llmview

Chrome DevTools for LLM APIs.

A local proxy that intercepts, inspects, replays, and budgets your LLM API calls — with a real-time dashboard. Zero code changes. One binary. Interactive debugging.

English | 中文

---

Quick Start

# Install (pick one)
go install github.com/uucz/llmview@latest
# or: download from https://github.com/uucz/llmview/releases
# or: docker run -p 4700:4700 ghcr.io/uucz/llmview

# Run
llmview

# Point your AI tools at it
export OPENAI_BASE_URL=http://localhost:4700/proxy/openai/v1
export ANTHROPIC_BASE_URL=http://localhost:4700/proxy/anthropic

Open http://localhost:4700 — every LLM call now streams through your dashboard.

Why

I spent $47 debugging an AI agent last week. Blind. No idea which calls were burning tokens, which prompts were bloated, or where the agent was looping.

Existing tools are either:

• Cloud-hosted (Helicone, AgentOps) — your prompts leave your machine
• SDK-based (Langfuse, Phoenix) — requires code changes and framework lock-in
• CLI-only (llm-interceptor) — no UI, just raw logs

llmview sits between your AI tools and the API. You change one environment variable. That's it. Your prompts never leave your machine. Everything shows up in a real-time dashboard.

Features

Feature	Description
Real-time timeline	Watch API calls stream in as they happen
Live token streaming	See the response being generated token-by-token
Request/response viewer	Parsed message threads + syntax-highlighted JSON bodies
Filter & search	Filter by provider, status code, model. Search across all calls.
Request replay	Re-send any call with modified parameters or a different model
Budget enforcement	Set a session cost ceiling — proxy returns 402 when exceeded
Export	Download session data as JSON or CSV
Cost tracking	Per-call and session-total cost with per-model pricing
Multi-provider	OpenAI, Anthropic, Ollama — all through one dashboard
Zero code changes	Just set an environment variable
Single binary	One ~10MB file. No database to install. No Docker required.
Local & private	SQLite storage. Nothing leaves your machine.
Dark theme	Because you're probably running this at 2am

Supported Providers

Provider	Environment Variable	Works With
OpenAI	OPENAI_BASE_URL=http://localhost:4700/proxy/openai/v1	GPT-4o, o1, o3, any OpenAI model
Anthropic	ANTHROPIC_BASE_URL=http://localhost:4700/proxy/anthropic	Claude Opus, Sonnet, Haiku
Ollama	OLLAMA_HOST=http://localhost:4700/proxy/ollama	Llama, Mistral, Qwen, any local model

Works with any tool that uses these SDKs: Claude Code, Cursor, Aider, LangChain, CrewAI, OpenAI Python/Node SDK, Anthropic SDK, and more.

How It Works

Your Agent / IDE / Script
         │
         ▼  (just an env var change)
    ┌─────────┐
    │ llmview │ ← intercepts, logs, calculates cost
    └────┬────┘
         │
         ▼  (forwards to real API)
   OpenAI / Anthropic / Ollama

llmview is a reverse proxy. It receives the request, records it, forwards it to the real API, records the response, calculates the cost, and pushes everything to the dashboard via WebSocket. Streaming responses are forwarded chunk-by-chunk with zero added latency.

Configuration

# Custom port (default: 4700)
llmview --port 8080

# Custom database path (default: ~/.llmview/llmview.db)
llmview --db /path/to/data.db

# Set a budget ceiling (proxy returns 402 when exceeded)
llmview --budget 5.00

Model Pricing

llmview ships with built-in pricing for popular models (GPT-4o, Claude Sonnet, etc.). Local models (Ollama) are tracked as free. Pricing updates with new releases.

REST API

llmview exposes a JSON API for programmatic access:

# Current session stats
curl http://localhost:4700/api/session

# List recent calls
curl http://localhost:4700/api/calls?limit=20&offset=0

# Get full request/response for a specific call
curl http://localhost:4700/api/calls/{id}

# Replay a call with overrides
curl -X POST http://localhost:4700/api/replay \
  -H 'Content-Type: application/json' \
  -d '{"call_id":"abc123","overrides":{"model":"gpt-4o-mini"}}'

# Get config (budget info)
curl http://localhost:4700/api/config

# Health check
curl http://localhost:4700/api/health

Building from Source

git clone https://github.com/uucz/llmview.git
cd llmview
make build    # builds UI + Go binary
make test     # runs all tests

Requires: Go 1.25+, Node.js 18+ (for UI build). No C compiler needed — pure Go.

Troubleshooting

macOS: "permission denied" when running the binary

Downloaded binaries don't have execute permission by default. Fix:

chmod +x llmview-darwin-arm64
./llmview-darwin-arm64

macOS: "cannot be opened because the developer cannot be verified"

macOS Gatekeeper blocks unsigned binaries. Go to System Settings > Privacy & Security, scroll down, and click "Open Anyway". Or remove the quarantine attribute:

xattr -d com.apple.quarantine llmview-darwin-arm64

"command not found: llmview"

If you downloaded the binary, you need to use the full path or ./ prefix:

# Run from current directory
./llmview-darwin-arm64

# Or move to a directory in your PATH
sudo mv llmview-darwin-arm64 /usr/local/bin/llmview
llmview

If you used go install, make sure $GOPATH/bin (usually ~/go/bin) is in your PATH:

export PATH="$HOME/go/bin:$PATH"

Common mistakes

# WRONG: "go" is for compiling source code, not running binaries
go llmview

# WRONG: "run" is not a system command
run llmview

# WRONG: just the filename doesn't work without PATH setup
llmview-darwin-arm64

# CORRECT:
./llmview-darwin-arm64

Roadmap

• Real-time proxy + dashboard
• Token cost tracking
• Multi-provider support
• Request/response detail viewer with message thread parsing
• Export sessions to JSON/CSV
• Filter & search calls by provider, status, model
• Request replay with parameter overrides
• Budget enforcement with real-time progress bar
• Prompt diff viewer (compare request variations)
• VS Code extension
• Breakpoints (pause before dangerous operations)
• Session history browser

License

MIT

---

If llmview saved you from a surprise API bill, consider giving it a star.