Cost

Path: /admin/cost

Where you go to answer "how much are we spending on Claude, and which agent is burning it?" One page, plain numbers, no spreadsheets. The standalone agentx usage serve process was retired and folded into this page — same data, now living next to everything else in /admin.

The point of this page is not to make you a cost analyst. It's to put the three or four numbers you actually need on one screen, so a non-technical operator can answer "is spend healthy?" in under a minute.

What you'll see

• Today / Last 7 days / Last 30 days: three big number tiles at the top. Each tile shows total spend in USD plus the task count it covers, so you can tell whether a high number reflects more work or more expensive work.
• Tier-2 surcharge KPI: a single number — how many extra dollars you paid this month because sessions crossed the tier-2 token threshold. Anthropic charges 1.5x once a session goes long; this KPI is the chunk of your bill that came from not rotating in time. If it's creeping up week over week, agents are running long without rotating and you're paying the surcharge unnecessarily.
• 30-day trend chart: one bar per day, height = spend. Hover any bar for the exact figure. The fastest way to spot a spike before it becomes a habit — three days of growing bars is the cue to investigate.
• Per-agent table: one row per agent. Columns: agent id, tasks run, total spend, and the slice of that spend that came from the tier-2 surcharge. Sortable on every column — click "Spend" to find the biggest spender; click "Tier-2" to find the worst rotators.

The numbers come from /api/admin/observability/cost, which reads the task_history table in .agentx/db.sqlite. No live polling of Anthropic — costs are computed locally from token counts using a built-in rate card. That means the page works offline and there's no API quota to worry about.

What you can do

• Sort the per-agent table by any column to find your top spender or your worst tier-2 offender.
• Click an agent's row to open that agent's page (/admin/agents/<id>) and dig into individual tasks. From there you can see exactly which task drove the cost.
• Hover the trend chart to compare any two days at a glance.
• Read the tier-2 KPI as a leading indicator: if it's growing, lower session.tierTwoThresholdTokens so sessions rotate sooner. The cost reduction is usually immediate.
• Switch the per-agent table between Today / 7d / 30d to localise a problem to a specific window.
• Bookmark a deep link for any view — the time range and sort column are encoded in the URL, so you can save "/admin/cost?range=7d&sort=tier2" as your morning check.

The page is read-only. There's nothing here you can break by clicking around. To actually change cost-related settings — model, rotation thresholds, agent budgets — go to Admin → Agents or Admin → Advanced.

Anthropic rate card (built in)

Costs are computed using these per-million-token rates:

Model	Input	Output	Cache read
Opus	$15	$75	$1.50
Sonnet	$3	$15	$0.30
Haiku	$0.25	$1.25	$0.025

Output is always 5x more expensive than input, so an agent that talks a lot is more expensive than one that listens a lot. Cache reads land between $0.025/M and $1.50/M depending on model — they're an order of magnitude cheaper than fresh input, which is why prompt caching is on by default. If Anthropic publishes new prices, the rate table lives in src/observability/pricing.ts — one edit and every figure on this page updates.

Common tasks

1. "Which agent burned the most this week?" Switch the per-agent table to "Last 7 days," sort by Spend descending. Top row wins. Click into them to see which tasks drove the spend.

2. "We had a tier-2 hotspot — who?" Sort the per-agent table by Tier-2 surcharge. The agent at the top is rotating too late. Two fixes: lower session.tierTwoThresholdTokens for that agent specifically, or audit its prompt for runaway context growth (huge system prompts, accumulating tool outputs, etc.).

3. "Did today cost more than yesterday?" Look at the trend chart's last two bars. The tile at the top shows today's running total; hover the previous bar for yesterday's final. If today is on track to exceed yesterday and it's only noon, that's worth investigating.

4. "Finance wants a monthly export." The CLI is the right tool here: agentx usage report --days 30 produces a CSV-friendly summary. The dashboard is for scanning; the CLI is for piping.

5. "Spend doubled overnight — what changed?" Open the trend chart, find yesterday's bar, then go to Health → Activity for the same date to see if task volume went up. If volume is flat but spend doubled, someone changed a model setting (Sonnet → Opus, or cache disabled). Check the Agents tab for recent edits.

6. "Set a soft budget alert for an agent." The page itself doesn't alert, but the data feeding it does. Pair it with agentx usage today --json in a cron job that pings you when an agent's daily spend crosses your threshold. The dashboard then becomes the place you look once the alert fires.

7. "Compare two clients for monthly billing." Sort the per-agent table, but also cross-reference Business → Projects to see which agents map to which client. The cost page itself doesn't aggregate by client today — for that, lean on agentx usage report plus the projects table.

Troubleshooting

• Spend says $0 across the board. Either the daemon hasn't run any tasks yet (so task_history is empty), or the dashboard is pointed at a different .agentx/db.sqlite than the daemon writes to. Run agentx db tasks from the same working directory — if that returns rows, the page is reading the wrong DB; pass --cwd or run the dashboard from the project root.
• Per-task cost looks suspiciously low or zero. Token counts exist but the model name didn't match the rate table. Check src/observability/pricing.ts — model aliases (claude-3-5-sonnet-latest etc.) need an entry there.
• Tier-2 KPI is huge and you don't know why. Open Health → Rotations and filter for reason tier-2. The rotation log shows which sessions kept growing past the threshold and when.
• Trend chart has gaps. Days where no tasks ran show up as zero-height bars, not gaps. If you see literal gaps, the daemon was offline that day — cross-check with systemctl status agentx.
• The 30-day total disagrees with what Anthropic invoiced. Local computation uses the rate card in src/observability/pricing.ts and the token counts the daemon recorded. Anthropic's invoice may include charges for sessions outside agentx (Claude Code dev sessions, ad-hoc API calls), or for usage from a different API key. The dashboard is "what agentx spent," not "what your Anthropic account spent."
• Page loads but every chart is blank. The browser couldn't reach /api/admin/observability/cost. Open dev tools → Network and look at the response — usually a 401 (token missing) or a 500 (DB locked because another process is writing). The daemon log under Health → Logs names the cause.

A note on what the dashboard does not do

It doesn't budget. It doesn't enforce caps. It doesn't email finance when you cross a number. agentx treats the dashboard as a passive observability surface — the loop "see spend → adjust config → save" is human-driven by design. If you want hard limits, wire agentx usage today --json into a cron with a comparison and a Slack notifier; the dashboard is for the sit-down review, not the realtime guardrail.

CLI parity

• agentx usage today — same today / 7-day rollup as the top tiles, printed to terminal. Useful in shell scripts and cron alerts.
• agentx usage report --days 7 — full Python-analyzer report; richer than the dashboard if you want per-day, per-model, or per-session breakdowns.
• agentx db tasks — raw task_history rows, useful when you suspect the dashboard is misreading something.
• agentx db rotations — raw rotation events, the source of the tier-2 surcharge calculation.