See what your coding-agent spend buys.
Octic shows each engineer’s coding-agent sessions, the prompts behind them, and the spend split by kind of work. It runs on every host with an Octic sensor, with no per-app setup, for Claude Code and Codex today.
Coding-agent usage
Claude Code and Codex · Acme Corp · last 30 days, September 2026
$6,752.15 across 545 sessions from 4 engineers
Daily coding-agent spend over September 2026, the team with Priya Raman’s share lit. Priya Raman’s busiest day was September 14, at $293.42.
What did Priya Raman’s $3,012.40 buy? Kind of work Share Spend Feature work $1,684.20 Refactors $512.10 Tests $401.30 Docs $236.80 Debugging $178.00 Priciest session: Add idempotency keys to the refund endpoint and cover retries $41.80
$3,012.40 across 214 sessions: feature work took 56%.
Every session, with the prompt behind it
Agentic Coding Insights lists each engineer’s coding-agent sessions and rebuilds each one: the prompt, the reply, every tool call with its command, output, and wall time, and what the session cost, down to cache reads. A line on the bill turns into work you can recognize.
Any host running an Octic sensor reports its sessions on its own. Claude Code and Codex are supported today, and more coding agents are coming.
Priya Raman’s sessions
4 of 214 this month, costliest first
Add idempotency keys to the refund endpoint and cover retries
Priya Ramanpayments-apiopus-5Feature workautoHigh effortSep 16, 10:42 AM · 52m
- Turns
- 6
- Tool calls
- 31
- Tokens in
- 18.4k
- Tokens out
- 61.2k
- Cost
- $41.80
Prompt · turn 0
Add idempotency keys to the refund endpoint and cover retries
Assistant · turn 1
I’ll accept an Idempotency-Key header on POST /refunds, store each key with its response in refund_requests behind a unique index, and replay the stored response when a retry arrives.
First 5 of 31 tool calls
- Readsrc/
routes/ refunds.ts41ms - Grep"idempotency"
src/118ms - Editsrc/
routes/ refunds.ts16ms Output · first lines
PASS test/refunds.test.ts ✓ replays the stored response for a retried key (41 ms) ✓ rejects a reused key with a different body (12 ms) ✓ refunds once under concurrent retries (88 ms) Tests: 14 passed, 14 total
- Bashgit
diff --stat63ms
Tool latency
- Read · refunds.ts41ms
- Grep · idempotency118ms
- Edit · refunds.ts16ms
- Bash · npm test18.4s
- Bash · git diff63ms
Bars are on a log scale.
Cost breakdown
| Input tokens | 18,400 |
|---|---|
| Output tokens | 61,200 |
| Cache reads | 1,694,000 |
| Cache writes | 66,400 |
| Total opus-5 | $41.80 |
1,840,000 tokens, as the session reported them. The total is the billed figure.
Spend split by kind of work, engineer by engineer
Octic splits spend by model, by repository, and by kind of work, such as feature work, refactors, tests, docs, and debugging, so “what did $3,000 of tokens buy?” has a real answer for every engineer on the team.
Where the money goes
$6,752.15 over September 2026
- Feature work$2,997.80 · 203 sessions
- Refactors$1,282.95 · 105 sessions
- Debugging$952.00 · 90 sessions
- Tests$913.50 · 88 sessions
- Docs$605.90 · 59 sessions
Feature work takes 44% of spend across 203 sessions.
Work mix by engineer
- Priya Raman$3,012.40By kind of work: Feature work $1,684.20, Refactors $512.10, Tests $401.30, Docs $236.80, Debugging $178.00.
- Marcus Webb$1,846.75By kind of work: Feature work $612.40, Refactors $368.15, Tests $214.60, Docs $96.30, Debugging $555.30.
- Lena Okafor$1,204.10By kind of work: Feature work $402.50, Refactors $318.40, Tests $176.20, Docs $214.60, Debugging $92.40.
- Sam Ito$688.90By kind of work: Feature work $298.70, Refactors $84.30, Tests $121.40, Docs $58.20, Debugging $126.30.
Sessions on personal accounts get flagged
Octic checks the identities reporting sessions against the engineers you enrolled. When a session runs on a personal, non-corporate Claude or Codex account, Octic flags it with the host it ran on, so you see work that happens outside the accounts your company manages.
Identities reporting sessions
4 engineers enrolled, 5 identities reporting: one is a personal account.
- Priya RamanClaude Code, Codexpriya.raman@acme.comCorporate seat
- Marcus WebbClaude Codemarcus.webb@acme.comCorporate seat
- Marcus WebbCodexPersonal accountFlagged
Why is the staging plan trying to recreate the NAT gateway?
Codexinfra-terraformDebuggingWIN-ENG-07, a Windows host$9.35
Sent to Remediate as a finding
- Lena OkaforClaude Codelena.okafor@acme.comCorporate seat
- Sam ItoClaude Code, Codexsam.ito@acme.comCorporate seat
SDK telemetry for the AI you build
Coming soonWhen the Octic SDK ships, you’ll add it to your applications and they’ll report the AI activity inside them: each AI agent run, with its model calls and tool calls.
Runs will land on the same map as everything Octic discovers, next to each AI agent’s owner and trust state. Today, the kernel-level sensors that find AI on a host already watch its processes, so observability doesn’t wait for the SDK.
- support-triage-agent, an AI agent in Acme's own application, runs on ip-172-31-67-241. Owner @support-eng. Approved.
- Its run run_4d1e, as the Octic SDK will report it, makes 3 model calls through the Anthropic SDK 1.3.0 package and 2 tool calls to the mcp-jira MCP server.
- The run used 18,420 tokens over 6.8 seconds and cost $0.11.
Observe shows what AI does once it’s on the map
Discover finds AI agents and coding agents, and Trust decides whether each is approved. Observe shows what they do and what they cost, and hands anything wrong, like a Codex session on a personal account, to Remediate as a finding with an owner.
Put your own estate on the map
Install a sensor on a Linux or Windows host and coding-agent sessions start showing up, with no extra setup. Connect your APIs and the rest of the map fills in.