Capturly

Record, replay, and debug any HTTP API — with built-in AI traffic intelligence

Capturly is an HTTP proxy that captures, replays, and inspects API traffic. It works with any REST/GraphQL endpoint but goes deeper on AI protocols — automatically detecting OpenAI, AGUI (Agent GUI), and generic Bedrock Converse traffic, extracting prompts, combining supported streaming chunks, and surfacing tool calls.

Quick Start

# Install
pip install capturly

# Point at OpenAI and watch traffic
capturly --mode log --backend https://api.openai.com --combine-chunks

# Open dashboard at http://localhost:9090
# See your system prompts, tool definitions, and complete conversations

Use Cases

Debug AI Applications

See exactly what your app sends to OpenAI:

capturly --mode log --backend https://api.openai.com --combine-chunks
# Configure your app to use http://localhost:9999 instead of api.openai.com

The dashboard extracts:

• System prompts (including multi-part array content)
• Tool definitions and tool calls
• Token usage
• Complete conversation history (streaming chunks combined)
• AGUI agent runs: messages, tool calls, reasoning, and state

Inspect Bedrock Converse Traffic

Capturly inspects generic Bedrock Converse JSON and ConverseStream binary Amazon EventStream traffic for Claude and other compatible models, including AWS /model/{id}/converse and /model/{id}/converse-stream endpoints and compatible gateways. This does not include the native Anthropic Messages API. The dashboard shows system and message history, new inputs, tools and results, generated text, usage, and stop reason. EventStream inspection is always enabled; --combine-chunks applies to SSE only.

Capturly forwards requests but does not provide credentials or re-sign requests with AWS SigV4. Use authentication that is already valid for the upstream endpoint, or a gateway that handles authentication; do not assume local SDK signing works automatically through the proxy.

Create Deterministic Test Fixtures

Record real API responses, replay them in tests:

# Record once
capturly --mode record --backend https://api.stripe.com
curl http://localhost:9999/v1/charges

# Replay forever (no network, instant, deterministic)
capturly --mode replay
pytest tests/

Performance Testing

Stress test your service with cached responses:

# Replay with zero delay (simulate fast API)
capturly --mode replay --delay 0

# Or simulate slow upstream
capturly --mode replay --delay 2000

Development Speedup

Cache common endpoints, record new ones automatically:

capturly --mode hybrid --backend https://api.example.com
# First request: hits backend, caches response
# Subsequent requests: instant replay from cache

Dashboard

The built-in web dashboard (default: http://localhost:9090) shows a newest-first traffic list with filters and tag chips (AI, AGUI, SSE, Tools, Tool Results), plus per-request detail views that surface system prompts, new inputs vs. conversation history, tool calls, token usage, and AGUI agent runs. See the Dashboard Guide.

Modes

Configuration

CLI arguments:

capturly --mode log --backend https://api.openai.com --port 8080

Environment variables:

export CAPTURLY_RECORDINGS_DIR=./recordings
capturly --mode record --backend https://api.example.com

Installation

pip install capturly

Requirements: Python 3.9+

Documentation

• Getting Started — 5-minute tutorial
• Mode Reference — Detailed behavior of each mode
• Dashboard Guide — Traffic list, filters, detail views, API
• Configuration — All options, all sources

Contributing

Contributions welcome! See CONTRIBUTING.md for guidelines.

License

MIT License - see LICENSE file.