Skip to the install command

See the 5-hour limit coming. Know when to /compact.

Live plan limits, spend and context for every Claude Code and Codex session. Free, open source, runs on your Mac.

Install commands: 
brew install konvuinc/tap/konvu-telemetrykonvu-telemetry setup

Requires macOS, Homebrew, and Claude Code or Codex. Works with CLI and Desktop. Cursor next.

Konvunow

Codex session running hot

🔥 $24.6 forecast for the next 10 prompts

💸 $46.8 spent so far

Coding Agents monitoring

Live sessions

6

Sessions with activity in the last 20 minutes; running and idle state uses a 5-minute freshness window.

All agents Claude Codex
Spent beyond plan$59.10across 2 billing sessions
Most expensive session$46.80Migrate authentication flows
Biggest forecast+$24.60Migrate authentication flows
Average context used57%across 2 billing sessions
Costliest subagent$6.20Migrate auth… · Auth mapper
Spending real money·2 sessions
Usage limitweek 100%Resets in 2d 4hcredits 88%Resets in 18d 6h
Sort  Highest forecast⌄
SessionSpend / next 10 promptsResponsible forContextSubagentsActivity / age
Migrate authentication flowsGPT-5 Codex · high effort
$46.80+$24.60 in next 10 prompts
79% of all money spent
91%AI review off outside plan
4 spawned · 2 live$9.70 spent
0s agoAge unavailable
Investigate CI failuresGPT-5 Codex · high effort
$12.30+$2.10 in next 10 prompts
21% of all money spent
24%AI review off outside plan
1 spawned · 0 live$1.30 spent
45s agoAge unavailable
Within your plan·4 sessions
Usage limit5h 68%Resets in 1h 52mweek 74%Resets in 3d 9h
Sort  Highest forecast⌄
SessionPlan / next 10 promptsResponsible forContextSubagentsActivity / age
Trace vulndb enrichment gapsClaude Opus 4.1 · high effort
Included in your planNext 10 prompts: +27% of 5-hour limit
17% of your 5-hour limit
84%Next AI review in 3 prompts
3 spawned · 2 live
3m agoAge unavailable
Bump lockfile for transitive CVEClaude Sonnet 4 · low effort
Included in your planNext 10 prompts: +17% of 5-hour limit
3.6% of your 5-hour limit
14%Next AI review in 7 prompts
0 spawned · 0 live
3m agoAge unavailable
Implement subscription-limit dashboardClaude Opus 4.1 · high effort
Included in your planNext 10 prompts: +12% of 5-hour limit
24% of your 5-hour limit
63%AI review due now
2 spawned · 1 live
1m agoAge unavailable
Review pull-request feedbackClaude Sonnet 4 · medium effort
Included in your planNext 10 prompts: +2.4% of 5-hour limit
9.8% of your 5-hour limit
24%Next AI review in 6 prompts
0 spawned · 0 live
2m agoAge unavailable

Why live

Find the session driving your usage.

Each session shows its share of your spend or plan limit, what its next 10 prompts could cost, and how much of its context is stale.

Notifications

When is it time to check my usage?

Get a browser notification when a provider limit crosses 50%, 80%, or 100%, or a session spending beyond your plan is forecast to use at least $10 over the next ten prompts.

Usage limit

Konvunow

Claude 5h limit at 80%

80% of your 5h limit is used.

Spending forecast

Konvunow

Codex session running hot

🔥 $24.6 forecast for the next 10 prompts

💸 $46.8 spent so far

Illustrative browser notifications. Appearance depends on your browser and OS.

Usage summaries also appear where you work.

  • Claude Code CLIStatus line
  • Codex CLIStop hook
  • Claude DesktopEnd of reply
  • Codex DesktopEnd of reply

Context review

Is it time to /compact?

Every 10 prompts, a small model rates each session’s context: still needed, or finished work. It runs on your own Claude or Codex login and costs a few cents per review.

When enough of the window is finished work, the status line shows /compact suggested. Click it for a command that says what to keep and what to drop.

● Included· 5h reset: 1h 52m 68% · Week reset: 3d 9h 74%

Context 84% · /compact suggested

This session 17.0% of 5h limit ━━━▶ 44.0% forecasted in next 10 prompts

🔗 Open live dashboard

  • Needed
  • Not needed now
  • Unreviewed
  • Free space

Compact this session

41% finished or replaced
Claude Code
/compact Preserve: vulndb enrichment gap trace; enrich_advisories.py edits; failing test_epss_backfill error; user’s choice to skip withdrawn advisories. Drop: finished Docker cleanup; rejected full re-ingest approach.

Illustrative session. The suggestion appears once finished or replaced work fills a fifth of the window.

Fan-out

Which subagents are costing the most?

See how many subagents a session has spawned, how many are still live, and what they have spent. The dashboard highlights the costliest subagent and its parent session.

Open the session view to inspect the context each subagent received and its estimated API-equivalent cost. Subagent spend belongs to the parent session, so do not add it to the session total again.

Fan-out · Codex · GPT-5.6 Sol

One task, as a tree

codex_b2cc0f6a-28be-… · 1,049 calls

  • Root sessionspawned 451.2M
    own 51.2M209 calls
  • Subagent 1102M
    inherited 51.2Mown 50.4M248 calls
  • Subagent 291.3M
    inherited 51.2Mown 40.1M197 calls
  • Subagent 396.5M
    inherited 51.2Mown 45.2M223 calls
  • Subagent 486.2M
    inherited 51.2Mown 35M172 calls
inherited contextown tokens

Tokens

427M

×3.4 the sequential run

Wall time

2,939 min

235 min slower than sequential

API-equivalent cost

$486

list price, GPT-5.6 Sol

Fan-out to 2 to 5 subagents used 2.6x to 5.9x the input tokens of sequential work, and was never faster (Systima, 2026-07-22).

Try it on your next session.

Two lines to install. Free, open source, and it runs on your Mac.

Install commands: 
brew install konvuinc/tap/konvu-telemetrykonvu-telemetry setup

Requires macOS, Homebrew, and Claude Code or Codex. Works with CLI and Desktop. Cursor next.

How it works

  1. 01

    Claude Code and Codex write session transcripts to your Mac as they run.

  2. 02

    A local collector reads them and fetches provider-reported account limits and reset times using the logins already on your Mac. It does not retain credentials.

  3. 03

    The dashboard groups live sessions into spending beyond your plan and usage included in it. Each row shows its contribution, a forecast for the next 10 prompts, context, and subagents.

  4. 04

    With the context review on, a small model rates each live session’s context every 10 prompts, through that session’s own provider login. Konvu never receives the excerpts.

  5. 05

    Usage appears in each client and at 127.0.0.1:7824. Browser notifications are optional.

What it is not

Konvu reads transcripts after your agent writes them. It does not proxy agent requests or change how the agent runs. It contacts provider usage endpoints, sends anonymous product events by default, and sends short session excerpts to that session’s own provider for the context review.

Data sources

Coverage follows what each agent writes to disk and which surfaces it offers.

Claude Code and Desktop

Transcripts under ~/.claude/projects. Statusline merged into ~/.claude/settings.json.

Live account limits and reset times from Anthropic, session and subagent usage, and a context map reviewed by Claude Haiku. Status line in the CLI; usage box at the end of a Desktop reply, after tool calls by default.

Codex CLI and Desktop

Transcripts under ~/.codex/sessions. Stop hook merged into ~/.codex/hooks.json.

Live account limits and reset times from the Codex app-server, session and subagent usage, and a context map reviewed by a small GPT model. Stop hook summary in the CLI; usage box at the end of a Desktop reply. Both appear after tool calls by default.

Cursor Coming

Local workspace state written by the Cursor app.

Cursor writes its own cost figure and no per-request token split, so it gets its own path. Follow the repository for the release.

Subscription limits come from the providers. Dollar costs and Codex credit equivalents are estimates, not invoices. macOS only.

Questions

What leaves my machine?

Nothing with session content goes to Konvu. The collector uses your existing Claude and Codex logins to fetch account limits every two minutes, without storing credentials or raw responses. The context review, on by default within your plan, sends short excerpts of a session to that session’s own provider: Claude to Anthropic, Codex to OpenAI, with tools disabled. Run konvu-telemetry context-analysis off to stop it. Anonymous product events go to PostHog by default. They contain no session content or token usage. Run konvu-telemetry telemetry off to disable them and delete unsent events.

Does it slow my agent?

Konvu does not sit in the agent request path. Its status line and hooks read local usage data. The collector fetches account limits and runs the context review in the background, outside the agent’s session.

How is it different from CodeBurn or ccusage?

CodeBurn and ccusage itemize what you spent after the session, and do it well. Konvu reads the session while it is running and warns you before the next prompt. It also tells you when a session’s context is stale enough to compact.

How do notifications reach me?

The local dashboard sends browser notifications. Allow them there and under macOS Notifications, then keep a dashboard tab open with your Mac awake. The panel can send a test notification.

How does it decide what to compact?

Claude Haiku, for Claude sessions, or a small GPT model, for Codex, rates each piece of context from 0 to 10 against the session’s open goal. Only finished, replaced or abandoned work counts toward a /compact suggestion. Context that is not needed now but could be looked up again stays. The ratings are estimates, so read the command before you send it.

What does the context review cost?

A few cents per review at API prices. Within your plan it comes out of your allowance. It makes at most 30 model calls per session and 60 overall per hour, and pauses when a plan limit is 90% used. Sessions billed beyond your plan are reviewed only after konvu-telemetry context-analysis on --allow-paid. Each run’s cost shows on its session in the dashboard.

Does it change my agent’s settings?

konvu-telemetry setup starts a per-user LaunchAgent and registers the Claude Code status line plus hooks for Codex CLI and both Desktop apps. In Codex, open /hooks and trust the Konvu hooks. Provider transcripts are never modified.

Is it free and open source?

Yes. MIT licensed at github.com/KonvuInc/konvu-telemetry, so you can check what it reads. No card, no account.

How do I remove it?

konvu-telemetry uninstall removes the service, the Konvu entries in your agent settings, your data under ~/.konvu/telemetry and the Homebrew package. It leaves your agent transcripts and your own hooks in place.

Who builds this?

Konvu, a company in Brooklyn, New York. We also build a separate product for security teams at konvu.com.

Install it before the next long session.

See what your sessions are using now and what the next 10 prompts could consume. Follow spend, plan limits, subagents, and stale context from one local dashboard.

Install commands: 
brew install konvuinc/tap/konvu-telemetrykonvu-telemetry setup

Requires macOS, Homebrew, and Claude Code or Codex. Works with CLI and Desktop. Cursor next.