diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index 05d0c73..81d2885 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -77,7 +77,7 @@ jobs:
         include:
           # Linux builds use ubuntu-22.04 (glibc 2.35) rather than a manylinux
           # container: manylinux ships a static CPython without libpython*.so,
-          # which PyInstaller cannot embed. The glibc-2.35 floor still covers
+          # which Nuitka cannot embed. The glibc-2.35 floor still covers
           # Ubuntu 22.04+, Debian 12+, RHEL 9+, and all current modern distros.
           - target: linux-amd64
             runner: ubuntu-22.04
@@ -95,10 +95,13 @@ jobs:
             runner: macos-latest
             ext: ''
             archive: tar.gz
-          - target: windows-amd64
-            runner: windows-latest
-            ext: '.exe'
-            archive: zip
+          # Windows build temporarily disabled: MSVC path can't use ccache,
+          # so every run is a cold compile and takes >30 min on the default
+          # GitHub runner. Re-enable after switching to sccache or MinGW-w64.
+          # - target: windows-amd64
+          #   runner: windows-latest
+          #   ext: '.exe'
+          #   archive: zip
 
     steps:
       - uses: actions/checkout@v4
@@ -111,23 +114,74 @@ jobs:
         with:
           python-version: ${{ env.PYTHON_VERSION }}
 
-      - name: Install project + PyInstaller
+      - name: Install Nuitka build toolchain (Linux)
+        if: startsWith(matrix.target, 'linux-')
+        shell: bash
+        run: |
+          sudo apt-get update -qq
+          sudo apt-get install -y -qq patchelf ccache
+
+      - name: Install ccache (macOS)
+        if: startsWith(matrix.target, 'darwin-')
+        shell: bash
+        run: brew install ccache
+
+      # Persist ccache across runs. Key hashes pyproject.toml + build-binary.sh
+      # so a Nuitka pin bump or a build-script flag change buckets the cache
+      # into a fresh partition. ccache itself is content-addressed, so stale
+      # .o reuse is not a concern: if the generated C changes, it misses.
+      # Skipped on Windows because Nuitka uses MSVC there and ccache does not
+      # integrate with cl.exe.
+      - name: Restore ccache
+        if: ${{ !startsWith(matrix.target, 'windows-') }}
+        uses: actions/cache@v4
+        with:
+          path: |
+            ~/.cache/ccache
+            ~/Library/Caches/ccache
+          key: ccache-${{ matrix.target }}-py${{ env.PYTHON_VERSION }}-${{ hashFiles('pyproject.toml', 'scripts/build-binary.sh') }}
+          restore-keys: |
+            ccache-${{ matrix.target }}-py${{ env.PYTHON_VERSION }}-
+            ccache-${{ matrix.target }}-
+
+      - name: Configure ccache
+        if: ${{ !startsWith(matrix.target, 'windows-') }}
+        shell: bash
+        run: |
+          ccache --max-size=2G
+          ccache --zero-stats
+
+      - name: Install project + Nuitka
         shell: bash
         run: |
           python -m pip install --upgrade pip
           python -m pip install -e .
-          python -m pip install pyinstaller
+          python -m pip install "nuitka>=2.4" "zstandard>=0.22"
 
       - name: Build binary
         shell: bash
         run: |
-          pyinstaller --clean --noconfirm agentrun.spec
+          bash scripts/build-binary.sh
           ls -lh dist/
 
+      - name: Print ccache stats
+        if: ${{ !startsWith(matrix.target, 'windows-') }}
+        shell: bash
+        run: ccache --show-stats
+
       - name: Smoke test binary
         shell: bash
         run: |
-          ./dist/agentrun${{ matrix.ext }} --version
+          BIN="./dist/agentrun${{ matrix.ext }}"
+          "$BIN" --version
+          "$BIN" --help >/dev/null
+          "$BIN" config --help >/dev/null
+          "$BIN" model --help >/dev/null
+          "$BIN" sandbox --help >/dev/null
+          "$BIN" skill --help >/dev/null
+          "$BIN" super-agent --help >/dev/null
+          "$BIN" tool --help >/dev/null
+          echo "All 8 subcommand invocations OK on ${{ matrix.target }}"
 
       # --- Package (Unix) -----------------------------------------------
       - name: Package tar.gz (Unix)
diff --git a/.gitignore b/.gitignore
index b0eb629..87497df 100644
--- a/.gitignore
+++ b/.gitignore
@@ -12,7 +12,11 @@ venv/
 .idea/
 .vscode/
 *.spec
-!agentrun.spec
+# Nuitka build artifacts
+*.dist/
+*.build/
+*.bin
+*.onefile-build/
 .coverage
 coverage.json
 htmlcov/
diff --git a/AGENTS.md b/AGENTS.md
index 061bf6d..4f4a4f6 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -20,7 +20,7 @@ make test                 # Run all tests
 .venv/bin/pytest tests/test_cli_basic.py::TestConfigCommands::test_set_and_get -v  # Single test
 
 # Build standalone binary
-make build                # PyInstaller binary → dist/ar
+make build                # Nuitka --onefile binary → dist/agentrun (warm-start cache: ~/.agentrun/cache/)
 make build-all            # macOS + Linux (via Docker)
 ```
 
diff --git a/Makefile b/Makefile
index e53dd98..fce03d9 100644
--- a/Makefile
+++ b/Makefile
@@ -5,9 +5,7 @@ VENV     := .venv
 BIN      := $(VENV)/bin
 PIP      := $(BIN)/pip
 PIP_MIRROR := -i https://mirrors.aliyun.com/pypi/simple/
-PYINST   := $(BIN)/pyinstaller
 APP_NAME := agentrun
-SPEC     := agentrun.spec
 VERSION  := $(shell $(PYTHON) -c "from src.agentrun_cli import __version__; print(__version__)" 2>/dev/null || echo "0.1.0")
 
 help: ## Show this help message
@@ -21,7 +19,6 @@ install: ## Install the package in editable mode
 dev: ## Install with dev dependencies
 	$(PYTHON) -m venv $(VENV)
 	$(PIP) install $(PIP_MIRROR) -e ".[dev]" || $(PIP) install $(PIP_MIRROR) -e .
-	$(PIP) install $(PIP_MIRROR) pyinstaller
 
 lint: ## Run ruff linter
 	$(BIN)/ruff check src/ tests/
@@ -45,9 +42,8 @@ clean: ## Remove build artifacts
 	rm -rf build/ dist/ *.spec __pycache__
 	find . -type d -name __pycache__ -exec rm -rf {} + 2>/dev/null || true
 
-build: ## Build binary for the current platform (uses agentrun.spec)
-	DISABLE_BREAKING_CHANGES_WARNING=1 \
-	$(PYINST) --clean --noconfirm $(SPEC)
+build: ## Build single-file binary for the current platform (uses Nuitka)
+	bash scripts/build-binary.sh
 	@echo ""
 	@echo "Binary built: dist/$(APP_NAME)"
 	@ls -lh dist/$(APP_NAME)
@@ -55,8 +51,9 @@ build: ## Build binary for the current platform (uses agentrun.spec)
 build-macos: build ## Alias for build (on macOS, just run 'make build')
 	@echo "macOS binary ready at dist/$(APP_NAME)"
 
-# Cross-compiling Python to Linux is not supported by PyInstaller.
-# Use this target inside a Linux environment (Docker / CI).
+# Nuitka compiles Python to native C, so cross-compiling is not supported.
+# Use this target inside a Linux environment (Docker / CI) to produce a
+# Linux binary when you're on a non-Linux host.
 build-linux: build ## Build Linux binary (run inside Linux or Docker)
 	@echo "Linux binary ready at dist/$(APP_NAME)"
 
@@ -69,10 +66,9 @@ build-all: ## Build for all platforms (macOS local + Linux via Docker)
 	tar cf - --exclude=.venv --exclude=.git --exclude=build --exclude=dist --exclude=__pycache__ --exclude='*.pyc' . | \
 	docker run --rm -i -v $(PWD)/dist:/out python:3.10-slim sh -c \
 		"mkdir /build && cd /build && tar xf - && \
-		 apt-get update -qq && apt-get install -y -qq binutils >/dev/null 2>&1 && \
-		 pip install $(PIP_MIRROR) -e . && pip install $(PIP_MIRROR) pyinstaller && \
-		 DISABLE_BREAKING_CHANGES_WARNING=1 \
-		 pyinstaller --clean --noconfirm $(SPEC) && \
+		 apt-get update -qq && apt-get install -y -qq binutils patchelf ccache gcc >/dev/null 2>&1 && \
+		 pip install $(PIP_MIRROR) -e . && pip install $(PIP_MIRROR) nuitka zstandard && \
+		 bash scripts/build-binary.sh && \
 		 cp dist/$(APP_NAME) /out/$(APP_NAME)"
 	@mkdir -p dist/linux && cp dist/$(APP_NAME) dist/linux/$(APP_NAME)
 	@echo ""
diff --git a/README.md b/README.md
index 15f902d..3e2a5c0 100644
--- a/README.md
+++ b/README.md
@@ -18,7 +18,7 @@ agents that you configure declaratively without writing or deploying any runtime
 - **Multiple output formats** — `json` (default), `table`, `yaml`, and `quiet` for shell piping.
 - **Agent-friendly** — JSON-by-default output, deterministic exit codes, no interactive prompts when stdin isn't a TTY.
 - **Rich sandbox primitives** — code execution, file system, process management, and CDP/VNC-backed browser automation.
-- **Single-file distribution** — PyInstaller produces standalone `ar` / `agentrun` binaries for Linux, macOS and Windows (x86_64 + arm64).
+- **Single-file distribution** — Nuitka `--onefile` produces standalone `ar` / `agentrun` binaries for Linux, macOS and Windows (x86_64 + arm64) with warm-start caching under `~/.agentrun/cache/`.
 
 ## Installation
 
diff --git a/README_zh.md b/README_zh.md
index caf12c3..addc97f 100644
--- a/README_zh.md
+++ b/README_zh.md
@@ -17,7 +17,7 @@ Agent）**：一种由平台托管、用户只需声明配置、无需编写或
 - **多种输出格式** — 默认 `json`，支持 `table` / `yaml` / `quiet`（适合 shell 管道）。
 - **对 Agent 友好** — 默认 JSON 输出、确定性退出码、非 TTY 下不弹交互提示。
 - **完整沙箱能力** — 代码执行、文件系统、进程管理、CDP/VNC 浏览器自动化。
-- **单文件分发** — PyInstaller 产出 Linux / macOS / Windows（x86_64 + arm64）上的独立 `ar` / `agentrun` 二进制。
+- **单文件分发** — 基于 Nuitka `--onefile` 产出 Linux / macOS / Windows（x86_64 + arm64）上的独立 `ar` / `agentrun` 二进制，热启动复用 `~/.agentrun/cache/` 缓存。
 
 ## 安装
 
diff --git a/agentrun.spec b/agentrun.spec
deleted file mode 100644
index f4cd6f0..0000000
--- a/agentrun.spec
+++ /dev/null
@@ -1,89 +0,0 @@
-# -*- mode: python ; coding: utf-8 -*-
-#
-# PyInstaller spec — single source of truth for binary builds.
-# Used by `make build` locally and by `.github/workflows/release.yml` in CI.
-# Keep the EXCLUDES list in sync with the CLI's actual runtime needs.
-
-from PyInstaller.utils.hooks import collect_data_files
-
-# Everything the CLI does NOT need at runtime but gets pulled in transitively
-# via agentrun-inner-test[core]. Excluding these keeps the binary small.
-EXCLUDES = [
-    'litellm',
-    'tablestore',
-    'agentrun_mem0ai',
-    'agentrun_mem0',
-    'alibabacloud_bailian20231229',
-    'alibabacloud_gpdb20160503',
-    'tiktoken',
-    'tokenizers',
-    'numpy',
-    'grpcio',
-    'torch',
-    'tensorflow',
-    'transformers',
-    'PIL',
-    'matplotlib',
-    'scipy',
-    'sklearn',
-    'pandas',
-    'pytz',
-    'pygments',
-    'sqlalchemy',
-    'Crypto',
-    'pycryptodome',
-    'rich',
-    'markdown_it',
-    'mysql',
-    'MySQLdb',
-    'oss2',
-    'posthog',
-    'jinja2',
-    'qdrant_client',
-    'huggingface_hub',
-    'hf_xet',
-    'fsspec',
-    'h2',
-    'regex',
-    'future',
-    'google',
-]
-
-datas = []
-datas += collect_data_files('certifi')
-
-a = Analysis(
-    ['src/agentrun_cli/main.py'],
-    pathex=[],
-    binaries=[],
-    datas=datas,
-    hiddenimports=[],
-    hookspath=[],
-    hooksconfig={},
-    runtime_hooks=[],
-    excludes=EXCLUDES,
-    noarchive=False,
-    optimize=0,
-)
-pyz = PYZ(a.pure)
-
-exe = EXE(
-    pyz,
-    a.scripts,
-    a.binaries,
-    a.datas,
-    [],
-    name='agentrun',
-    debug=False,
-    bootloader_ignore_signals=False,
-    strip=False,
-    upx=False,
-    upx_exclude=[],
-    runtime_tmpdir=None,
-    console=True,
-    disable_windowed_traceback=False,
-    argv_emulation=False,
-    target_arch=None,
-    codesign_identity=None,
-    entitlements_file=None,
-)
diff --git a/docs/en/index.md b/docs/en/index.md
index 5eab1b1..8d56275 100644
--- a/docs/en/index.md
+++ b/docs/en/index.md
@@ -69,6 +69,14 @@ make build        # local binary → dist/agentrun
 After installation, both `ar` and `agentrun` are available as entry points and behave
 identically. `ar` is shorter; the examples in this manual use it.
 
+### Binary startup cache
+
+The prebuilt binary is a Nuitka `--onefile` executable. On first launch it extracts its payload into `~/.agentrun/cache/agentrun-<version>/` (about 20 MB); subsequent launches reuse the cache, bringing warm start-up below 300 ms.
+
+- **Safe to delete.** Remove `~/.agentrun/cache/` at any time; the next invocation re-extracts.
+- **Upgrades.** A new binary version writes to a new subdirectory; old ones stay until you clean them up.
+- **Read-only `$HOME`.** If `~/.agentrun/cache/` is not writable, the bootstrap falls back to `$TMPDIR` with full re-extraction on every run (~2 s). Either grant write access or run from a shell where `$HOME` points somewhere writable.
+
 ## Authentication
 
 The CLI resolves credentials from three sources, in this order:
diff --git a/docs/zh/index.md b/docs/zh/index.md
index 0d381c4..2cddf33 100644
--- a/docs/zh/index.md
+++ b/docs/zh/index.md
@@ -68,6 +68,14 @@ make build        # 本地打独立二进制 → dist/agentrun
 安装完成后 `ar` 和 `agentrun` 都是入口点，行为完全一致。`ar` 更短，文档里的示例
 默认用 `ar`。
 
+### 二进制启动缓存
+
+预编译二进制是一个 Nuitka `--onefile` 可执行文件。首次运行会把内置 payload 解压到 `~/.agentrun/cache/agentrun-<version>/`（约 20 MB），之后每次启动复用缓存，热启动耗时低于 300 ms。
+
+- **可以随时删除。**`~/.agentrun/cache/` 删掉后下一次运行自动重建。
+- **升级行为。**新版本二进制会写入新的子目录，老目录保留直到手动清理。
+- **`$HOME` 只读的情况。**若 `~/.agentrun/cache/` 不可写，bootstrap 会回落到 `$TMPDIR` 且每次都完整解压（~2 秒）。请确保目录可写，或从 `$HOME` 指向可写位置的 shell 启动。
+
 ## 认证
 
 CLI 按以下顺序解析凭证（上面优先）：
diff --git a/pyproject.toml b/pyproject.toml
index 980a259..09079cd 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -38,7 +38,8 @@ Issues = "https://github.com/Serverless-Devs/agentrun-cli/issues"
 dev = [
     "pytest>=8.0.0", "pytest-cov>=6.0.0",
     "pytest-asyncio>=1.2.0",
-    "pyinstaller>=6.0.0",
+    "nuitka>=2.4",
+    "zstandard>=0.22",
     "ruff>=0.14.0",
     "mypy>=1.11.0",
     "types-PyYAML>=6.0",
@@ -48,7 +49,8 @@ dev = [
 dev = [
     "pytest>=8.0.0", "pytest-cov>=6.0.0",
     "pytest-asyncio>=1.2.0",
-    "pyinstaller>=6.0.0",
+    "nuitka>=2.4",
+    "zstandard>=0.22",
     "ruff>=0.14.0",
     "mypy>=1.11.0",
     "types-PyYAML>=6.0",
diff --git a/scripts/bench-startup.sh b/scripts/bench-startup.sh
new file mode 100755
index 0000000..353072b
--- /dev/null
+++ b/scripts/bench-startup.sh
@@ -0,0 +1,49 @@
+#!/usr/bin/env bash
+# Measure agentrun binary cold-start + warm-start times.
+#
+# Usage: bash scripts/bench-startup.sh [binary-path]
+# Default binary: ./dist/agentrun
+#
+# Prints per-run `real` time and the median of runs 2..10 (warm runs).
+
+set -euo pipefail
+
+BINARY="${1:-./dist/agentrun}"
+if [ ! -x "$BINARY" ]; then
+  echo "Binary not found or not executable: $BINARY" >&2
+  exit 2
+fi
+
+echo "=== Benchmark: $BINARY --help ==="
+echo "(run 1 = cold; runs 2..10 = warm)"
+echo ""
+
+# Portable high-resolution timer. macOS BSD `date` lacks %N, so use python3
+# (required anyway as a build dep). Returns seconds as a float.
+now() { python3 -c 'import time; print(time.perf_counter())'; }
+
+TIMES=()
+for i in $(seq 1 10); do
+  START=$(now)
+  "$BINARY" --help >/dev/null
+  END=$(now)
+  ELAPSED=$(awk "BEGIN{print $END - $START}")
+  TIMES+=("$ELAPSED")
+  printf "Run %2d: %.3f s\n" "$i" "$ELAPSED"
+done
+
+# Stats over runs 2..10 (warm). Report min as well as median: on dev
+# machines with on-access AV scanners attached to unsigned binaries the
+# scanner adds ~10s to some runs, polluting the median. `min` reflects
+# Nuitka's true warm-start cost when the scan is cached; `median` is the
+# authoritative figure on clean environments (Linux CI, signed releases).
+WARM=("${TIMES[@]:1}")
+SORTED=$(printf '%s\n' "${WARM[@]}" | sort -n)
+N=$(echo "$SORTED" | wc -l | tr -d ' ')
+MID=$(( (N + 1) / 2 ))
+MIN=$(echo "$SORTED" | sed -n '1p')
+MEDIAN=$(echo "$SORTED" | sed -n "${MID}p")
+
+echo ""
+printf "Warm min    (runs 2..10): %.3f s\n" "$MIN"
+printf "Warm median (runs 2..10): %.3f s\n" "$MEDIAN"
diff --git a/scripts/build-binary.sh b/scripts/build-binary.sh
new file mode 100755
index 0000000..63abc7e
--- /dev/null
+++ b/scripts/build-binary.sh
@@ -0,0 +1,116 @@
+#!/usr/bin/env bash
+# Build the agentrun CLI as a single-file binary using Nuitka --onefile.
+#
+# Called by `make build` and by .github/workflows/release.yml.
+# Output: ./dist/agentrun (or ./dist/agentrun.exe on Windows).
+#
+# Exit codes: 0 success, non-zero = Nuitka failure (propagates).
+
+set -euo pipefail
+
+# ----- Config -----------------------------------------------------------
+# EXCLUDES: 1:1 port of the old agentrun.spec EXCLUDES list.
+# Any transitive dep of agentrun-sdk[core] that the CLI does NOT use at
+# runtime goes here. Keep sorted for review clarity.
+EXCLUDES="\
+agentrun_mem0,\
+agentrun_mem0ai,\
+alibabacloud_bailian20231229,\
+alibabacloud_gpdb20160503,\
+Crypto,\
+fsspec,\
+future,\
+google,\
+grpcio,\
+h2,\
+hf_xet,\
+huggingface_hub,\
+jinja2,\
+litellm,\
+markdown_it,\
+matplotlib,\
+MySQLdb,\
+mysql,\
+numpy,\
+oss2,\
+pandas,\
+PIL,\
+posthog,\
+pycryptodome,\
+pygments,\
+pytz,\
+qdrant_client,\
+regex,\
+rich,\
+sklearn,\
+scipy,\
+sqlalchemy,\
+tablestore,\
+tensorflow,\
+tiktoken,\
+tokenizers,\
+torch,\
+transformers"
+
+# Version: read from setuptools-scm generated file if present, else fall
+# back to git describe, else "0.0.0+unknown".
+VERSION=$(python3 -c "from src.agentrun_cli import __version__; print(__version__)" 2>/dev/null \
+  || git describe --tags --always --dirty 2>/dev/null \
+  || echo "0.0.0+unknown")
+
+# Nuitka's --product-version requires a strict N(.N)* form with all numeric
+# components. Derive a sanitized variant by stripping PEP 440 pre-release /
+# dev / local suffixes and padding to 3 components, so
+# "0.1.0rc5.dev2+g1713d9f79.d20260421" normalizes to "0.1.0" and
+# "0.3.0rc1" also normalizes to "0.3.0" (stable cache-dir naming).
+PRODUCT_VERSION=$(python3 -c "
+import re, sys
+v = sys.argv[1]
+# Drop local segment (anything after '+') and any non-numeric trailing
+# components (rc, dev, a, b, post, ...).
+v = v.split('+', 1)[0]
+parts = v.split('.')
+out = []
+for p in parts:
+    if re.fullmatch(r'[0-9]+', p):
+        out.append(p)
+    else:
+        break
+# Pad to 3 numeric components so '0.3.0', '0.3.0rc1', and '0.3.0.dev2'
+# all normalize to '0.3.0' (stable cache-dir naming under ~/.agentrun/cache/).
+out = (out + ['0', '0', '0'])[:3]
+print('.'.join(out))
+" "$VERSION")
+
+# Cache directory (per design doc, co-located with ~/.agentrun/config.json).
+# Nuitka expands {HOME} and {VERSION} at bootstrap time. Each binary
+# version gets its own subdirectory so upgrades don't collide.
+TEMPDIR_SPEC='{HOME}/.agentrun/cache/agentrun-{VERSION}'
+
+# Output filename differs on Windows (.exe suffix handled by Nuitka).
+mkdir -p dist
+
+# ----- Build ------------------------------------------------------------
+echo "=== Building agentrun (version: $VERSION) with Nuitka --onefile ==="
+python3 -m nuitka \
+  --onefile \
+  --assume-yes-for-downloads \
+  --output-filename=agentrun \
+  --output-dir=dist \
+  --include-package=agentrun_cli \
+  --include-package-data=certifi \
+  --onefile-tempdir-spec="$TEMPDIR_SPEC" \
+  --product-name=agentrun \
+  --product-version="$PRODUCT_VERSION" \
+  --python-flag=-O \
+  --lto=no \
+  --nofollow-import-to="$EXCLUDES" \
+  --remove-output \
+  src/agentrun_cli/main.py
+
+# ----- Post-build report ------------------------------------------------
+BINARY="dist/agentrun"
+[ -f "${BINARY}.exe" ] && BINARY="${BINARY}.exe"
+echo ""
+echo "=== Build complete ==="
+ls -lh "$BINARY"