Latitude for Claude Code - See where Claude Code burns tokens. Hit your limits less.
Trace every Claude Code session. See the full system prompt, every tool call, every subagent, and token cost per turn. One command to install, free, your traces stay in your account. Receive a weekly report with your stats.
Replies
Latitude
Hey everyone,
It's Cesar, founder of Latitude.
If you keep hitting your Claude Code limits faster than you'd expect, this shows you why. Full session trace: system prompt, every tool call, subagent spawns, per-turn token cost. You see where your context actually went and which actions burn the most.
Same thing works if you're using Claude Code as a harness for your own agent: track cost and latency per session, and get recurring failures auto-flagged across runs.
Install is one command:
It's free, traces stay in your Latitude account.
Happy to answer technical questions in the thread.
@latitude @heycesr Love it! I'd love to try it. Does it work with non-anthropic models? For example I'm using auriko.ai to run deepseek or kimi model in claude code. Would it work?
Uselink
this is genuinely what we need, especially being able to spot which sessions burn way more tokens than others lol.
Does this support the Claude Team plan, or API-only for now?
Latitude
@nathan_tran2 both are supported!
Build Check
Oh man! Needed it. Claude is making me crazy, cause I never know where my tokens go. Thanks for helping on this and wish you all the best here!
Does it work for Claude Cowork projects? I mainly burn my limits on Cowork sessions.
Latitude
@michael_vavilov We only support Claude Code at the moment. But if you want to hit your limits less, I suggest you to try Claude Code. It's not that different from Cowork and it can do the same things without spending as many tokens.
Token-burn observability for Claude Code is the kind of thing you only think to build after you blow a session on a regex loop. Curious how you handle the per-tool breakdown when an agent calls bash ten times for the same file read. Is it parsed from stream events or from the response after the fact?
Does the recurring failure detection use heuristics, or are you applying some ML based pattern analysis?
Latitude
@xavier_hernandez2 Both
congratulations, how much overhead does telemetry collection add to latency during active sessions?
Latitude
@olivia_bennett7 It adds no overhead at all. Telemetry data is sent asynchronously and adds no latency to runtime sessions.
- On macOS the installer writes
~/Library/LaunchAgents/so.latitude.claude-code-telemetry.plist which runs
launchctl setenv BUN_OPTIONS=--preload=... on every login. That sets
BUN_OPTIONS for every Bun process on your machine, not just claude — so any
other Bun-based tool you run will also load their preload shim. Wider blast
radius than "just Claude Code."
^^ Might be worthwhile reeling that in a bit. Every Bun process is a bit overreaching.
Latitude
@robert_douglass Thanks, taking a look asap
Latitude
@robert_douglass Unfortunately this was the only reliable way to capture telemetry for the claude desktop app in mac OS. The impact in other Bun processes is neglibile, though. Feel free to DM me if you need more details.
@geclos_latitude makes sense. I'm not overly worried (not a Bun power user :P) but it looked either like an oversight or a forced compromise. You've confirmed the latter.
This is solving a real problem. I run automated agents (social media engagement, security audits, competitor monitoring) and the hardest part isn't building them — it's knowing when they silently fail. I built a custom "doctor" module that diagnoses and self-heals agent errors, but a proper observability layer would have saved me weeks.
The "auto-generated evals from production failures" is the killer feature here. How granular is the token cost tracking per agent task?