every subcommand is also /<name> in the TUI. TUI-only: /plan /goal /steer /agents /worktrees /distill /code-review. Full list: /help.

Interface language: /lang <en|zh> switches the human-facing text (hints, notices, command feedback) and persists to ~/.seek/config.json; SEEK_LANG=zh seek overrides per-launch; an unset config auto-detects from LC_ALL/LANG. Model-facing text and the status bar stay English by design — see docs/prd/feature-i18n.md.

---

Built for real, not a weekend build

~85k lines of Go (~44k non-test) across 66 packages, -race tested on macOS / Linux / Windows in CI. PRD-driven — docs/prd/ holds the v0–v7 design history; a pitfalls log and a behavioral eval harness keep it honest. Zero external deps in the DeepSeek client; stdlib-first throughout. Current release: v0.10.3.

Docs: docs/ · contributing: CONTRIBUTING.md / AGENTS.md

---

Open source

MIT. No region lock, no signup, no mandatory telemetry — builders anywhere are welcome. Inspired by earendil-works/pi (MIT); attribution in NOTICE.