# Opencode Build Tool

> Build custom tools and plugins using the OpenCode SDK packages. Use when creating custom tools, implementing tool schemas, building plugins, managing OpenCode sessions, using opencode-ai/sdk or opencode-ai/plugin, creating a tool.ts file, writing an execute handler, using tool.schema, ToolContext, createOpencode, plugin hooks, custom auth, automating OpenCode from a script, building a CI tool, or integrating external APIs. Keywords: Zod schema, execute handler, event streaming, session management, abort signal, tool discovery, multiple tools per file, tool factory.

- Skill: `pantheon-org/opencode-build-tool` (Agent Skill, multi-file: 20 files)
- Install (CLI): `npx skillmds@latest add pantheon-org/opencode-build-tool`
- Raw SKILL.md: https://api.skillmd.com/api/skills/pantheon-org/opencode-build-tool/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Product & Planning
- Author: pantheon-org (https://skillmd.com/u/pantheon-org)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/pantheon-org/opencode-build-tool

---


# OpenCode SDK Development

## Quick Start: Custom Tool in 3 Steps

1. Create `.opencode/tool/my-tool.ts`:

```typescript
import { tool } from "@opencode-ai/plugin"

export default tool({
  description: "Brief description of what the tool does",
  args: {
    query: tool.schema.string().describe("Search query")
  },
  async execute({ query }, { abort }) {
    return `Result for: ${query}`
  }
})
```

1. Restart OpenCode — tools are auto-discovered.
1. Filename = tool name: `my-tool.ts` → `my-tool`.

## Mindset

- `@opencode-ai/plugin` — **Tools + Hooks**: Add LLM-callable tools, intercept execution, react to events
- `@opencode-ai/sdk` — **Client API**: Control OpenCode from external scripts, session management

**When to use**: Adding new capabilities the agent doesn't have, automating OpenCode from CI/scripts, or integrating external APIs as first-class tools.
**When NOT to use**: Built-in tool exists (`client.tool.list()`). Need shell (use `$` from Bun). Logic belongs in system prompt (`AGENTS.md`). Need MCP.

> Full decision guide in [advanced-patterns.md](references/advanced-patterns.md).

## Tool File Locations

```bash
.opencode/tool/*.ts          # project-specific
~/.config/opencode/tool/*.ts  # global
```

Multiple named exports per file → multiple tools: `filename_exportname`

## Anti-Patterns

NEVER return non-string from `execute`. WHY: The tool runtime expects a string response; returning an object throws at runtime.
```typescript
// BAD - throws at runtime
async execute({ query }) { return { results: [] } }
// GOOD - serialize first
async execute({ query }) { return JSON.stringify({ results: [] }) }
```

NEVER `import { tool } from "@opencode-ai/sdk"`. WHY: The SDK package has no `tool()` export. The tool factory is only in `@opencode-ai/plugin`.
```typescript
// BAD - import error at runtime
import { tool } from "@opencode-ai/sdk"
// GOOD
import { tool } from "@opencode-ai/plugin"
```

NEVER ignore the `abort` signal in long-running operations. WHY: Ignoring abort leaves processes dangling after the user cancels, consuming resources.
```typescript
// BAD - dangling fetch after cancel
async execute({ url }) { return await fetch(url).then(r => r.text()) }
// GOOD - propagate abort
async execute({ url }, { abort }) { return await fetch(url, { signal: abort }).then(r => r.text()) }
```

## Reference Files

- [advanced-patterns.md](references/advanced-patterns.md) — patterns, anti-patterns, SDK examples; [sdk-api.md](references/sdk-api.md) — SDK client API; [plugin-api.md](references/plugin-api.md) — plugin hooks
- [basic-tool.ts](references/basic-tool.ts) — simple tool; [full-plugin.ts](references/full-plugin.ts) — complete plugin

Run `bun run opencode run "list tools"` to verify your tool is discovered and named correctly.

```yaml
---
description: custom tools opencode SDK plugin execute handler abort signal Zod schema session management event streaming
---
```

## Eval Scenarios

- [Scenario 0: Create a custom todo-search tool with schema validation](evals/scenario-0/task.md)
- [Scenario 1: Use SDK client to create session and send prompt](evals/scenario-1/task.md)
- [Scenario 2: Implement abort signal handling in long-running tools](evals/scenario-2/task.md)