switch-trust-cc-plugin — Claude Code usage & cost reporter

See what Claude Code is actually costing you.

A Claude Code plugin that reports your usage — tokens, model, and computed cost — to the Switch Trust platform, so your Switch Trust cost/ROI dashboards reflect real Claude Code consumption, per user and per repo, over time. Once installed, it runs quietly in the background; no dependencies to install, nothing to run day-to-day.

Install

/plugin marketplace add https://github.com/sandbox-quantum/switch-trust-cc-plugin
/plugin install switch-trust-cc-plugin@switch-trust-cc-plugins

Claude Code will prompt you for the settings below — enter at least your Switch Trust API Key. You can revisit these any time via /config; your key is stored in your OS's secure credential store, never in plaintext. These settings come from /config only — a repository you open can only lower the Report Mode or make PII Redaction stricter (never the reverse, or change your key, endpoint or identity).

Setting	Required	Description
Switch Trust API Key	yes	Your Switch Trust API key (sk_...)
Switch Trust Ingestion Endpoint	no	URL of the ingestion service (defaults to production instance). Must be https:// on an allowed Switch Trust host, or http://localhost / http://127.0.0.1 for local development — anything else is refused and logged, never sent.
User Email Override	no	Override the per-user identity used for attribution (else auto-detected locally: git config user.email → OS login name; no cloud credentials are used — see Data & Privacy)
Report Mode	no	condensed, text, full (default), or none (disables reporting) — see Report modes
PII Redaction	no	default (default), strict, or off — best-effort replacement of card numbers, US SSNs, IBANs and labelled IDs; see PII redaction

• Hooks and the /switch-trust-status command activate as soon as the install finishes. If the install summary asks you to, run /reload-plugins (a full restart is not required).
• No dependency install step needed — the plugin is pure Python 3 standard library.

To opt out entirely without uninstalling, set Report Mode to none via /config.

Keeping it updated

Claude Code pins the version you installed — you won't get security or bug fixes until you update. Third-party marketplaces have auto-update off by default, so turn it on:

/plugin

Open the Marketplaces tab, select switch-trust-cc-plugins, and choose Enable auto-update. From the shell instead:

claude plugin marketplace update switch-trust-cc-plugins --auto-update

Org admins: force it for everyone via managed settings:

{
  "extraKnownMarketplaces": {
    "switch-trust-cc-plugins": {
      "source": {
        "source": "github",
        "repo": "sandbox-quantum/switch-trust-cc-plugin"
      },
      "autoUpdate": true
    }
  }
}

Once set there, users can't turn auto-update off with /plugin.

Support

Questions or issues: open a GitHub issue or email support@switchagents.ai.

Report modes

The Report Mode setting controls how much conversation detail is sent.

Mode	What's collected
full (default)	Real prompts and responses, plus tool calls and tool results (commands run, file contents, tool output).
text	Real prompts and responses. No tool calls or results.
condensed	Token counts and a one-line AI-generated session summary only. No real conversation content.
none	Nothing — reporting is fully disabled.

See Data & Privacy below for exactly what each mode sends and why full is the default.

PII redaction

Before anything is sent, the plugin also replaces structured personal data on your machine, in every mode (including the condensed title, repo and branch string). This is pattern-based and best-effort: it reduces exposure, it does not guarantee removal. The PII Redaction setting (switch_trust_pii_redaction, set via /config) chooses what is replaced:

Setting	What's replaced
default	Payment card numbers with a known issuer prefix (Visa, Mastercard, Amex, Discover) and a valid Luhn check digit — in prose, CSV, SQL and JSON, including JSON numbers. US Social Security numbers in 123-45-6789 form that pass the SSA validity rules, or as 9 digits right after a label such as SSN, social security number, taxpayer_id or TIN, or under such a JSON key. IBANs with a valid checksum. Passport, driver's licence and medical-record numbers, and dates of birth, only right after a label such as passport no., driver's license, MRN, DOB or date of birth, or under a matching JSON key in tool input.
strict	Everything in default, plus phone numbers in +<country code><number> form and phone numbers right after a label such as phone, mobile or fax.
off	Nothing. Secret redaction still applies.

Each value is replaced with a numbered placeholder such as [card-1] or [ssn-2]. Within one report (everything one hook run sends) the same value gets the same placeholder and different values get different ones, so repeated tool calls can still be told apart. Placeholders are plain sequence numbers, not hashes, so they cannot be reversed; numbering restarts with each report.

It does not detect: names, email addresses, postal addresses, phone numbers (outside strict), national ID numbers from countries other than the US, ID numbers and dates of birth with no label next to them, or values split up or obfuscated. Free-text special-category data — health conditions, religion, political opinions, ethnicity, sexual orientation — is not detected at all; patterns cannot recognize it reliably. It also replaces some look-alikes: well-known test card numbers, about 2–3% of random 16-digit order numbers, most values shaped like 123-45-6789 (such as part numbers), unless a label such as order id or part sits right before them, and card-like digit runs inside mostly-numeric hex IDs (an instance ID such as i-4743355369860f07b can become i-[card-1]f07b).

A repository you open can make this setting stricter, but never lower it or turn it off.

Data & Privacy

This is usage reporting, not anonymous telemetry — it is attributed to you and your repository, and it runs on every Stop and SessionEnd. In every mode except none, each report sent to the ingestion endpoint (default https://ingestion.switchagents.ai, always a switchagents.ai subdomain) includes your identity (email), your repository's name and a one-way hash of its git origin URL (used to group usage by repo without storing the URL itself), the Claude Code session ID, timestamps, the model name, and token counts.

condensed additionally includes the current git branch and an AI-generated one-line session title, combined into a single summary string — no real conversation text.

text and full additionally include the real prompt and response text for each turn. full also includes tool calls and tool results — commands run, file contents, and other tool output — as they appear in the transcript, including any third-party data that happens to be in them. Redaction of recognizable secrets and of structured personal data (see PII redaction) is best-effort — it reduces exposure, it does not guarantee removal.

full is the default mode, and it sends real conversation content — prompts, responses, and tool inputs/outputs such as shell commands and file contents. Switch Trust's ROI recommendations are built from that content, which is why full is the default. If your organization's data-handling policies require sending less, switch the Report Mode setting via /config — see Report modes above for what each mode sends.

Who can see it: reports are visible to every member of your Switch Trust organization, whatever their role. Your prompts, responses and (in full mode) tool output can be read by your colleagues there.

To turn this off: set Report Mode to none via /config, or uninstall the plugin. There is no per-session consent prompt — the setting applies to every session until you change it.

Exactly what each field contains, by mode:

Field	full (default)	text	condensed	none
session_id	Claude Code session id	same	same	not sent
agent	Claude Code — <your email> (per-user attribution)	same	same	not sent
model	message.model from the transcript (e.g. claude-opus-4-8)	same	same	not sent
input_tokens	effective tokens for that turn	same as full	effective tokens for the whole session, folded to reflect real billing (cache reads/writes priced in)	not sent
output_tokens	output tokens for that turn	same as full	summed output tokens for the whole session	not sent
prompt	the real user message text for that turn, plus any tool results fed back to the model (command output, file contents)	the real user message text for that turn	<ai-title> (<repo>:<gitBranch>) — no real conversation text	not sent
response	the real assistant message text for that turn, plus any tool calls the assistant made (commands run, files edited)	the real assistant message text for that turn	never sent (left empty)	not sent
app_info	repository name, and a one-way SHA-256 hash of its git origin URL (sent as slotid)	same	same	not sent
timestamp	time of that turn	same as full	last message time in the reported window	not sent

Identity (agent's <your email>): resolved entirely locally, in order — the User Email Override /config setting, then git config user.email, then your OS login name, then left anonymous if none resolve. No cloud provider or identity service (gcloud, AWS, etc.) is ever contacted to look this up.

Prohibited Data: our terms of service prohibit sending protected health information, payment-card data, government ID numbers, or other special-category personal data through this plugin. In full mode (the default), the plugin forwards whatever text appears in the files and command output it is shown, filtered only by best-effort, pattern-based redaction of secrets and of the structured personal data listed in PII redaction. Free-text health or other special-category data is not detected. If your repository or shell output may contain any of this data, use condensed or text mode, or none.

More: Privacy Policy, Terms of Service, Data Handling, Data Protection Addendum.

Observability / debugging

The plugin keeps all state and a log under Claude Code's own per-plugin data directory, ${CLAUDE_PLUGIN_DATA} (e.g. ~/.claude/plugins/data/<plugin-id>/ — set for every hook invocation, survives plugin updates, cleaned up on uninstall). Outside a hook invocation it falls back to ~/.switch-trust/roi_cc_plugin/:

<state dir>/
├── ccusage.log            # what ran, tokens reported, HTTP result, errors
├── <session_id>.cursor    # last processed transcript message per session
├── <session_id>.opened    # SessionOpen-sent marker
├── <session_id>.last_emit # throttle timestamp
├── <session_id>.seq       # next sequence_number (report modes text/full only)
└── user_email             # cached identity

Quick checks:

• Status: run /switch-trust-status in Claude Code — shows where state lives, whether config is set, active sessions, and the tail of the log.
• Check the files manually: inspect <state dir>/ccusage.log and the other files listed above directly.

If the Switch Trust API Key setting isn't set, the log will say no switch_trust_api_key configured; skipping — that's the first thing to check.

License

Source available, not open source: Apache 2.0 as limited by the Commons Clause license condition, which excludes selling the software or a derivative of it. See LICENSE for the full terms.