Skip to content
llmprof

llmprof

pprof for your LLM context. See where every token and dollar goes - system prompt, tool schemas, RAG chunks, history - as a flame graph, and get told what to cut.
Try the live demo Quickstart View on GitHub

You profile CPU and memory. Why are you flying blind on the most expensive resource in your AI app - the context window?

llmprof is a local, zero-config profiler for LLM calls. Point your client’s base_url at it; it forwards requests to the real provider (your API key passes straight through), and on the way it breaks each request’s prompt tokens into the components that make it up, prices the call, and flags what is wasteful.

llmprof dashboard showing a context flame graph, optimization findings, and reclaimable cost for a single call

Open the interactive demo → - the real dashboard on a recorded session, no install.

What you get

Section titled “What you get”

Context flame graph

One request’s tokens broken down by component, with per-tool drill-down. The fat bar is usually the waste.

Reclaimable dollars

A waste detector that flags duplicated content, unused tool schemas, and uncached prefixes - with a “$X/mo reclaimable” headline.

Context timeline

Watch context creep across an agent’s turns: history balloons while the system prompt and tools stay flat.

Cost leaderboard

Which prompt template actually drives your bill - grouped by system prompt and tool set, ranked by total cost.

30-second try

Section titled “30-second try”
Terminal window
pipx install llmprof
llmprof up
# then point your client at http://localhost:4000/v1 and open http://localhost:4000

No Python? npx llmprof up runs it with no Python install - see Installation.

Runs fully local. Your prompts never leave the machine. Works with any OpenAI- or Anthropic-compatible client, including Claude Code and the Codex CLI.

Where to next

Section titled “Where to next”