opencode

opencode integrates with TokenOps on three planes — pick as many as you want.

Passive reader (zero wiring)

opencode stores sessions in a local SQLite database. TokenOps reads it read-only and attributes every assistant turn per project and per session — no proxy, no keys, no change to how you run opencode.

tokenops vendor-usage enable opencode      # reads ~/.local/share/opencode/opencode.db
tokenops start                             # restart to pick up the poller
tokenops vendor-usage status               # confirm the source is emitting

The reader is multi-provider: whichever model/provider a session used (anthropic, openai, github-copilot, google, openrouter, or anything else opencode routes to) is attributed correctly. Use --root to point at a non-default database path.

MCP tools (agent-side)

opencode is an MCP host, so it can call TokenOps directly:

// opencode.json
{ "mcp": { "tokenops": { "command": "tokenops", "args": ["serve"] } } }

The agent can then ask for tokenops_session_budget, tokenops_burn_rate, tokenops_plan_headroom, and call tokenops_status to discover what data sources are live and how to upgrade signal quality.

Proxy (ground-truth metering)

Point opencode's provider base URL at the TokenOps proxy to meter every request/response with exact token counts and live optimization. For an OpenAI-compatible provider:

TOKENOPS_STORAGE_ENABLED=true tokenops start

Set the provider's baseURL in opencode.json to http://127.0.0.1:7878/<provider>/v1 (e.g. /openai/v1, /openrouter/v1, /groq/v1). See Coverage for the full provider list.