Cloud quickstart
Traza Cloud runs the open traza-server engine for you. Everything below talks
to it through the gateway at https://api.trazadb.com with a project API key;
every engine route a project may use works there exactly as the
engine reference describes.
1. Sign in and create a project
Sign in at app.trazadb.com with GitHub, or with your email address: the email carries a sign-in link and a six-digit code you can paste instead.
Create an organization, then a project. A project is one engine tenant on one of our cells, with its own data, retention and API keys. The Free plan allows three projects.
2. Create an API key
On the project’s Keys tab, create a key. For this walkthrough choose the
full scope, which can send spans, read them and change run fuses. The key is
shown once; the console stores only its hash.
export TRAZA_API_KEY=tzk_…
API keys explains the three scopes. In production, give an
exporter an ingest key and a dashboard or agent a read key.
3. Send a span
The quickest check is one span in the engine’s native JSON. session.id makes
it part of a run, which is what Guard and Replay act on.
curl -s https://api.trazadb.com/v1/spans \
-H "Authorization: Bearer $TRAZA_API_KEY" \
-H 'Content-Type: application/json' \
-d '[{"trace_id": "trace-1", "span_id": "span-1", "name": "search_orders.tool",
"service": "support-agent", "status": "ok",
"start_time_unix_nano": 1792062221000000000,
"end_time_unix_nano": 1792062221250000000,
"attributes": {"session.id": "sess-1", "gen_ai.operation.name": "execute_tool",
"gen_ai.tool.name": "search_orders"}}]'
{"accepted":1,"durability":"wal"}
accepted counts the spans stored; durability names the mode that
acknowledged them. The console’s onboarding screen flips from “waiting for
your first span” as soon as it lands.
4. Point your application at it
Any OpenTelemetry SDK exports to Traza Cloud with three variables:
export OTEL_EXPORTER_OTLP_ENDPOINT=https://api.trazadb.com
export OTEL_EXPORTER_OTLP_HEADERS="Authorization=Bearer%20$TRAZA_API_KEY"
export OTEL_EXPORTER_OTLP_PROTOCOL=http/protobuf
Per-SDK guides: Python, TypeScript, OpenAI Agents SDK, LangGraph, Claude Code and curl. Read the OpenTelemetry overview first: it covers the three things that trip people up.
5. Look at it
Open explorer on the project page opens the engine’s own trace browser for your project: waterfalls, sessions, token analytics, search and live tail.
6. Hold runs to limits
Give the service a policy. This one stops a run of support-agent that spends
two dollars, takes forty steps or repeats the same call four times:
curl -s https://api.trazadb.com/v1/fuses/policies \
-H "Authorization: Bearer $TRAZA_API_KEY" \
-d '{"name": "support", "match": {"service": "support-agent"},
"limits": {"max_cost_usd": 2.0, "max_steps": 40, "max_repeats": 4},
"action": "stop"}'
Then have the agent ask before each step, and route stops to your phone, Slack or a webhook. Guard on Traza Cloud.
7. Replay a run
On the project’s Replay tab, pick a recent run. The console creates a replay and shows the two environment lines that point the OpenAI or Anthropic SDK at it. Replay on Traza Cloud.
8. Let your coding agent read it
claude mcp add --transport http traza https://api.trazadb.com/v1/mcp \
--header "Authorization: Bearer $TRAZA_API_KEY"
MCP lists the tools and what a key can reach.
9. Next
- Foresight: once an agent has completed runs, its Fate Map shows which paths end well, and a running run gets a verdict.
- Steer: Mission Control, nudges, and approving or killing a run from your phone or Slack.
- X-Ray: the context window any model call saw, and why its prompt cache broke.
- Claude Code and Codex: turn the coding sessions on your machine into runs, from the transcripts they already write.