I kept wondering why my Claude Code sessions got so expensive. I couldn’t find a tool that showed me what was exactly in the context window. So I built one.
Zero dependencies. Sits between your tool and the API, captures everything, gives you a web UI to explore it. I’ve tested with Claude, Codex and Aider.
npx context-lens claude "your prompt"
npx context-lens codex "your prompt"
npx context-lens aider --model claude-sonnet-4
npx context-lens -- python my_agent.py
This starts the proxy (port 4040), opens the web UI (http://localhost:4041), sets the right env vars, and runs your command. Multiple tools can share one proxy, just open more terminals.
- Composition treemap — visual breakdown of what’s filling your context: system prompts, tool definitions, tool results, messages, thinking, images
- Cost tracking — per-turn and per-session cost estimates across models
- Conversation threading — groups API calls by session, shows main agent vs subagent turns
- Agent breakdown — token usage and cost per agent within a session
- Timeline — bar chart of context size over time, filterable by main/all/cost
- Context diff — turn-to-turn delta showing what grew, shrank, or appeared
- Findings — flags large tool results, unused tool definitions, context overflow risk, compaction events
- Auto-detection — recognizes Claude Code, Codex, aider, and others by source tag or system prompt
- LHAR export — download session data as LHAR (LLM HTTP Archive) format (doc)
- State persistence — data survives restarts; delete individual sessions or reset all from the UI
- Streaming support — passes through SSE chunks in real-time
Findings
Diff view
Drill-down details
npm start
# Port 4040 = proxy, port 4041 = web UI
ANTHROPIC_BASE_URL=http://localhost:4040 claude "your prompt"
OPENAI_BASE_URL=http://localhost:4040 codex "your prompt"
Add a path prefix to tag requests by tool:
ANTHROPIC_BASE_URL=http://localhost:4040/claude claude "prompt"
OPENAI_BASE_URL=http://localhost:4040/aider aider "prompt"
Codex with a ChatGPT subscription needs mitmproxy for HTTPS interception (Cloudflare blocks reverse proxies). The CLI handles this automatically — just make sure mitmdump is installed:
pipx install mitmproxy
npx context-lens codex "your prompt"
Context Lens sits between your coding tool and the LLM API, capturing requests in transit.
Reverse proxy (Claude Code, aider, OpenAI API tools)
Tool ──HTTP──▶ Context Lens (:4040) ──HTTPS──▶ api.anthropic.com / api.openai.com
│
▼
Web UI (:4041)
The CLI sets env vars like ANTHROPIC_BASE_URL=http://localhost:4040 so the tool sends requests to the proxy instead of the real API. The proxy buffers each request body, parses the JSON to extract context structure (system prompts, tools, messages), forwards the raw bytes upstream with all original headers intact, then captures the response on the way back. The tool never knows it’s being proxied.
Forward HTTPS proxy (Codex subscription mode)
Some tools can’t be reverse-proxied. Codex with a ChatGPT subscription authenticates against chatgpt.com, which is behind Cloudflare. A reverse proxy changes the TLS fingerprint, causing Cloudflare to reject the request with a 403. For these tools, Context Lens uses mitmproxy as a forward HTTPS proxy instead:
Tool ──HTTPS via proxy──▶ mitmproxy (:8080) ──HTTPS──▶ chatgpt.com
│
mitm_addon.py
│
▼
Web UI /api/ingest
The tool makes its own TLS connection through the proxy, preserving its native TLS fingerprint. The mitmproxy addon intercepts completed request/response pairs and posts them to Context Lens’s ingest API. The tool needs https_proxy and SSL_CERT_FILE env vars set to route through mitmproxy and trust its CA certificate.
What the proxy captures
Each request is parsed to extract: model name, system prompts, tool definitions, message history (with per-message token estimates), and content block types (text, tool calls, tool results, images, thinking). The response is captured to extract usage stats and cost. Requests are grouped into conversations using session IDs (Anthropic metadata.user_id), response chaining (OpenAI previous_response_id), or a fingerprint of the system prompt + first user message.
Captured requests are kept in memory (last 100) and persisted to data/state.jsonl across restarts. Each session is also logged as a separate .lhar file in data/. Use the Reset button in the UI to clear everything.
MIT