TokenSpend

TokenSpend is a spend attribution platform for AI coding tools. It connects your AI usage, starting with Claude Code, to your GitHub pull requests, then tracks which tokens resulted in merged code. The output is a breakdown showing how much of your AI budget actually produced shipped features versus how much went to work that never landed. For engineering teams trying to justify AI tool costs or optimize spending, this answers the question that finance keeps asking, namely which AI investment delivered real value. The platform treats token consumption as an investment to be measured against outcomes rather than a cost to be minimized blindly.

The core insight is that raw token consumption tells you almost nothing useful. Knowing your team burned a million tokens last month does not say whether those tokens produced working code or were spent on experiments that went nowhere. TokenSpend categorizes spend into distinct buckets. Shipped spend is tied to merged pull requests and represents tokens that directly contributed to production code. In flight covers open PRs and recent work that has not landed yet, showing active investment still in progress. Unmatched is spend with no merged PR after 30 days, indicating work that may have been exploratory or abandoned. There is also an Other category for spend that falls outside these patterns. The breakdown reveals which portion of your AI budget actually turned into production code, and the platform tracks yield percentage alongside total spend so you can see efficiency trends over time. This categorization makes it possible to have informed conversations about AI tool value rather than guessing.

Attribution works by linking captured sessions to pull requests via branch and SHA. The system connects GitHub through a GitHub App that monitors merged code, then reconciles that against the AI spend data captured from Claude Code and other supported tools. Confidence scoring is visible on every attribution, so you know how solid the match is. Some attributions will be high confidence when the branch and commit references align cleanly, while others may be lower confidence when the mapping is less direct. The platform tracks at the team level rather than creating individual leaderboards, which keeps the focus on aggregate ROI rather than turning AI usage into a performance metric for individual engineers. This aggregate by default approach is intentional, avoiding the dysfunction that comes from making AI assistance feel like a surveillance tool. Engineers can see their own usage, but team metrics do not become a scoreboard.

Beyond spend tracking, TokenSpend includes a Token Router that provides a unified endpoint supporting Claude, GPT, Gemini, and open source models. The router currently tracks 52 models from 14 vendors in its live registry, including options like Claude Sonnet 5, GPT 5.6 Luna, and Gemini 3.5 Flash, each with their own pricing and context window specifications. Routing is completely free, and you do not need a provider API key to try it. Usage plus cost remain visible by engineer even while routing through the unified endpoint, so you can compare how different models perform for your team. The infrastructure has already routed over a million tokens, so the system is proven at moderate scale. For teams using multiple AI providers, the router consolidates access through a single integration point.

Privacy is baked in deeply. Only token counts, timestamps, and git references leave your machine. Prompts and code never do. This means your actual codebase and the queries you send to AI tools stay entirely local. Everything is encrypted in transit and at rest. There is a zero data retention option available with one click if you want an even stricter posture, which is useful for teams with specific compliance requirements. SSO is available for enterprise customers, and SOC 2 compliance is in progress. This architecture matters for teams in regulated industries or anyone who does not want their codebase touching another third party beyond what is strictly necessary for attribution.

The freemium model is built around pilot adoption. You can start with one repository and one team without paying anything. The routing is completely free, and the spend attribution works within those constraints. This means you can evaluate whether the tool delivers value before committing budget. Enterprise features unlock for teams that need broader coverage across multiple repositories and teams. Daily spend visualization broken out by model version, 30 day and 90 day period analysis, and custom date ranges are available as teams scale up. The interface provides charts showing spend over time so you can spot trends and anomalies. For teams evaluating whether AI coding tools are worth the investment, TokenSpend provides the data to make that decision rather than guessing. It also gives you the numbers you need when leadership asks for ROI on your Claude or OpenAI bills.

TokenSpend targets engineering teams and FinOps functions who need visibility into AI tool spending. If you are paying hundreds or thousands of dollars a month for coding assistants and cannot explain where that money went, this is the tool that closes that gap. The focus is narrow, specifically GitHub based workflows and AI coding tools, but the problem it solves is specific and increasingly common as AI coding tools become line items in engineering budgets. For organizations where AI spend is growing and accountability matters, this platform turns vague cost centers into measurable investments.