Skip to content

Data collected ​

This page states what Phillip sends from your machine, where it sends it, and which parts are dropped before storage. After reading it, you know which data leaves your machine, which data actually stays in Ryoma, and where limits remain.

Summary

Ryoma stores only counts and metadata. Phillip doesn't turn on the variables that record prompt content, responses, or tool results. Ryoma's Collector uses an allowlist: only the fields needed for measurement are written to disk, and emails, account IDs, log bodies, and any unknown field are dropped. One point needs a careful read in Tool parameters: tool parameters can travel over the wire to the Collector, and are dropped there.

When data is sent ​

Data is sent only when you run Claude Code through ph and are signed in to Tara. Then Phillip sets the OpenTelemetry environment variables for the session, and data goes to Ryoma by OTLP over HTTP, authenticated with your Tara token. If you aren't signed in to Tara, nothing is sent. Check the connection with ph forge doctor, on the Ryoma and Collector lines.

Exporter settings in the repo's .claude/settings.json have no effect, so a repo can't turn telemetry off. Conversely, if you run claude directly without ph, Phillip doesn't turn on telemetry for that session. Administrators can enforce it with managed settings.

The telemetry data flow. Claude Code sends OTLP through Phillip to the Collector. The Collector attaches an identity from the token, removes sensitive fields such as email, then keeps only the fields on the allowlist. The filtered data is written to ClickHouse, which internal Grafana and the API then read. Source code, prompt content, replies and tool output never leave your machine.Claude Codevia Phillip,OTLP/HTTPCollectorAttach identityfrom the verified tokenRemove sensitive fieldsemail, account ID, bodyAllowlistonly fields Ryoma readsClickHousecounters andmetadataGrafanainternal,row isolationAPI Ryomareads metricsby permissionNever leaves your machineYour source codePrompt content and repliesTool outputDropped by the CollectorEmail and account IDTool parameters and log contentAny field not on the allowlist

Swipe sideways to see the whole diagram.

The Collector runs these steps in order and rejects the whole batch if one step fails, so unfiltered data never reaches ClickHouse.

What does your machine send? ​

Claude Code emits two groups of data: metrics and events (logs). The tables below list what your machine sends to the Collector. What Ryoma stores is much narrower, and is covered separately in a later section.

Metrics ​

MetricContent
TokensToken counts by type (input, output, cache read, cache creation), model, query source, and the skill, plugin, and agent names if any
CostEstimated converted cost at API prices
SessionsNumber of sessions, active time
Raw outputLines of code changed, number of commits, number of PRs
Edit acceptanceHow many times you accepted or rejected a change

Events ​

EventContent
Request to the modelTokens by type, cost, model, processing time
Skill, plugin, subagentThe name of the resource that was loaded or used, and who triggered it (you typed it, or Claude called it on its own)
Tool, hook, and safety decisionsTool name, allow or deny result, number of hook runs
Guardrails decisionsEach block, ask, or mask: rule ID, decision, severity, policy name, repo, task ID, tool name. No content

Identity and context ​

Each record carries a few more attributes to tie it to a person, project, and task. These are the session ID that Claude Code provides, the repo address (Phillip turns on the variable that sends the repo so Ryoma can map a repo to a project and organization), and the task ID forge.issue, taken from the branch name when Phillip launches Claude Code. The sender's identity (forge.user_id, forge.org_id) is attached by the server from your Tara token, and Ryoma doesn't trust any field that the client declares by itself.

When you sign in to Claude with OAuth, Claude Code may attach user.email and the account ID to records. The Collector drops these fields, along with every field outside the allowlist, before writing, so they aren't in Ryoma's raw data. Metrics are computed from the identity in the token, not from email.

What Phillip doesn't turn on ​

Phillip doesn't turn on the four variables that let Claude Code send content. These are the content of your prompts, the content of the model's responses, the content of tool inputs and outputs, and the raw API request and response bodies. So source code, file contents, prompts, and responses aren't in Phillip's telemetry.

Tool parameters ​

To know which Tara skills, plugins, and agents are being used, Phillip turns on tool detail mode. This mode makes Claude Code attach tool names and MCP names, and it can also attach shell commands, tool parameters, and error content, a few KB each time. Those fields may contain secrets or customer data.

Filtering happens on the server side, so these fields can leave your machine and reach Ryoma's Collector. There the Collector applies the allowlist. Only the fields Ryoma needs are written to disk: session ID, model, token counts, cost, skill, plugin, and agent names, repo, task ID, Guardrails decisions (rule ID, decision, severity), and the identity the server attached. Everything else is dropped, including tool parameters, commands, errors, user.email, account IDs, and even the record body. Tool and MCP names aren't on the list, so they aren't stored either.

To be precise, the content of tool parameters isn't stored, but it can travel over the wire to the Collector. Because filtering uses an allowlist, a new field that a later Claude Code version adds is dropped by default, with no need to wait for anyone to update the list. Besides the allowlist, the Collector has a separate layer that removes known content keys, in case the allowlist is edited by mistake.

Because tool parameters can still travel over the wire, run Ryoma over HTTPS, and let Guardrails mask secrets before they appear in commands or tool results.

Where is data stored, and for how long? ​

Raw data lives in Ryoma's analytics store (ClickHouse), and administrators set the retention period with RYOMA_RAW_TTL. Aggregated metrics such as Flow, AI, and project health live in Ryoma's database. Ryoma also syncs tasks from GitHub or GitLab with a read-only service token, including titles, labels, owners, and timestamps. Source code content isn't synced to compute metrics.

What else does Tara store? ​

Tara counts downloads per version and writes an audit log for administrative changes. The Knowledge base holds only documents that an organization or project deliberately uploads, and is read according to the space's permissions. Notifications through Telegram, email, Slack, and Zalo contain only metadata such as the task ID, title, and link.

View your data ​

You can see your own AI usage by project and by day with ph ryoma, and you can check the telemetry sending status with ph forge doctor. To reconcile Ryoma's metrics with Claude Code's local log on your machine, run ph ryoma verify.

Next steps ​