@counted/opencode

Native OpenCode plugin for Counted — privacy-first analytics for AI agent dev sessions. No code, content, or PII ever leaves the machine.

Install

Add it to your opencode.json plugin array (OpenCode auto-installs it):

{
  "plugin": ["@counted/opencode"]
}

Or drop a re-export in .opencode/plugins/counted.ts (project) or ~/.config/opencode/plugins/counted.ts (global):

export { CountedPlugin } from "@counted/opencode";

Then set your project's client key in the environment:

export COUNTED_AGENT_KEY="ck_your_agent_project_client_key"
# Optional — defaults to https://app.counted.dev
export COUNTED_AGENT_HOST="https://app.counted.dev"

The plugin is a no-op until COUNTED_AGENT_KEY is set. Create the project with the agent dashboard template so the pre-built insights line up.

What it tracks

| OpenCode signal | Counted event | Props | | --- | --- | --- | | session.created | session_start | mode | | tool.execute.after (any) | tool_use | tool, outcome | | tool.execute.after (edit/write) | file_edit | filePath (repo-relative), action, language | | tool.execute.after (bash) | command_run | command (binary name only) | | session.idle / dispose | — (flush) | — |

Events run through a single long-lived @counted/sdk Analytics instance (batched, flushed on idle and on dispose).

Privacy

• File paths only (repo-relative; basename fallback), never contents or diffs.
• Command binary name only (git, bun, …) — no arguments, no output.
• No prompt text, no AI responses, no PII.

OpenCode surfaces tool failures on the event stream rather than the after-hook, so per-tool outcome is currently always success; failures aren't attributed to a specific tool yet.

Compare agent setups

Every event carries a setup fingerprint so you can break metrics down by agentic configuration:

• setupHash — a stable digest of your opencode.json setup: model, agent/ prompt definitions, tools, permissions, and sampling params. Only the digest is sent — prompt content never leaves the machine. setupHashVersion tracks the scheme.
• model — sent in the clear so breakdowns are readable.
• setupLabel — optional human bucket; set COUNTED_SETUP_LABEL="reviewer-v2".

Add a breakdown insight grouped by setupHash (or setupLabel) over tool_use outcome, command_run, or file_edit to compare setups.

Advanced

The package also exports the low-level helpers (init, track, trackToolUse, …) if you want to emit events from your own OpenCode plugin instead of using CountedPlugin directly.

License

MIT