C
codealisha99/tracelight
LLM observability: a streaming-aware proxy for Anthropic and OpenAI APIs, agent trace trees, cost and latency dashboard.
updated 2026-11-14
View on GitHub ↗Link checked 2026-11-17 · HTTP 200 · opens the vendor's own site
- Star count
- 0 ★ · codealisha99/tracelight · read from the GitHub search API · (checked 2026-11-19 20:00:48)
- Source
- read directly from its GitHub page · type: open-source project
- Harvest history
- first seen 2026-11-15 · last seen 2026-11-20 · seen 6× · listed by 1 sources
- Score
- score 103 · 1 sources · repo
FAQ (answers come from our records)
Who is it for?
We filed it under these tags: Open Source. The tags come from the entry itself, not a taxonomy we imposed.
Who maintains it?
repository
codealisha99/tracelight;first catalogued 2026-11-15;last seen 2026-11-20。There is no official certification or quality rating here — we record sources and dates only.README excerpt (English, untranslated) · Read the full README on GitHub →
# tracelight
Excerpted from the repository README (abridged)
Live demo: https://tracelight-demo.vercel.app. A read-only dashboard over a committed snapshot, captured by the real proxy from scripted agents against a mock upstream. The proxy and write endpoints return 403.
Landing page: https://tracelight-site.vercel.app. Source is in site/.
LLM and agent observability you can point an SDK at. Change one base URL and every Anthropic or OpenAI call is forwarded untouched and recorded: streamed responses are reconstructed from their SSE chunks, with time to first token, inter-chunk stalls, tokens per second, prompt-cache reads, computed cost, and anomaly flags. Agent runs (LLM calls, tool calls, sub-agents) show as one trace tree, with a dashboard for filtering, metrics, sessions, scores, and a "needs attention" queue.
Small Langfuse / Helicone / LangSmith in TypeScript, with SQLite storage, a Python SDK, and a React dashboard. Inspired by langfuse/langfuse, Helicone/helicone, Arize-ai/phoenix, and the OpenTelemetry GenAI conventions. The design is my own; no code is copied.
## Quickstart
Keyless demo first. It needs no API keys: a mock upstream stands in for api.anthropic.com and api.openai.com.
bash
make install # pnpm install + uv sync for the Python SDK tests
make test # vitest (unit, proxy, SDK, real-SDK compatibility) + pytest
make demo # mock upstream :4402, tracelight :4400, scripted agent, dashboard
# open http://127.0.0.1:4400 (Ctrl-C stops the demo)
make demo runs a support agent through the proxy using the official @anthropic-ai/sdk. It makes a tool-using plan, a slow-first-token answer, a sub-agent with a slow stream, a failing tool, and a rate-limited call, plus a second session with prompt caching. Then it attaches an eval score and a user score. In the dashboard you get the trace tree and waterfall, the flagged failures, and the metrics page.
### Real traffic
Run the server, then point your SDK at it. Upstreams default to the real APIs.
bash
pnpm server # tracelight on :4400 (TRACELIGHT_DB defaults to ./data/tracelight.db)
ts
import Anthropic from '@anthropic-ai/sdk';
const client = new Anthropic({ baseURL: 'http://127.0.0.1:4400/anthropic' }); // key stays in your env
python
from openai import OpenAI
client = OpenAI(base_url="http://127.0.0.1:4400/openai/v1")
Agent tracing, so calls and steps form one tree:
ts
import { configure, currentHeaders, span, trace } from './src/sdk/tracelight.js';
configure({ baseUrl: 'http://127.0.0.1:4400' });
await trace('support-agent', async () => {
await client.messages.create(params, { headers: currentHeaders() }); // LLM span attaches under the run
await span('search_docs', () => searchDocs(q), { input: { q } });
}, { sessionId: 'user-42' });
python
import tracelight
tracelight.configure("http://127.0.0.1:4400")
with tracelight.trace("support-agent", session_id="user-42"):
client.chat.completions.create(..., extra_headers=tracelight.headers())
Set TRACELIGHT_REDACT=email,credit_card to redact PII in stored payloads. See .env.example.
## Architecture
mermaid
flowchart LR
A[Agent code Anthropic / OpenAI SDK] -->|base URL| P[Proxy :4400]
T[TS / Python SDK trace · span] -->|POST /api/spans| API
P -->|raw bytes, no buffering| U[(Upstream API)]
Landing page: https://tracelight-site.vercel.app. Source is in site/.
LLM and agent observability you can point an SDK at. Change one base URL and every Anthropic or OpenAI call is forwarded untouched and recorded: streamed responses are reconstructed from their SSE chunks, with time to first token, inter-chunk stalls, tokens per second, prompt-cache reads, computed cost, and anomaly flags. Agent runs (LLM calls, tool calls, sub-agents) show as one trace tree, with a dashboard for filtering, metrics, sessions, scores, and a "needs attention" queue.
Small Langfuse / Helicone / LangSmith in TypeScript, with SQLite storage, a Python SDK, and a React dashboard. Inspired by langfuse/langfuse, Helicone/helicone, Arize-ai/phoenix, and the OpenTelemetry GenAI conventions. The design is my own; no code is copied.
## Quickstart
Keyless demo first. It needs no API keys: a mock upstream stands in for api.anthropic.com and api.openai.com.
bash
make install # pnpm install + uv sync for the Python SDK tests
make test # vitest (unit, proxy, SDK, real-SDK compatibility) + pytest
make demo # mock upstream :4402, tracelight :4400, scripted agent, dashboard
# open http://127.0.0.1:4400 (Ctrl-C stops the demo)
make demo runs a support agent through the proxy using the official @anthropic-ai/sdk. It makes a tool-using plan, a slow-first-token answer, a sub-agent with a slow stream, a failing tool, and a rate-limited call, plus a second session with prompt caching. Then it attaches an eval score and a user score. In the dashboard you get the trace tree and waterfall, the flagged failures, and the metrics page.
### Real traffic
Run the server, then point your SDK at it. Upstreams default to the real APIs.
bash
pnpm server # tracelight on :4400 (TRACELIGHT_DB defaults to ./data/tracelight.db)
ts
import Anthropic from '@anthropic-ai/sdk';
const client = new Anthropic({ baseURL: 'http://127.0.0.1:4400/anthropic' }); // key stays in your env
python
from openai import OpenAI
client = OpenAI(base_url="http://127.0.0.1:4400/openai/v1")
Agent tracing, so calls and steps form one tree:
ts
import { configure, currentHeaders, span, trace } from './src/sdk/tracelight.js';
configure({ baseUrl: 'http://127.0.0.1:4400' });
await trace('support-agent', async () => {
await client.messages.create(params, { headers: currentHeaders() }); // LLM span attaches under the run
await span('search_docs', () => searchDocs(q), { input: { q } });
}, { sessionId: 'user-42' });
python
import tracelight
tracelight.configure("http://127.0.0.1:4400")
with tracelight.trace("support-agent", session_id="user-42"):
client.chat.completions.create(..., extra_headers=tracelight.headers())
Set TRACELIGHT_REDACT=email,credit_card to redact PII in stored payloads. See .env.example.
## Architecture
mermaid
flowchart LR
A[Agent code Anthropic / OpenAI SDK] -->|base URL| P[Proxy :4400]
T[TS / Python SDK trace · span] -->|POST /api/spans| API
P -->|raw bytes, no buffering| U[(Upstream API)]
Listed badge (optional)
Copy this line into the README if you want the listing mark. Optional, and it does not affect being listed.
[](https://anclum.com/en/s/codealisha99-tracelight/)
- Source
- github · https://github.com/codealisha99/tracelight
- Browse by field
- Open source
- added
- today
More entries
Filed under
Every tag opens the full list for that filter. Counts are the live catalogue figures.