Everything Claude Code does and more. Any model. Zero subscription. Smarter with your tokens.

OpenCode X is a terminal AI coding agent forked from opencode with 280+ commits. It matches Claude Code's headline features (autonomous goals, background execution, parallel subagents, hooks) while adding what Claude Code can't offer: provider freedom, intelligent token compression, 3-tier context safety, configurable orchestration guardrails, and an open-source MIT codebase you actually own.

Your existing claude code hooks, plugins, agents and skills work immediately. Zero migration.

Autonomous multi-turn execution with real-time cost tracking

You're paying $200+/month per developer for a single-provider tool. OpenCode X gives you the same workflow with genuine advantages:

• $0/month — bring your own API keys, pay only for tokens you use
• Any model — Claude, GPT, Gemini, Mistral, Llama, DeepSeek, or any of 75+ providers
• 30-60% token savings — LLM compression on tool outputs before they reach your model (Claude Code sends raw output)
• Never lose context — 3-tier safety net auto-recovers at 75%, 97%, and 413 overflow (Claude Code only has auto-compact)
• Your hooks already work — same config, same env vars, same lifecycle events
• Configurable guardrails — doom loop detection, spawn limits, per-model concurrency (Claude Code doesn't expose these)
• Hybrid routing — automatically route compression/titles to cheap models, reasoning to premium ones
• Open source — MIT, no telemetry, no account, fork and extend freely

Same foundation you love. 280 commits of production-grade additions:

• Swarm execution — batch-parallel subagents with bounded concurrency using /swarm
• Autonomous goals — /goal with judge validation and 200-turn auto-continuation
• Persistent + session memory — agent learns across sessions, remembers across /clear
• Push to background — Leader+D detaches running session, keeps working
• Token intelligence — LLM compression, improved cache stability
• Context safety net — 3-layer auto-recovery, sessions never crash from overflow
• Doom loop detection — catches infinite tool loops before they burn tokens
• Parallel tool calls — concurrent read/grep/glob within single LLM round-trip
• Hybrid model routing — cheap model for lightweight work, premium for reasoning
• Claude Code compatibility — hooks, plugins, agents, skills, CLAUDE.md all work out of the box
• Orchestration guardrails — spawn depth limits, descendant caps, concurrency controls

OpenCode X doesn't pretend to replace everything Claude Code offers. Here's what's the same, what's better, and what's different:

npm install -g @sdeonvacation/opencode-x
opencode-x

Grab from Releases:

# macOS / Linux
chmod +x opencode-x-<platform>
sudo mv opencode-x-<platform> /usr/local/bin/opencode-x

# Windows (PowerShell as Admin)
move opencode-x-windows-x64.exe C:\Windows\opencode-x.exe

git clone https://github.com/sdeonvacation/opencode-x
cd opencode-x && bun install
./packages/opencode/script/build.ts --single
# Binary at: packages/opencode/dist/opencode-$(uname -s | tr A-Z a-z)-$(uname -m)/bin/opencode

# 1. Set any API key
export ANTHROPIC_API_KEY=sk-...    # or OPENAI_API_KEY, GOOGLE_API_KEY, etc.

# 2. Run
opencode-x

# 3. Type a prompt. That's it.

Already have Claude Code hooks in ~/.claude/settings.json? They work automatically.

Multiple systems work together to minimize cost — the single biggest advantage over Claude Code's raw-output approach:

LLM Compression — Large tool outputs are pre-processed by a fast model before reaching your main model. Three strategies selected automatically:

Falls back silently to raw output on error. Never corrupts results.

Cache Stability — System prompt split into stable prefix (cached across turns) and dynamic suffix (memory, goals, env). Dynamic content changes without invalidating the cached prefix.

Snapshot Gate — Skips expensive git diff tracking on read-only turns (grep, read, search).

Part Coalescer — Batches rapid streaming DB writes (300ms debounce window).

Sessions never crash from token overflow. Three layers trigger automatically:

Plus: reactive 413 recovery retries the same turn inline after emergency compaction. No user intervention needed.

/goal Implement authentication middleware with all tests passing

Agent works autonomously. After each turn, a separate judge model evaluates whether the condition is met. On "no," the agent keeps working with the judge's reasoning as guidance. Guards prevent runaway:

• 200-turn hard limit (configurable)
• Optional token budget
• Independent judge validation (prevents premature self-reported success)

One tool call spawns N parallel subagents with the same prompt template:

{ "experimental": { "swarm": true, "swarm_concurrency": 5 } }

• Items up to 128, concurrency 1-20
• Per-item failure isolation (one crash doesn't abort others)
• Background mode: dispatch and keep working, results auto-injected on completion
• Structured XML results with per-item status and timing

Mid-execution, press Leader+D. Running agent detaches to background. TUI immediately accepts new input. Toast notification fires on completion. Session remains in session list.

Route work to the right model automatically:

{
  "hybrid": {
    "enabled": true,
    "cheap_model": { "providerID": "anthropic", "modelID": "claude-haiku-4-5" }
  },
  "experimental": {
    "task_categories": { "review": { "providerID": "anthropic", "modelID": "claude-sonnet-4-6" } },
    "ultrawork_model": { "providerID": "anthropic", "modelID": "claude-sonnet-4-6" }
  }
}

• Compression routing — cheap model handles tool output compression, titles, compaction
• Category routing — tasks routed by type (review → sonnet, exploration → haiku)
• Ultrawork routing — use_ultrawork: true triggers premium reasoning model
• Concurrency limiter — per-model semaphore prevents API throttling

Configurable protections that Claude Code doesn't expose:

• Doom loop detection — ring-buffer hash catches repeated identical tool calls (default threshold: 5)
• Spawn depth limit — max 3 levels of nested subagents
• Descendant cap — max 50 total subagents per root session
• Per-model concurrency — semaphore prevents API rate limiting across parallel subagents
• Stream idle timeout — abort + retry stalled provider streams (default: 60s)

Persistent memory — cross-session facts in ~/.local/share/opencode/memory/. Three types: user preferences, project conventions, corrections. Injected into every session's system prompt. Agent learns from your corrections over time.

Session memory — SQLite-backed entries that survive /clear and compaction. Managed via /memory_add, /memory_edit, /memory_delete. Scoped to current session only.

Single LSP server per project, shared across all agent types — primary, subagents, background sessions, swarm workers, isolated worktrees. Path translation maps worktree paths to parent directory. 30-minute idle shutdown + auto-respawn.

No redundant server spawns. Consistent diagnostics everywhere. Zero LSP overhead multiplication.

Your existing setup works immediately with no migration:

Supported hook events: PreToolUse, PostToolUse, SessionStart, UserPromptSubmit, Stop

Safe concurrent execution of read-only tools within a single LLM round-trip. Permission-aware gate ensures only pre-approved, parallel-safe tools run concurrently:

Measured improvements over baseline opencode and raw LLM interaction:

Full reference in OPENCODE-X_GUIDE.md.

Key experimental features:

{
  "experimental": {
    "swarm": true,
    "swarm_concurrency": 5,
    "goal_system": true,
    "parallel_tool_calls": true,
    "microcompact": true,
    "context_collapse": true,
    "tool_result_budget": 50000,
    "prompt_split_caching": true,
    "worktree_isolation": true,
    "background_subagents": true,
    "loop_detector_threshold": 5,
    "max_subagent_depth": 3,
    "max_subagent_descendants": 50,
    "ultrawork_model": { "providerID": "anthropic", "modelID": "claude-sonnet-4-6" }
  },
  "hybrid": {
    "enabled": true,
    "cheap_model": { "providerID": "anthropic", "modelID": "claude-haiku-4-5" }
  }
}

Active development pushing further ahead:

• Runtime: Bun 1.3.11
• Language: TypeScript 5.8
• Framework: Effect-ts 4.0.0-beta
• TUI: Solid.js + @opentui
• Database: SQLite + Drizzle
• AI: Vercel AI SDK + 75+ providers

git clone https://github.com/sdeonvacation/opencode-x
cd opencode-x && bun install
bun run dev                              # TUI
bun run --cwd packages/opencode dev      # CLI
bun --cwd packages/opencode run build
bun --cwd packages/opencode test --timeout 30000

Requires Bun 1.3.11+.

• Repository: https://github.com/sdeonvacation/opencode-x
• Full Guide: OPENCODE-X_GUIDE.md
• Upstream: https://github.com/anomalyco/opencode
• Docs: https://opencode.ai/docs

MIT — See LICENSE.