AI coding subscriptions make experimentation easy, but they also make spending harder to understand. A developer may use Codex in one repository, Claude Code in another, and GitHub Copilot throughout the day without having a single view of the resulting usage. RepoSpend addresses that gap with a small, local-first dashboard that turns existing AI coding session data into a project-level cost map.
The tool is free and open source, and it runs on the developer’s machine rather than as a hosted analytics service. It does not require an account, does not send telemetry, and does not upload source code or prompts. Instead, RepoSpend reads session files already stored locally and presents the information in a browser at localhost:2005. That design is especially practical for developers who need visibility into usage but cannot justify sending project metadata to yet another cloud dashboard.
A project view of AI coding usage
RepoSpend’s main organizing idea is the Git root. Sessions found inside a repository, package, script directory, or other subfolder can be grouped under the project name developers already recognize. This is more useful than a raw stream of token counts: someone comparing a client project with a personal tool can quickly see which codebase is consuming more assistance.
The dashboard combines several views rather than reducing everything to one headline number. Users can filter by source and inspect totals, top repositories, individual sessions, models, and tools. It also surfaces cache reuse, command problem rates, and file-edit activity. Those signals do not explain every reason a session became expensive, but they can point to areas worth investigating. Repeated prompts, failed commands, or low cache reuse may indicate that a workflow needs tightening.
- Repository grouping connects local sessions to Git roots, making project-to-project comparisons easier.
- Session, model, and tool breakdowns help identify where the largest share of token usage came from.
- Cache reuse and command issue indicators provide clues about repeated or unsuccessful AI interactions.
- File-edit information adds a little operational context to otherwise abstract token totals.
For example, an independent developer juggling several small applications could use RepoSpend at the end of a week to find out whether a new feature branch, a debugging session, or a particular model is driving usage. It will not tell that developer whether the generated code is good, but it can make the resource trade-off visible before the next billing cycle.
The numbers are estimates, not invoices
This distinction matters more than the dashboard’s visual polish. RepoSpend reports API-equivalent cost estimates, not the amount charged by a subscription provider. It converts observed token usage using public model API rates, creating a common yardstick for comparing tools and models. A user should not treat the result as a replacement for a credit-card statement, a provider usage page, or the terms of an individual plan.
That approach is still valuable. Subscription pricing and usage rules can obscure the marginal cost of different workflows, while an API-equivalent figure gives developers a consistent reference point. It can show, for instance, that one model is responsible for most of a repository’s estimated consumption even when several assistants are active. The estimate may differ from actual spending because subscriptions, included quotas, discounts, rate changes, and product-specific accounting are not necessarily represented.
The project is also trying to keep its pricing assumptions current. Version 0.1.3 lowered the API-equivalent rates for GPT-5.6 Terra and Luna, including cached input and cache-write pricing, while retaining support for user-defined price overrides. It also improved folder grouping for repositories that have not been verified. These changes suggest a model that can be adjusted as pricing changes, rather than a permanently fixed calculator.
Installation is simple, but the project is young
Running RepoSpend requires Node 20+. Once that runtime is available, the basic launch path is a single npx repospend command, followed by opening localhost:2005 in a browser. There is no separate database or hosted account to configure. The project also provides a guided demo using fictional, Lord of the Rings-style data, which lets people explore the interface without exposing real repositories.
That low setup cost is one of RepoSpend’s strongest arguments. A developer who is comfortable with a terminal can install it, inspect the dashboard, and decide within minutes whether the collected data is useful. Someone who avoids command-line tools may find the Node requirement less welcoming, although the absence of server setup keeps the overall process relatively light.
RepoSpend remains an early 0.x project. Version 0.1.3 is functional enough for experimentation, but its community and ecosystem are still small. Support for Codex and Claude Code is described as the most accurate, while GitHub Copilot has initial support and Cursor remains experimental and must be enabled explicitly. Import behavior and cost precision should therefore be checked against the source tools before the numbers are used for serious budgeting.
- Use the dashboard for trends and comparisons, not as an authoritative billing record.
- Check which local session formats are available before assuming every tool has equal coverage.
- Consider setting custom prices when the default API assumptions do not match a team’s internal accounting.
For privacy-conscious developers who use multiple AI assistants, RepoSpend is a pragmatic utility: installable, inspectable, and focused on one narrow problem. Its biggest limitation is also clear. It offers a local view of estimated consumption, not a universal ledger of what every provider will charge.
Comments
No comments yet
Be the first to comment