Measure the productivity of Agentic Coding
Your team writes more code with AI agents every week, and the process behind it is invisible. AI Coding Insights connects token spend to what it actually produced: cost, adoption, and productivity in one view.
Start for free or Book a demo
Connect token spend to what is actually produced
AI agents are writing a growing share of your code. Output is up. But the process is invisible: no way to benchmark agents against each other, compare models on anything beyond raw token cost, or show business value and ROI of token spend to leadership.
The billing tab gives you a total and nothing else. When a model takes 5x more turns to reach a good result but costs 10x less per token, it might be the better economic choice. You can only know that with productivity data sitting next to cost — PRs created vs. deployed, commit to production time, and which teams are actually getting value.
Three views. One question: is this worth it?
Stop guessing where the tokens went
Token spend per session, per user, per repo, per model. Sort by cost to find sessions that ran in circles before doing real work. Even on flat-rate Team or Max plans, usage data shows whether each developer needs the plan they're on.
See where AI coding is taking hold
Active sessions over time, session depth by developer, model usage by team. Track whether the rollout is spreading or stuck in one corner. Normalize spend by active user to tell a better story than total spend alone.
See what the agents actually ship
PRs created vs. deployed, time from PR open to merge, code churn and rework rates. The numbers that turn cost comparisons into unit economics: cost per merged PR, cycle time per model, and which teams are getting real output from their AI investment.
Every session, replayable
Open AI Coding Insights to find the run you care about. Sort by cost, drill into the stats, and read the conversation exactly as Claude Code ran it. Turn count sits next to token cost, so you can see which sessions were efficient and which ran in circles.
Every action, traced.
Behind each prompt is a tree of activity: tool calls, MCP servers, permissions, failures. The plugin captures all of it; Dash0 makes it queryable.
Every MCP server, every invocation
Which server, which tool, how long it took, whether it errored. Catch slow or flaky MCP servers without instrumenting them yourself.
Which tools the agent actually uses
Bash, Edit, Grep, custom MCP tools. See what the agent reaches for, how often, and where it fails. Useful when tuning prompts or trimming permissions.
Every stop, every denied permission
PostToolUseFailure, StopFailure, and PermissionDenied events surface as their own panel, so you can see exactly what tripped the agent up.
Connects to your stack, not the other way around.
One open-source plugin, built on standard OpenTelemetry. Install it and session telemetry flows into Dash0 as standard OTel traces over OTLP. No proprietary agent, no lock-in.
Portable, not locked in
Everything is standard gen_ai.* and vcs.* OpenTelemetry over OTLP. If you ever leave Dash0, point the OTLP exporter at any other OTel-compatible backend and the instrumentation still works.
Install the plugin
Install dash0@claude-plugins-official in Claude Code. Chat and LLM-call spans with full token accounting, tool and MCP calls, sub-agent activity, and errors start flowing right away.
Tune what leaves your machine
On Claude Code, prompt and tool I/O are stripped by default (DASH0_OMIT_IO=true). Set DASH0_OMIT_USER_INFO=true to hash user.name and drop email.
Install from the Marketplace
Install dash0-agent-plugin from the Cursor Marketplace. Every span carries repository, branch, commit, and pull-request URL automatically, no extra setup.
Priced per user, not per token.
$10 per user per month. A fraction of what the agents themselves cost.
| Users | Monthly | Annual |
|---|---|---|
| 10 | $100 | $1,200 |
The factory is running.
AI is writing more of your codebase every week. AI Coding Insights shows you what it's building, what it costs, and whether the investment is working, per tool, per team, per model.