# SuperPenguin by Carrot Labs: complete product reference

> SuperPenguin is an AI spend attribution platform. Providers can each tell you what you spent with them, and several now publish their own split between coding-agent spend and API spend, but none of them knows your customers, features, teams, or pull requests. SuperPenguin connects every provider billing account into one total and then attributes that total to the thing in your business that caused it. SuperPenguin observes provider calls and reads provider billing APIs; it is not a proxy and is not in your request path.

## What SuperPenguin adds that a provider dashboard cannot

An AI bill can be read at four levels of detail. The platform you pay (OpenAI,
Anthropic, Cursor). The surface inside that platform (the Codex agent versus the
Platform API; Claude Code versus the Anthropic API). The thing in your business
that caused the spend (a customer, a feature, a team, a pull request). And the
individual request.

Providers publish the first two. Several of them separate agent spend from API
spend in their own consoles, and that separation is real. What no provider
publishes is the third level, because a customer id, a feature name, a team, and
a repository are your vocabulary rather than theirs. Reading them requires
correlating every provider at once against labels only you can supply. That is
the level SuperPenguin operates at, and it is one capability applied to two
domains.

The surface a request came through determines which questions can honestly be
asked of it. API spend serves an end customer, so it attributes to customers,
features, teams, and environments. Coding-agent spend does not: there is no
customer behind a Codex session, so it attributes to teams, projects, and merged
pull requests. SuperPenguin does not invent a customer for agent spend, and any
figure grouped by customer is explicitly a subset of the total rather than the
whole bill rescaled.

## Connect an AI agent with the CLI or MCP

Terminal-based agents can use the SuperPenguin CLI. Install on macOS or Linux:
curl -fsSL https://install.superpenguin.ai | sh
Then run superpenguin auth login --native interactively and pause while the human
signs in, chooses an organization, reviews scopes, and approves access. Run
superpenguin context --agent first, discover relevant operations with
superpenguin capabilities list --q "weekly spend" --agent, and inspect command help.
The --agent flag returns structured JSON and disables prompts; it does not
request Markdown. Use superpenguin usage summary --window month_to_date --agent
for SDK/OpenTelemetry spend, and superpenguin billing summary --window month_to_date
--agent for explicit provider invoice or reconciliation questions. Start coding
investigations with superpenguin coding usage-summary --tool cursor --scope self
--window this_week --agent; use coding sessions --status observed for drill-down.
For analytics SQL, inspect analytics schema, submit one bounded query with a UUID
idempotency key, and resume its handle rather than submitting again to poll.
CLI setup, field-selection examples, and troubleshooting: https://superpenguin.ai/docs/cli

SuperPenguin exposes an OAuth-authenticated remote Streamable HTTP MCP server at
https://app.superpenguin.ai/api/mcp. No API key or OAuth client secret is needed.
The MCP client starts discovery and opens SuperPenguin in the user's browser; the
human signs in, chooses an organization, reviews scopes, and approves access.
The grant inherits the user's current organization membership and role, and each
operation rechecks live permissions.

Setup and agent-ready instructions: https://superpenguin.ai/docs/mcp

Programmatic REST clients can use the same OAuth grant and operation contracts.
The REST base is https://app.superpenguin.ai/api/agent/v1 and the authenticated,
scope-filtered OpenAPI 3.1 document is at
https://app.superpenguin.ai/api/agent/v1/openapi. Setup and security guidance:
https://superpenguin.ai/docs/agent-api. SDK ingestion API keys are not Agent API credentials;
unattended service accounts and OAuth client credentials are not supported yet.

The server keeps the upfront tool catalog small. An agent calls get_context once,
uses discover to load only the relevant operation schemas and examples, and calls
query with an operationId returned by discovery. For durable analytics, it submits
once and resumes the returned handle instead of creating duplicate jobs to poll.
Missing cost is unknown, not zero; billed, SDK-observed, credit-inclusive,
pending, and API-equivalent values remain separately labeled.

## Products

### Attribution (https://superpenguin.ai/attribution)

What it is: Cost per customer, feature, and team
Setup cost: Two lines of code, once.

Tags every AI request with dimensions you define (customer, feature, team, environment) alongside ones we fill in (model, provider, prompt version), then reports cost in dollars grouped by any of them. There are no cardinality caps. Setup is two lines of code using the Python or TypeScript SDK, which supports OpenAI, Anthropic, Gemini, Bedrock, the Vercel AI SDK, and LiteLLM. Alerts to Slack, email, or Discord can be scoped to any of these dimensions, so a notification names the customer or feature that moved instead of reporting an organization-wide total; alerts are available from the Growth plan. By default the SDK records metadata only: token counts, model, latency, and your tags. Prompt and response text are captured only if an organization opts in, and are then sampled, redacted by default, encrypted at rest, and deletable by an org owner.

### Coding ROI (https://superpenguin.ai/coding-roi)

What it is: Coding agent spend, per pull request
Setup cost: Mac app and API key connections.

The same attribution capability applied to coding agents rather than to production API traffic. Reports what each merged pull request cost across Cursor, Claude Code, and Codex, broken down by project and across the team, and keeps visible the agent spend that never reached a merge. A coding-agent session has no end customer behind it, so agent spend attributes to teams, projects, and pull requests rather than to customers or product features. Because coding agents report cost differently, every figure is labeled with a cost basis (billed, API-equivalent, or estimated) and a match confidence tier rather than being averaged into one unlabeled number. Tools with a team API key need only that key plus a GitHub connection. Claude Code and Codex keep usage on the developer's machine, so those require the free macOS menu-bar app. PR attribution is available on every plan: Free shows 3 merged PRs total with no monthly reset, Growth includes the 20 most recent PRs merged this calendar month (UTC), and Pro includes unlimited PRs. AI coding analytics includes 1 person on Free, 3 Mac app users on Growth, 5 on Pro, and unlimited on Enterprise. The Mac app itself is free.

### One View (https://superpenguin.ai/one-view)

What it is: Every provider bill, as one total
Setup cost: Paste an admin key. No code.

The layer the other two products measure against, rather than a separate product to choose between them. Connects provider billing accounts and shows one total across all of them, in dollars, updated through today. Breaks that total down by provider, model, project, and API key, and reconciles what we tracked against the invoice the provider actually issued, naming the drift rather than smoothing it. Setup is pasting an admin key; no code. An end-of-month forecast is available from Pro, and projects the organization total rather than a figure per customer or per team.

## Providers

Billing sync covers 14+ providers and gateways wherever the
provider exposes an admin billing API, including OpenAI, Anthropic, Google
Gemini, AWS Bedrock, Azure OpenAI, Deepgram, ElevenLabs, OpenRouter, Vercel AI
Gateway, Together AI, and Fireworks AI. Per-request attribution through the SDK
additionally covers 100+ providers via LiteLLM.

## Pricing

Pricing is based on managed AI spend, meaning the AI spend SuperPenguin sees for
you in a calendar month across connected providers. It is not per seat and not
per request.

- Free: $0, up to $2,000 managed spend per month. For individuals getting started
- Growth: $30 per month, up to $5,000 managed spend per month. For small teams shipping AI
- Pro: $200 per month, up to $20,000 managed spend per month. For teams scaling AI in production
- Enterprise: Custom, no managed spend cap. For organizations with high volume

### Feature availability by plan

#### Limits

- Managed spend per month (Free: Up to $2K; Growth: Up to $5K; Pro: Up to $20K; Enterprise: Unlimited)
- Workspace members (dashboard access) (Free: 1; Growth: 3; Pro: 10; Enterprise: Unlimited)

#### Attribution

- Per-request cost in dollars (Free: yes; Growth: yes; Pro: yes; Enterprise: yes)
- Your own dimensions, no cardinality caps (Free: yes; Growth: yes; Pro: yes; Enterprise: yes)
- Python and TypeScript SDKs (Free: yes; Growth: yes; Pro: yes; Enterprise: yes)
- Custom model pricing (Free: no; Growth: yes; Pro: yes; Enterprise: yes)

#### Coding ROI

- Download the Mac app to see your own live agent spend (Free: yes; Growth: yes; Pro: yes; Enterprise: yes)
- People sharing Mac app activity for Coding ROI (Free: 1; Growth: 3; Pro: 5; Enterprise: Unlimited)
- PR cost attribution (Free: 3 merged PRs total; Growth: Latest 20 merged PRs this month; Pro: Unlimited PRs; Enterprise: Unlimited PRs)

#### One View

- Every provider in one total (Free: yes; Growth: yes; Pro: yes; Enterprise: yes)
- Breakdown by provider, model, project, and API key (Free: yes; Growth: yes; Pro: yes; Enterprise: yes)
- Reconciliation against your invoice (Free: yes; Growth: yes; Pro: yes; Enterprise: yes)
- 90-day historical backfill (Free: no; Growth: yes; Pro: yes; Enterprise: yes)
- End-of-month spend forecast (Free: no; Growth: no; Pro: yes; Enterprise: yes)

#### Optimization

- Opt-in prompt and outcome capture (Free: no; Growth: no; Pro: yes; Enterprise: yes)

#### Alerts and access

- Slack alerts (Free: no; Growth: yes; Pro: yes; Enterprise: yes)
- Email alerts (Free: no; Growth: yes; Pro: yes; Enterprise: yes)
- Discord alerts (Free: no; Growth: yes; Pro: yes; Enterprise: yes)

## Common questions

**What is SuperPenguin?**

SuperPenguin is an AI spend intelligence platform: it turns what you spend across every AI provider and coding tool into cost you can attribute, understand, and control. It lets teams see what each customer, feature, team, or pull request actually costs, catch spikes before they reach the invoice, and forecast where the month is heading, without ever sitting in the path of a request.

**How do I try SuperPenguin?**

Start free, no credit card. Pick the entry point that fits:

**Do you store my prompts or responses?**

By default, no. SuperPenguin records cost metadata only: token counts, model, latency, and the tags you set, never your prompt or response text. Content capture is off unless your organization turns it on in Settings, and even then it is sampled rather than complete, text-only with images and audio stripped, redacted by default, and encrypted at rest. An owner can delete captured content at any time.

**Can I use this with my team?**

Yes. SuperPenguin is a shared workspace, not a personal account. Connect your providers and SDK once, invite your teammates, and manage AI spend together, with roles that control who can add connections or change billing.

**How is this different from my provider's dashboard?**

Your provider's dashboard shows you their bill, by their project, their model, and their API keys. It cannot show you the other providers you pay, cannot group by your customer or your feature, and cannot tell you what a pull request cost. SuperPenguin puts every provider in one total and lets you slice it by dimensions you define.

**Will the SDK slow down my API calls?**

No. Your calls go straight to the provider; SuperPenguin is not a proxy and is not in your request path. Telemetry is sent after your response has already returned, which adds under 10ms on non-streaming requests and under 5ms to time-to-first-token on streaming ones.

**How can I track OpenAI or Anthropic spend by customer?**

Install the SuperPenguin Python or TypeScript SDK, wrap your provider client, and pass a customer id in the metadata on each call. Your spend is then broken down by customer, in dollars, down to the request. You can also track by feature, team, environment, and prompt version the same way.

**How do I catch an unexpected spike in my AI bill?**

Set spend alerts and get notified in Slack, Discord, or email. You can set a daily threshold, a spike detector that compares against yesterday, or a monthly budget cap. Alerts fire within minutes of a breach, not at the end of the month.

**Can I see how much each pull request cost in AI spend?**

Yes. Connect GitHub so pull requests can be matched, and have each engineer install the free Mac app. Cursor, Claude Code, and Codex all keep the per-session detail that ties spend to a pull request on the developer's machine, so the app is what syncs it. Coding ROI then prices each merged pull request and rolls the spend up by repository.

**Which providers are supported?**

OpenAI, Anthropic, Google Gemini, AWS Bedrock, Azure, Deepgram, ElevenLabs, Together AI, Fireworks AI, Modal, OpenRouter, Vercel AI Gateway, LiteLLM, and Cursor, with more added regularly. Per-request attribution through the SDK reaches 100+ more via gateways like OpenRouter and LiteLLM.

### Billing questions

**What counts as managed spend?**

The AI spend SuperPenguin sees for you in a calendar month, across every connected provider. It is what your providers bill you, not what SuperPenguin bills you. Your SuperPenguin subscription itself never counts toward the limit.

**What happens if I go over my plan's managed spend?**

Nothing breaks and nothing gets deleted. We tell you, and you can move up a plan when it suits you.

**Is the Free plan a trial?**

No. Free is a plan, not a countdown. It covers up to $2K in managed spend, AI coding analytics for 1 person, and attribution for 3 merged PRs total. The PR view has no monthly reset, and the Free plan does not expire.

**Do I need a credit card to start?**

No. You can connect a provider and see your own numbers without entering payment details.

**Is the Mac app included?**

The Mac app is free to download and shows you your own agent spend on its own. Free includes AI coding analytics for 1 person and PR attribution for 3 merged PRs total. Growth includes 3 Mac app users and the latest 20 merged PRs this month (UTC). Pro includes 5 Mac app users and unlimited PR attribution. GitHub is required for PR attribution.

**Can I change or cancel my plan?**

Yes, at any time, from Settings. Upgrades take effect immediately and downgrades take effect at the end of the current billing period.

## LLM cost management as a category

LLM cost management is the practice of measuring what AI costs an organization,
attributing it to the thing that caused it, and acting before the invoice
arrives. It is made of four jobs, in order: monitor it (every provider bill in
one running total, including spend nothing instrumented), attribute it (each
dollar labeled with the business object behind it), control it (an alert when
spend moves, priced at the rate the organization actually pays), and optimize
it (what to change next, written from the team's own traffic).

It covers two halves of AI spend, split by who the AI was working for, and most
tools see only one:

- **Serving your customers**: the model calls the product makes on a customer's
  behalf, through providers like OpenAI, Anthropic, Gemini, Bedrock, and
  Fireworks AI. Attributes to customer, feature, team, and environment.
- **Your team's own work**: AI a person at the company drives themselves, which
  SuperPenguin covers today across Cursor, Claude Code, and Codex. There is no
  end customer behind an agent session, so it attributes to team, project, and
  merged pull request instead. This half is not confined to engineering: anyone
  who ships work through those agents is measured the same way.

The accounting labels COGS and R&D map onto this split only roughly. Calls made
to serve a customer behave like cost of goods sold and set AI gross margin.
Spend a team drives itself is operating expense but not always R&D, because a
designer or a support lead doing day-to-day work is not doing research and
development. Traffic from dev and staging environments is neither.

It is a different category from LLM gateways, which sit in the request path and
can reject a call once a budget is crossed, and from LLM observability, which
traces and evaluates output quality and logs an estimated cost alongside latency.
Both see only the traffic routed through them, so neither reaches the second
half. Cost management reads billing and usage data instead of the request path,
which is what lets it reach both. Most teams run it alongside a gateway rather
than instead of one.

## Canonical URLs

- Homepage: https://superpenguin.ai
- Attribution: https://superpenguin.ai/attribution
- Coding ROI: https://superpenguin.ai/coding-roi
- One View: https://superpenguin.ai/one-view
- LLM cost management: https://superpenguin.ai/llm-cost-management
- Pricing: https://superpenguin.ai/pricing
- Documentation: https://superpenguin.ai/docs
- Use the SuperPenguin CLI: https://superpenguin.ai/docs/cli
- CLI installer: https://install.superpenguin.ai
- Connect an AI agent with MCP: https://superpenguin.ai/docs/mcp
- Agent API guide: https://superpenguin.ai/docs/agent-api
- MCP endpoint: https://app.superpenguin.ai/api/mcp
- Agent API base: https://app.superpenguin.ai/api/agent/v1
- Authenticated Agent API OpenAPI: https://app.superpenguin.ai/api/agent/v1/openapi