omp-ollama-cloud

oh-my-pi (omp) plugin that registers the Ollama Cloud provider with the model catalog from srnoob2570/ollama-cloud-catalog (official rate card, thinking efforts), adds a live streaming-stats widget (tok/s, TTFT), and fetches real account quota from ollama.com's usage API for omp's /usage surfaces.

Ported from @srnoob2570/opencode-ollama-cloud. omp ships its own /model command, so that one is not ported.

Install

omp plugin install @srnoob2570/omp-ollama-cloud

For development:

omp plugin link /path/to/omp-ollama-cloud

The plugin only supplies models and pricing; omp resolves the credential itself. Set OLLAMA_CLOUD_API_KEY, or run omp /login and pick Ollama Cloud.

Models

The plugin replaces omp's built-in discovery list for ollama-cloud with the published catalog, so omp models shows the real lineup: context and max-output limits, thinking efforts, and image support per model.

What it does

• Registers the ollama-cloud provider through pi.registerProvider, with fetchDynamicModels fed by the published catalog artifact. The catalog's models replace omp's built-in discovery list for this provider.
• Attaches each model's official rate card (USD per 1M tokens) from the artifact, so the cost counter, /usage, and omp-stats show real spend. These are the off-peak rates Ollama Cloud publishes; the cache-read rate maps to cost.cacheRead.
• Renders one line below the editor, 42.3 tok/s · TTFT 812 ms, from omp's own AssistantMessage timing (duration, ttft); it never touches the wire. TPS is total output tokens over total stream time. Ollama Cloud flushes output in bursts, so a decode-only window would report burst speed, sometimes 500+ tok/s. TTFT is the simple mean per step. Only main-conversation ollama-cloud assistant steps count. Subagents run in their own runners, and title generation, auto-thinking, and compaction never reach the main runner as assistant messages.
• Probes the npm registry once per boot (npm installs only; dev checkouts are skipped). When a newer release exists, a second widget line names the version and the install command. The plugin never mutates its own install.
• Fetches the account's real quota from GET https://ollama.com/api/usage and feeds omp's normalized usage pipeline: /usage in the TUI and ACP sessions, the status-line footer (polled every 5 min), recorded usage history (omp usage --history), and credential health checks. Session and weekly windows render as percent bars with per-model request notes; a nonzero four-week activity cost adds a USD row. omp resolves the credential itself (stored login key or OLLAMA_CLOUD_API_KEY). The plugin sends the key only as a Bearer header to a pinned origin. It re-verifies the fetch target as https://ollama.com/api/usage immediately before every call, ignores params.baseUrl redirections, and never logs or persists the key. Every consumed response field is validated before use, and nothing from the response is executed or written to disk.
• Adds /ollama-recost, a one-shot sweep over every saved session file under ~/.omp/agent/sessions (nested subagent transcripts included). It re-prices the $0 ollama-cloud lines from the rate card, then drops omp-stats' incremental offsets so the dashboard's next sync re-parses the files (its upsert updates the cost columns). The offset reset covers every session file, including files that needed no rewrite, because a file can be fully priced while its dashboard rows were synced from an older snapshot, and re-parsing is idempotent. Reports files rewritten and requests re-priced.

Usage surfaces

The omp usage CLI command prints the account's quota windows with per-model request counts. It reads what an open omp session has already fetched, so the numbers only appear while an omp window is open and has queried the usage at least once. With no live session it shows the built-in stub's "no limits reported". Use omp usage --history instead; it works from the recorded snapshots alone.

omp usage --history records snapshots and renders the trend per window:

Knobs

Environment variables, read once at startup. off, 0, false, and no all turn a knob off.

Variable	Default	Effect
OMP_OLLAMA_CLOUD_PRICING	on	off zeroes every cost block (counter stays at $0.00)
OMP_OLLAMA_CLOUD_STATS	on	off disables the widget and the collector
OMP_OLLAMA_CLOUD_COST_FIX	on	off disables the session-cost patch (see below)
OMP_OLLAMA_CLOUD_USAGE	on	off restores omp's built-in stub (reports "no quota API")
OMP_OLLAMA_CLOUD_BASE_URL	https://ollama.com	provider endpoint, e.g. a local proxy. Trailing slashes are stripped
OMP_OLLAMA_CLOUD_API	ollama-chat	wire API the provider speaks. openai-completions needs a base URL ending in /v1

omp loads ~/.omp/agent/.env at startup, so a line such as OMP_OLLAMA_CLOUD_BASE_URL=http://127.0.0.1:11435 routes every request through a local proxy over the native Ollama dialect. For a proxy that speaks the OpenAI dialect, like ollama-cloud-meter, keep the /v1 suffix and add the API knob:

OMP_OLLAMA_CLOUD_BASE_URL=http://127.0.0.1:11435/v1
OMP_OLLAMA_CLOUD_API=openai-completions

The quota fetcher keeps calling https://ollama.com/api/usage directly; only model traffic is redirected.

omp caches the dynamic model list per provider in ~/.omp/agent/models.db with a 24 h TTL. After flipping OMP_OLLAMA_CLOUD_PRICING, run omp models refresh to force a fresh fetch and see the change immediately.

The extension system only runs inside sessions, so the standalone omp usage CLI command keeps showing the built-in stub's "no limits reported" for Ollama Cloud regardless of this knob; the session surfaces (/usage, footer, --history) show the real numbers.

Session cost patch

omp's ollama-chat adapter never prices usage. It writes usage.cost as $0 for every request, so the cost counter, /usage, and omp-stats all show $0 for Ollama Cloud. The plugin re-prices the affected session lines at turn end from the model's rate card. OMP_OLLAMA_CLOUD_COST_FIX=off disables the patch. The plugin never touches lines that already carry a non-zero cost, so the patch becomes a no-op once upstream omp fixes the adapter.

Because the patch rewrites lines in place (byte lengths change), the plugin also resets the patched file's incremental offset in omp-stats' stats.db, so the dashboard's next sync re-parses it. /ollama-recost does the same for the full history. Lines from the turn in flight are priced when the turn ends.

Catalog upstream

Scheduled GitHub Actions in srnoob2570/ollama-cloud-catalog publish the artifact: models.dev shape plus the x_ollama extension (reasoning_options, quantization, rate card, hash gate). Two mirrors serve the same file; the loader races them in parallel with a 5 s timeout each and takes the first response that passes isCatalog. On total failure it falls back to the disk cache at ~/.cache/omp-ollama-cloud/catalog.json, refreshed on every successful fetch. With no network and no cached copy the provider lists no models until a fetch succeeds.

Development

bun install
bun test
bun run typecheck

omp plugin link . for a live checkout; changes reload with a new omp process.

License

MIT