Surfaces¶

A tour of the Ogma dashboard. Each surface reads from the same hash-chained, blockchain-anchored record — the difference is the lens it puts on your AI usage.

Ogma audits AI usage from outside the vendor's perimeter. It pulls from each vendor's management APIs with an admin key — no agents, no proxies, no sidecars — across Claude (Anthropic) and OpenAI today. (If you need to govern autonomous agents inline rather than audit usage after the fact, that's the separate Tyr product — see the Tyr quick start.)

New here? Connect a vendor first: Connect Anthropic, Connect OpenAI, Connect Claude (MCP).

API Usage¶

The cross-vendor view of API key consumption — tokens and estimated cost, built on each vendor's daily usage report.

• Vendor filter: All, Anthropic, or OpenAI. All gives you one cost view across both vendors; switch to a single vendor to drill in.
• Per model and per project/workspace. Rows break down by model and by the project (OpenAI) or workspace (Anthropic) the spend landed in.
• Cost estimate. Spend is computed against each vendor's published rates, using the rate that was in effect when the usage happened. Cached input is priced at the reduced cache-read rate. Where a model can't be priced (legacy aggregate rows with no model), the cost column shows — rather than a guessed number.
• Cache efficiency panel. Above the table, a per-model verdict on prompt caching: an overall cache hit rate plus the models where caching is being paid for but under-reused, each with a plain-English recommendation (for example, "Low reuse — stabilize the cached prefix or raise the cache TTL"). When every model looks healthy it collapses to a one-line note.

OpenAI note: OpenAI exposes an authoritative billed-cost report (used for project totals) in addition to token usage (used for per-user attribution), so OpenAI project spend reflects actual billed amounts, including non-token line items like fine-tune training.

Sessions¶

Per-developer, per-day activity — who did work with AI, and on which day.

A session is one (day, developer) roll-up. It's assembled from:

• Claude Code activity (Code Analytics), and
• Claude (chat) activity captured through the MCP connector.

A single session can draw from more than one source, so each row carries a badge per contributing source (for example, "12 events — 8 chat, 3 API key, 1 admin event") rather than collapsing everything under one label. Open a session to see its events broken down by source on a timeline.

Users¶

One person, every surface, both vendors — in a single row.

A single human typically shows up as several disconnected identities across a vendor's own console: an API caller, a Claude Code user, a Claude Desktop user, an OpenAI user. Ogma stitches them into one person by email, so a user's row spans Claude and OpenAI activity together.

Each user surfaces:

• Spend across vendors — their share of API cost, Claude and OpenAI combined.
• An activity heatmap — a GitHub-style grid of activity density per day, tone-shifted by source, with a per-day breakdown on hover.
• Top topics — for Claude chat activity captured via MCP, a breakdown of what kinds of work a person spends their Claude time on. Topics are inferred from Claude's own short, self-reported interaction summaries (never raw transcripts) and constrained to a fixed category set; the panel shows how many interactions were classified so partial coverage is always honest.

Unmapped activity¶

When an incoming record's identity can't be auto-matched to a known user by email, it lands in an Unmapped activity panel for an admin to link by hand. Once you link an identity manually, that link is protected from being overwritten by later automatic matching.

Insights¶

A page of analysis cards built on top of your usage data — the layer that turns raw events into something to act on. Cards appear in a fixed order; a card with nothing to flag still shows present-tense copy describing what it watches for, so the page never empties out.

Current cards include:

• Cost forecasting — projects month-end spend from your recent trend, per vendor, and flags when you're on pace to cross a budget cap (for example, "on pace to exceed the cap by the 24th"). The projection is a deliberately simple trend line; the value is the early warning, not decimal precision, and the card says so.
• Model mix — how your spend is distributed across models, including cross-vendor shift (e.g. spend moving between Claude and OpenAI over time).
• Project / workspace cost attribution — where spend is concentrated by project (OpenAI) or workspace (Anthropic).
• Cache-efficiency recommendations — the same prompt-caching analysis surfaced on API Usage, distilled into actionable callouts.

Content¶

When content capture is enabled, Ogma stores Claude chat message text in encrypted, hash-chained, per-message records. The Content surface lets you:

• View chats — the message transcript for a chat. PII is masked by default (emails, phone numbers, SSNs, card numbers, API keys, IPs are replaced with [redacted:<type>]). Any tenant member can view masked content; the per-message redaction summary (counts and types, never the values) is shown to everyone.
• Reveal PII (admins, logged) — an admin can un-mask a chat. Every reveal appends a tamper-evident content_reveal record to the chain noting who revealed it and when. Unmasking is treated as a privileged, audited action.

• eDiscovery export — download a tenant's captured content as a verifiable bundle. Choose the format:

  Format	Output
  zip (default)	A machine-readable JSON bundle — chats, a per-record chain proof, and a manifest with the chain-verification result.
  pdf	A courtroom-readable, Bates-numbered transcript (VARGATE-000001…) with an integrity attestation cover page and a chain-proof appendix.
  both	The JSON bundle and the PDF in one ZIP.

  The bundle's differentiator is its independent verifiability: an auditor can confirm the exported text is exactly what was recorded (SHA-256 of the plaintext matches the chained content_hash) and that the underlying chain is intact — without trusting Ogma. Exports redact PII by default; a full-content export is an admin action and is itself logged.

• Delete / crypto-shred — remove captured content on request (right-to-erasure, retention, offboarding) without breaking tamper-evidence:

  Scope	What it removes
  Per-chat	One chat's message content.
  Per-user (DSR)	All of one data subject's content, across every chat.
  Per-tenant (offboarding)	Everything — all content and the sealed keys — by destroying the tenant encryption key (crypto-shred). Terminal and irreversible.

  Deletion is chain-safe: Ogma never modifies or removes a chain record. It makes the content unreadable and appends a new, append-only deletion event — so the record that content existed, and that it was deleted both survive, and chain verification stays green. A deleted chat re-renders as a tombstone (a "content purged" banner with the date and reason; each message shows [content deleted]). All deletes are admin-only and require a reason that's recorded in the audit trail.

Budgets & alerts¶

Spend budgets with threshold and forecast alerting.

• Thresholds. A budget fires as current-period spend crosses 70%, 85%, and 100% of its cap. Each threshold fires at most once per period, so crossing all three sends three distinct alerts — not one cumulative one.
• Forecast alerts. A budget can also alert when spend is projected to cross a threshold by period-end on the current trend — the early-warning companion to the after-the-fact threshold alert. Both ride the same alert pipeline.

• Channels. Each budget configures who's notified per channel:

  Channel	What's sent	Where the credential comes from
  Email (default)	A branded budget-alert email.	Just an address.
  Slack	A message to an incoming webhook.	Slack → Apps → Incoming Webhooks → copy the https://hooks.slack.com/services/… URL.
  PagerDuty	An Events API v2 event (critical at 100%, warning otherwise).	A PagerDuty service → Integrations → Events API v2 → copy the routing key.

• Test before you rely on it. Each budget has a Send test alert button (admins) that fires a clearly marked [TEST] through every configured channel, so you can confirm a webhook delivers before a real threshold is ever crossed.

Related¶

• Connect Anthropic (Admin key) — usage, members, workspaces.
• Connect OpenAI (Admin key) — cross-vendor cost and per-user attribution.
• Connect Claude (MCP) — per-turn chat activity for Sessions and Users.
• Compliance Access Key — unlock the Content surface.