OpenCode telemetry dashboard

This is a zero-dependency Node 24+ dashboard for aggregate telemetry from opencode.db. It uses the built-in node:sqlite module and opens the database read-only. The server binds only to 127.0.0.1 and makes no prompt, plugin, model, or external network requests.

Run

Requires Node.js 24 or newer. Clone and run it from any path:

git clone https://github.com/sandvich1337/opencode-telemetry-dashboard.git
cd opencode-telemetry-dashboard
npm ci
npm start

Then open http://127.0.0.1:4173. Use --db PATH to select a database and --port PORT to select a loopback port. Pass --open to launch the default browser; the .cmd launcher opens it automatically. GET /health returns a database-independent health check.

The database can also be selected with OPENCODE_DB_PATH. Without that variable, the dashboard checks the standard OpenCode data locations under the user profile.

The visible page refreshes every 60 seconds only while the document is visible; hidden tabs do not poll. The selector lists root sessions only. Selecting a root includes that root and every descendant, so its cards summarize the complete session run. Parent-child sessions are shown in a compact spawn topology; clicking an individual child scopes the cards and tables to that child alone. Tool usage is split into collapsible agent blocks in both views.

The dashboard navigation keeps the legacy summary and sortable tables visible while exposing Architecture, Throughput, Investigate, and Environment views. For finite ranges, the investigation views request an equal-length preceding period with the same session and privacy options. Throughput comparisons show the denominator and sample evidence; unavailable analytics remain explicitly unavailable rather than becoming zero.

The Export session contents control is separate from the aggregate dashboard. It is enabled only for one selected root and requires an explicit confirmation that the resulting unredacted, raw SQLite snapshot may contain prompts, responses, identifiers, paths, tool inputs/outputs, and secrets. The archive is snapshot-only: it does not claim live telemetry, retries, causation, or unpersisted runtime events. Normal pages continue to hide raw content and existing aggregate downloads do not include it. Archives are generated in the browser with deterministic JSON and an uncompressed ZIP; users are responsible for deleting any downloaded copies.

Session archives use the structured session-contents-v2 format. The response contains eight canonical UTF-8 JSON files: manifest.json, the three typed raw-table snapshots, timeline.json, metrics.json, transcript.json, and chat.json. The manifest records the selected root tree, snapshot consistency, file sizes, and SHA-256 digests. chat.json contains ordered session messages, unattached parts, and semantic references back to the verified raw rows; missing tool metadata is omitted rather than inferred.

Navigation and time-range controls are keyboard accessible and remain usable at narrow widths; changing a filter keeps the latest request authoritative. The configured-inventory opt-in is withdrawn from the visible view immediately, before the refreshed request completes.

Configured environment inventory is shown only after both the local checkbox and the server request opt in. The Environment view can save a redacted local baseline and compare it with later data, or download the allow-listed aggregate inventory as deterministic JSON or principal CSV. Downloads are generated in the browser and never send data to a remote service.

Metrics

Only session, message, and part are read. The dashboard reports runs and subagents grouped by agent/model, token splits, estimated API-equivalent cost, tool/read/error counts, duration, metadata file additions/deletions, and aggregate reviewer-slot totals. Reviewer summaries include ISSUE/PASS/UNKNOWN percentages per slot and the ISSUE rate for the second reviewer after a first PASS.

The dashboard displays no reviewer rows and does not surface raw message text, tool input/output, IDs, file paths, database paths, prompts, or plugin data by default. The optional Show root titles control groups root sessions by their OpenCode title; child titles are never sent. Root titles can contain prompt-derived text, so this opt-in is off on every page load. Schema capability checks are surfaced instead of guessing when a required table or column is unavailable.

Input and output cards show their aggregate token totals separately. OpenCode records provider usage per model step, not per tool. The dashboard therefore splits each token bucket evenly across calls in a conserved step; if step data is incomplete it falls back to all tool calls in that message. Steps without a tool remain Unattributed / no tool. This is an allocation estimate, not causal per-tool metering, and the rows always conserve the recorded total. Aggregate results are cached for the current view and refreshed on the visible-only interval; the cache is read-only and does not modify opencode.db.

Cache-write totals distinguish a reported zero from missing provider telemetry. When no selected message reports a cache-write field, the dashboard shows Not reported instead of presenting the missing data as a measured zero.

Estimated USD

The dashboard reports a current API-equivalent estimate, not an invoice:

USD = (input × input_rate
     + cache_read × cache_read_rate
     + cache_write × cache_write_rate
     + (output + reasoning) × output_rate) / 1,000,000

Rates are USD per million tokens. The catalog date is 2026-08-17 and its source is https://platform.openai.com/docs/pricing. Rates are listed as input / cache-read / cache-write / output:

Model	Short context	Long context
luna	0.20 / 0.02 / 0.25 / 1.20	0.40 / 0.04 / 0.50 / 1.80
terra	2.00 / 0.20 / 2.50 / 12.00	4.00 / 0.40 / 5.00 / 18.00
sol	5.00 / 0.50 / 6.25 / 30.00	10.00 / 1.00 / 12.50 / 45.00

The *-fast catalog variants use twice the corresponding rate. The long context tier applies when input + cache-read + cache-write exceeds 272,000 tokens (or when long context is explicitly selected). Supported gpt-5.6-luna, gpt-5.6-terra, and gpt-5.6-sol names map to these catalog entries.

Privacy is local and aggregate-only: no prompts, raw tool payloads, IDs, paths, plugin data, or database contents are sent anywhere. The dashboard adds zero prompt, plugin, model, or network overhead.

CLI and npm

npm start
npm run metrics -- --json
npm run metrics -- --db path/to/opencode.db --range 7d

The metrics script emits aggregate JSON. The metrics:dashboard script starts the loopback dashboard and opens the browser. Neither command modifies the database.

Verify

npm ci
npm test