Make opencode-go as capable as Claude Code or Codex — autonomous agents, 15 MCP servers, spec-driven dev, lossless token reduction, battle-tested coding-agent patterns. One-command install.
A portable, shareable installer that provisions opencode with everything it needs to be a top-tier autonomous coding agent — locked to the opencode-go subscription only. No Anthropic, OpenAI, Google, or any other provider is referenced.
One command installs:
- 8 custom agents with per-model orchestration (build, plan, reviewer, explorer, data-engineer, frontend-engineer, devops-engineer, doc-writer)
- 7 slash commands (/review, /plan, /test, /ship, /debug, /refactor, /docs)
- 12–15 MCP servers (playwright, github, context7, sequential-thinking, fetch, filesystem, memory, time, git-utils, repomix, code-review-graph, pakt, + more in full tier)
- 22 hand-authored skills covering reasoning disciplines, spec adherence, token awareness, Windows tooling, domain patterns
- 4 third-party skill packs (superpowers, CodeRabbit, farmage, addyosmani) in full tier
- Spec-driven-development spine (GitHub Spec Kit + superpowers gated workflow + skillgate DoD gate)
- Lossless token-reduction stack (prompt caching, compaction prune, subagent delegation, ast-outline, code-review-graph ~82x reduction, PAKT lossless compression)
- External CLIs (ast-grep, ast-outline, gtr git-worktree-runner)
- Model toggle for single↔multi-model switching
- Uninstaller that safely reverts to stock opencode
- opencode installed
- opencode-go subscription authenticated (
/connect→ select OpenCode Go → paste API key) git,node/npm(required);python/uvx,gh(recommended)
git clone https://github.com/iamwaqargulzar/opencode-go-power-pack.git
cd opencode-go-power-pack
.\install.ps1
.\verify.ps1
# Restart opencodegit clone https://github.com/iamwaqargulzar/opencode-go-power-pack.git
cd opencode-go-power-pack
chmod +x install.sh verify.sh
./install.sh
./verify.sh
# Restart opencode# Windows
.\install.ps1 -Tier full # maximum capability (15 MCP, 4 skill packs)
.\install.ps1 -Tier lean # minimal (10 MCP, superpowers + custom skills)
.\install.ps1 -Tier standard # default (12 MCP, CodeRabbit, Spec Kit, DoD gate)
.\install.ps1 -SkipMcp # skip MCP prefetch
.\install.ps1 -DryRun # preview without installing# macOS / Linux
./install.sh --tier full
./install.sh --tier lean
./install.sh --skip-mcp
./install.sh --dry-run| Component | lean | standard (default) | full |
|---|---|---|---|
| MCP servers | 10 | 13 | 18 |
| Custom skills | 22 | 22 | 22 |
| Third-party skill packs | superpowers | + CodeRabbit | + farmage(66) + addyosmani(24) + cherry-pick(10) |
| External CLIs | gtr | + ast-outline + ast-grep | + jscodeshift + libcst + aider |
| Spec-driven spine | AGENTS.md only | + GitHub Spec Kit | + APM (multi-agent) |
| DoD gate | AGENTS.md only | skillgate | + donegate (CI) |
| Drift detector | — | — | coherence |
| Reasoning MCP | sequential-thinking | + think | + aleph |
| Lossless context MCP | repomix | + code-review-graph + PAKT | + arbor + provenant + headroom |
See docs/tiers.md for the full breakdown.
| Agent | Mode | Model | Purpose |
|---|---|---|---|
build |
primary | opencode-go/glm-5.2 |
Everyday coding, default agent |
plan |
primary | opencode-go/deepseek-v4-pro |
Read-only analysis & planning |
reviewer |
subagent | opencode-go/qwen3.7-max |
Code review, security audit |
explorer |
subagent | opencode-go/deepseek-v4-flash |
Fast cheap codebase recon |
data-engineer |
subagent | opencode-go/kimi-k2.7-code |
Data/ML/notebooks/SQL |
frontend-engineer |
subagent | opencode-go/glm-5.2 |
React/Vue/Next + Playwright |
devops-engineer |
subagent | opencode-go/deepseek-v4-pro |
Docker/K8s/Terraform/CI |
doc-writer |
subagent | opencode-go/mimo-v2.5 |
Docs/ADRs/READMEs |
See docs/agents.md for full agent reference.
| Command | Agent | What it does |
|---|---|---|
/review |
reviewer | Review diff for bugs, security, style |
/plan |
plan | Produce step-by-step implementation plan |
/test |
build | TDD — write failing tests, implement, verify |
/ship |
build | Lint + typecheck + test + commit + push |
/debug |
build | SWE-agent 5-step debug protocol |
/refactor |
build | Analyze → propose → apply with ast-grep |
/docs |
doc-writer | Generate README sections, ADRs, API docs |
| Server | Purpose |
|---|---|
| playwright | Browser automation, screenshots, clicks |
| github | Repo/PR/issue ops (token from gh auth token) |
| context7 | Library docs on-demand (never web-search for APIs) |
| sequential-thinking | Structured step-numbered reasoning |
| think | Anthropic "think" tool (+1.6% SWE-bench, d=1.47) |
| fetch | JS-rendered web fetch |
| filesystem | Cross-directory file ops |
| memory | Persistent cross-session knowledge graph |
| time | Current time / timezones |
| git-utils | Git operations via MCP |
| repomix | AST-compressed repo-to-text packing |
| code-review-graph | Blast-radius analysis (~82x token reduction) |
| pakt | Lossless JSON/YAML/CSV/MD compression (p=0.50) |
See docs/mcp-servers.md for the full list including full-tier servers.
Custom skills: opencode-go-workflow, windows-powershell-tooling, autonomous-execution, token-budget-awareness, repo-context-packing, mcp-orchestration, git-worktree-parallelism, test-first-verification, docs-as-code, data-pipeline-patterns, docker-compose-first, ci-pipeline-design, secrets-hygiene, error-recovery-protocol, progressive-refinement, ears-requirements, property-based-verification, reflexion-loop, tot-decide, swe-debate-self-critique, swe-agent-5-step-protocol, submit-time-self-review.
Third-party packs (standard+): CodeRabbit (code-review, autofix). Full tier: superpowers (14), farmage (66), addyosmani (24), cherry-picked top 10 from skills.sh.
See docs/skills.md for the full reference.
Switch between multi-model orchestration and single-model simplicity:
# Windows
.\toggle-models.ps1 status # show current mode
.\toggle-models.ps1 multi # per-agent models (default)
.\toggle-models.ps1 single # all agents use top-level model
.\toggle-models.ps1 single opencode-go/kimi-k2.7-code # single + pick model# macOS / Linux
./toggle-models.sh status
./toggle-models.sh multi
./toggle-models.sh single
./toggle-models.sh single opencode-go/kimi-k2.7-codeSee docs/model-toggle.md for details.
Safely revert to stock opencode:
# Windows
.\uninstall.ps1 # restore config, keep skills
.\uninstall.ps1 -Full # complete removal
.\uninstall.ps1 -Full -DryRun # preview# macOS / Linux
./uninstall.sh # restore config, keep skills
./uninstall.sh --full # complete removal
./uninstall.sh --full --dry-run # previewThe uninstaller reads the manifest (~/.config/opencode/.power-pack-manifest.json) written by the installer and removes only what was installed. If a backup exists, it restores it; otherwise it writes a minimal stock config.
The standard tier installs GitHub Spec Kit as the spec spine:
cd your-project
specify init . --integration opencodeThis adds /speckit.* slash commands: constitution → specify → clarify → plan → tasks → implement → converge (reconciles code-vs-spec). Combined with superpowers' gated workflow and skillgate's deterministic DoD gate, this keeps code aligned with spec.
See docs/spec-driven.md for the full workflow.
The pack uses a layered approach that reduces tokens without hindering capabilities:
- Prompt caching (
setCacheKey: true) — 90% input cost reduction on cache hits, truly lossless. - Compaction prune — drops old tool outputs from context.
- Subagent delegation — explorer/reviewer run in child sessions; only results return.
- ast-outline —
digest→outline→showprogressive disclosure (truly lossless, exact signatures + exact source). - code-review-graph — blast-radius analysis, ~82x median token reduction (38-528x), benchmarked.
- PAKT — lossless JSON/YAML/CSV/MD compression, comprehension-evaluated (p=0.50 sign test).
Excluded: LLMLingua (lossy — drops tokens, conflicts with "no capability hindrance" requirement).
See docs/token-reduction.md for the full stack.
enabled_providers: ["opencode-go"] in opencode.json hard-locks the setup to the opencode-go subscription. Even if other provider credentials exist in your environment, opencode ignores them. No Anthropic, OpenAI, Google, xAI, or any other provider is referenced anywhere in the pack.
- Windows: MCP commands use
["cmd","/c","npx",...]form. The installer handles this automatically. - macOS/Linux: The installer rewrites MCP commands to
["npx",...]on install. - RTK plugin: On native Windows, the auto-rewrite hook doesn't run (falls back to CLAUDE.md-style injection). Full hook support requires WSL. The
rtkCLI still works manually. - gtr (git-worktree-runner): On native Windows, may need WSL. The installer prints instructions.
See CONTRIBUTING.md for how to add skills, agents, MCP servers, and improve the installer.
See CHANGELOG.md.
MIT — see LICENSE. Third-party packs bundled by the installer retain their original licenses (MIT, Apache-2.0, MPL-2.0).
opencode opencode-go ai-agents mcp skills spec-driven coding-agent llm developer-tools