Frequently asked questions

Each person connects the Centrail CLI on the device where they use Claude Code, then syncs supported local logs. For a paid team, invite members and have them choose which repositories to share. A team's Overview counts eligible shared usage, not everything on each person's machine. Per-member breakdown visibility depends on the team's settings.

Start with your own free account, then compare team plans. Read how team numbers work for sharing rules and history boundaries.

Centrail captures supported local logs from Claude Code, Codex, and GitHub Copilot CLI. It records reported tokens and estimates cost using model pricing. Coverage depends on the logs available on each device; GitHub Copilot CLI support does not mean every Copilot surface is supported.

See surfaces and providers and how usage is parsed. This is a record of captured usage, not a complete provider account or invoice reconciliation.

Sync usage with repository context, then review tokens and estimated cost in your personal Overview or the team's eligible shared view. Repository selection determines what can sync, and team sharing determines what colleagues can see. Developer-level comparisons need the appropriate team breakdown permissions.

Use the numbers to investigate where spending is concentrated. They do not establish which developer is more productive or which model produces better work. Read how team numbers work.

Yes. Centrail reads supported local coding-agent logs through its CLI; you do not have to route model traffic through a gateway or instrument your application with an SDK. You still need the CLI, a Centrail account, and supported logs. Allowed usage and Git metadata are uploaded to that account.

See local token capture for what leaves the device and get started for setup.

Centrail can allocate captured usage to Git commits when matching evidence is available. Cost per Commit is an estimated allocation, not the complete cost of a feature. A feature can span multiple commits and people, and some usage remains unattributed. Attribution does not prove an agent caused a commit or that the work reached production.

Use the record as evidence for a cost investigation, alongside repository context and unattributed usage. Read Git attribution and cost, output, and value.

Captured usage and estimated cost can help you spot spending patterns and ask better budget questions. Centrail does not guarantee savings, identify the best model for a task, or provide a forecast through this workflow. Estimated usage cost is not necessarily your provider bill or subscription price.

The available record is a starting point for your decisions, not proof of business value. See cost, output, and value.

Prompts, completions, source code, and diffs are not uploaded as usage data. Usage and Git metadata are uploaded, and metadata such as repository names and commit identifiers can still be sensitive. Team members separately choose which repositories and earlier history to share.

Read the privacy policy before connecting a work device. The CLI's inspect --last command lets you review its last upload locally; do not share that payload publicly without checking its metadata.

No. Begin with device pairing and a manual sync. Optional hooks or the Claude Code plugin can capture session context and automate sync after an agent turn. Review the setup choices and your organization's policies before changing agent settings.

The optional GitHub App supplies independent push and pull-request evidence. It supplies no token usage and does not share members' usage with a team. See the guide's optional automation steps and how team numbers work.

Centrail is built for teams working in private repositories, with privacy and controlled sharing at the core. Prompts, completions, source code, and diffs are not uploaded as usage data. Usage and allowed Git metadata are uploaded to your account, and repository names, branch names, and commit identifiers can still be sensitive. CLI capture scope and team sharing are separate choices; connecting a device does not automatically disclose its usage to a team.

You control which repositories may sync and which usage your team can see. Use inspect --last to check the CLI's last upload locally, and read the privacy policy and local capture details for the data boundaries. For a team rollout or questions from your security team, contact support. We can help you review the setup without asking you to send private logs, credentials, or source code.

Yes. Email support@centrail.org with the problem you want to solve, the coding tools you use, and what a useful result would look like. For a team workflow, include the approximate team size and explain where the current setup falls short. Describe the need without sending confidential code or raw logs.

Requests help shape priorities; submitting one does not guarantee implementation or a delivery date. The same support contact is available for bugs and feedback.

The direction is a system of record for AI labor: connecting what agents consume to attributable work and, ultimately, completed outcomes. Coding agents and Git are the starting point because they leave useful records on both sides of that connection.

Today, Centrail records supported coding-agent usage, estimated cost, and associated Git evidence. The longer-term ambition is to help account for completed features, fixes, releases, and eventually work beyond software development. Those are directions to explore, not promises of available features, guaranteed savings, or delivery dates. Forecasting and business-outcome valuation are not advertised as available in this workflow.

If a particular outcome matters to you, share the use case with us. Start with the get-started guide for what the current product does.

Pairing authorizes a device; it does not create usage. Run sync, read its summary, and confirm that supported logs exist on the connected device. Check that you are viewing the paired account and an appropriate date range. In a team view, also check membership and repository sharing.

Missing usage is different from missing commit attribution: usage can exist without a matching commit. Follow the get-started troubleshooting steps or contact support with redacted error details, never credentials or private logs.

GreenAI starts with captured tokens, separates input, output, cache reads, and cache writes, and applies research-based energy rates and model factors. It then derives CO₂ and water estimates from that energy using provider/grid assumptions.

Electricity is shown in kWh, water in liters, and carbon in kg of CO₂; these are environmental footprint estimates, not electricity or water charges. Dollar spend is calculated separately from model pricing. The environmental methodology publishes the formulas, coefficients, assumptions, and sources.

They are not metered readings from the data center that served your requests. Agent logs do not identify all the hardware, location, efficiency, or grid conditions involved. GreenAI therefore reports a central estimate with a low-to-high range from the method's assumptions and sources. The range is not a guarantee that the actual footprint falls within it.

Read why the range is wide and the AI and energy background.

No. Centrail distinguishes input, output, cache reads, and cache writes where supported logs report them. Model pricing assigns these token classes different dollar rates; the footprint method also uses different energy assumptions for each class. Two sessions with the same total tokens can therefore have different estimated costs and footprints.

See how usage is parsed for token classes and the environmental methodology for the energy assumptions.

Captured token usage is priced using model rates, including separate cache rates where applicable. The result is estimated usage cost, not a reconciliation of your provider invoice, subscription, credits, or negotiated pricing. Environmental footprint and dollar spend describe different things.

Read cost, output, and value and the first-number definitions in the get-started guide.

The coding tool and model provider are different. GitHub Copilot CLI can use models from different makers; Centrail groups captured model usage by the company that makes the model. A model used through Copilot can therefore appear under OpenAI or Anthropic without being counted twice.

See surfaces and providers for the distinction and how usage is parsed for the captured fields.

Agent economics is accounting for what AI agents consume, what they cost, and the work associated with that usage. Centrail starts with coding-agent logs and Git: a ground-level record of captured usage, estimated cost, and work evidence.

That record supports questions about AI labor. It does not establish the complete economic value of an agent's output. Get started to see what the current product records.