🦍 ApesCode

Multi-provider AI coding agent for your terminal. 4-40× cheaper than Claude Code, with smart routing across 4 providers, MCP, sandbox, and REPL with branches.

Why ApesCode?

💰 Massive economy via smart routing

Empirical measurements on equivalent tasks (2026-04, smoke suite):

| Task type | Sonnet 4.6 | Kimi K2.6 | Gemini Flash | Haiku 4.5 | |---|---|---|---|---| | Simple coding | $0.015 | $0.003 (5×) | — | — | | File edit | $0.010 | $0.002 (5×) | — | — | | Trivial prompt | $0.002 | — | $0.0005 (4×) | $0.0002 (10×) | | Orchestration | $0.030 | — | $0.001 (30×) | — |

Classifier Haiku cost per routing decision: ~$0.00015 (negligible).

Monthly cost: 25-45 €/month vs 80 €/month for Claude Pro, measured empirically on a typical agentic workload.

🔌 4 providers integrated

• Anthropic — Sonnet 4.6 / Opus 4.7 / Haiku 4.5 (orchestration + baseline)
• Moonshot — Kimi K2.6 (coding sweet spot)
• xAI — Grok 4.20 non-reasoning (research / fresh facts)
• Google — Gemini 2.5 Flash (opt-in, cheapest worker; A/B orchestration)

Each provider is opt-in: set the API key env var, enable in settings.json, and the router starts using it.

🛡️ Defense-in-depth security (4 layers)

1. Denylist — unified dangerous command patterns (rm -rf /, curl | sh, etc.).
2. Permission gate — interactive prompt on suspect tool calls; cached per-session.
3. LLM classifier (auto mode) — Haiku pre-approves safe tool calls silently.
4. OS sandbox — bubblewrap (Linux) / sandbox-exec (macOS) wraps bash; Windows passthrough.

🌲 REPL with branches

/branch, /fork, /retry (Ctrl+R), /compact — explore multiple approaches without losing history. Sessions persist to SQLite; --resume re-enters any previous session.

✅ Feature parity with Claude Code

MCP, hooks (PreToolUse / PostToolUse / UserPromptSubmit / SessionStart / SessionEnd / Stop), subagents via Task tool, plan mode (readonly with approval), slash commands, web_search, web_fetch.

Install

npm install -g apescode

Requires Node.js ≥ 20. Works on macOS, Linux, and Windows.

Quick start

# Set at least one provider API key (Anthropic is the baseline)
export ANTHROPIC_API_KEY="sk-ant-..."

# Launch interactive REPL
apescode

# One-shot mode (runs once, exits)
apescode "Create a hello.txt file"

# Resume a previous session
apescode --list         # shows recent sessions
apescode --resume <id>  # rehydrates by id

Providers setup

Anthropic (required for baseline)

Provision at console.anthropic.com.

export ANTHROPIC_API_KEY="sk-ant-..."

Moonshot Kimi K2.6 (optional — huge economy on coding)

Provision at platform.moonshot.ai.

export MOONSHOT_API_KEY="sk-..."

Enable in .claude/settings.json:

{
  "routing": {
    "enabled": true,
    "specialists": { "simple-coding": "kimi-k2.6" }
  }
}

xAI Grok 4.20 (optional — research / fresh facts)

Provision at console.x.ai.

export XAI_API_KEY="xai-..."

Google Gemini 2.5 Flash (opt-in experimental — cheapest worker)

Provision at aistudio.google.com (free tier 1M tokens/day).

export GOOGLE_API_KEY="AIza..."

Enable the experimental provider in .claude/settings.json:

{
  "routing": {
    "enabled": true,
    "experimentalProviders": { "google": true }
  }
}

Tavily (optional — AI-optimized WebSearch)

Provision at tavily.com (free tier 1000 searches/month).

export TAVILY_API_KEY="tvly-..."

If TAVILY_API_KEY is not set, web_search falls back silently to DuckDuckGo HTML scraping.

Commands & shortcuts

Slash commands

| Command | Description | |---|---| | /help | list commands | | /compact | compact conversation history to save tokens | | /branch <name> | save the current session state as a branch | | /fork <name> | branch from the current state with auto-backup | | /branches | list saved branches | | /retry | re-run the last user turn (Ctrl+R) | | /new | clear session context (fresh conversation) | | /exit, /quit | graceful shutdown |

Custom slash commands (project-level) can be added via .claude/commands/*.md.

REPL keyboard shortcuts

• Tab — autocomplete slash commands
• Ctrl+R — retry last user message
• Ctrl+C — interrupt streaming / graceful shutdown on idle
• Shift+Enter or Ctrl+J — newline in multiline input
• Up / Down — history navigation (session-local)

CLI flags

• -m, --model <id> — override default model for the turn
• --provider <id> — force-route everything through one provider (anthropic | moonshot | xai | google)
• --mode <mode> — default | acceptEdits | plan | bypassPermissions
• --resume <id> — rehydrate a previous session
• --list — list persisted sessions
• --yes — auto-approve permission prompts (required for non-TTY runs)
• --one-shot — force one-shot mode even in a TTY

Configuration

Settings are read from ~/.claude/settings.json (user-level) with per-project overrides at <cwd>/.claude/settings.json. Shallow merge; project wins.

Minimal (baseline, no routing)

{
  "sandbox": { "mode": "permission-prompt" }
}

Recommended routing config (save ~60% on LLM costs)

{
  "routing": {
    "enabled": true,
    "orchestrator": "claude-sonnet-4-6",
    "specialists": {
      "trivial": "claude-haiku-4-5",
      "simple-coding": "kimi-k2.6",
      "research": "grok-4.20-0309-non-reasoning",
      "complex-reasoning": "claude-opus-4-7",
      "agentic": "claude-sonnet-4-6"
    },
    "fallbackChain": {
      "simple-coding": ["kimi-k2.6", "claude-sonnet-4-6", "claude-haiku-4-5"],
      "research": ["grok-4.20-0309-non-reasoning", "claude-sonnet-4-6"]
    },
    "classifierThreshold": 0.7
  }
}

Web tools

{
  "webSearchEnabled": true,
  "webFetchEnabled": true,
  "webSearch": { "fallbackEnabled": true },
  "webFetch": {
    "allowedDomains": ["*.github.io", "docs.python.org", "nodejs.org"],
    "maxBodyBytes": 1000000,
    "maxOutputChars": 10000,
    "timeoutMs": 15000
  }
}

Architecture

┌──────────────────────────────────────────────────────────────┐
│  User input                                                   │
└────────────────────────┬─────────────────────────────────────┘
                         ▼
┌──────────────────────────────────────────────────────────────┐
│  Router (3 layers)                                            │
│  1. Heuristic fast path (trivial, coding, research patterns)  │
│  2. Haiku classifier (ambiguous prompts, LRU cache + cost cap)│
│  3. Sonnet orchestrator (AGENTIC, delegates via Task tool)    │
└────────────────────────┬─────────────────────────────────────┘
                         ▼
┌──────────────────────────────────────────────────────────────┐
│  Provider (Anthropic / Moonshot / xAI / Google)               │
│  + Circuit breaker per-provider                               │
│  + Fallback chain per category                                │
└────────────────────────┬─────────────────────────────────────┘
                         ▼
┌──────────────────────────────────────────────────────────────┐
│  Tool use (bash, read, write, edit, grep, glob, ls,           │
│           web_search, web_fetch, Task, MCP tools)             │
│  + 4-layer sandbox (denylist, permission, classifier, OS)     │
└──────────────────────────────────────────────────────────────┘

Project status

Phase 6 complete — multi-provider routing production-ready, 4 providers empirically validated. Phase 7 in progress — polish + distribution (current).

• 684 unit tests + 38 smoke scenarios green
• Monthly cost empirically 2-3× lower than Claude Pro on equivalent workload
• Works on macOS, Linux, Windows (Windows sandbox-exec is passthrough)

See CHANGELOG.md for the commit-by-commit history.

License

MIT — see LICENSE.

Credits

Inspired by Claude Code by Anthropic. Built as a learning project + to optimize multi-provider costs. AI SDK layer by Vercel AI SDK v5.

Contributions welcome — open an issue at https://github.com/stephencvd/apescode/issues.