JamJet Cloud — Developer Quickstart

Add governance, cost tracking, and audit trails to any OpenAI or Anthropic agent in two lines.

You'll have an instrumented agent reporting traces in under five minutes.

Note: JamJet Cloud is the hosted, paid product alongside the open-source JamJet runtime + Engram. Open-source JamJet stays free forever; Cloud adds the multi-tenant dashboard, retained audit trails, hosted memory, and policy/approval surfaces. See the roadmap for what's shipped and what's next.

1. Get an API key

1. Sign in at app.jamjet.dev (GitHub OAuth, Google OAuth, or magic link).
2. Open Settings → Projects and create a project.
3. Copy the API key shown once at creation. It looks like jj_xxxxxxxxxxxx.

The key is not retrievable later. If you lose it, create a new project.

2. Install the SDK

pip install jamjet

The jamjet.cloud submodule ships in the main jamjet package starting at 0.6.0.

Optional: install the LLM SDK you use so JamJet can auto-instrument it.

pip install jamjet[openai]      # auto-instrument OpenAI
pip install jamjet[anthropic]   # or Anthropic
pip install jamjet[cloud-all]   # both

3. Initialize and run

Add two lines at process start, before any LLM call:

import jamjet.cloud as jamjet

jamjet.configure(api_key="jj_xxxxxxxxxxxx", project="my-agent")

Now your existing OpenAI calls are captured automatically:

from openai import OpenAI

client = OpenAI()
client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "hello"}],
)

Open app.jamjet.dev/dashboard/traces — the call appears within ~5 seconds with model, token counts, cost, and duration.

4. Add governance (optional)

Block tools by name

jamjet.policy("block", "payments.*")
jamjet.policy("require_approval", "delete_*")

Blocked tools are filtered out before the model sees them. require_approval pauses execution until a human approves in the dashboard.

Cap spend

jamjet.budget(max_cost_usd=5.00)

Calls that would exceed the budget raise BudgetExceeded instead of running.

Trace your own functions

@jamjet.trace
def lookup_customer(customer_id: str) -> dict:
    ...

Each call becomes a span in the same trace as the LLM calls around it.

Human-in-the-loop approvals

approval_id = jamjet.require_approval(
    action="charge_card",
    context={"amount_usd": 200, "customer": "cust_42"},
    timeout_seconds=300,
)

Blocks until a reviewer approves or rejects in Approvals in the dashboard.

5. Configuration

jamjet.configure(...) accepts:

The same options can be set via env: JJ_API_KEY, JJ_PROJECT, JJ_API_URL.

6. What gets sent (and what doesn't)

By default, JamJet Cloud captures metadata: model name, token counts, latency, cost estimate, tool names, status. Prompt and response content is NOT sent unless you set capture_io=True.

The SDK is fail-open: if api.jamjet.dev is unreachable, your agent keeps running. Failed batches are retried with exponential backoff and dropped after 5 consecutive failures (circuit breaker).

7. What's coming next

Active development. See the roadmap for full timing.

• Multi-agent visibility — agent identity, cross-agent trace propagation, network graph view of how your agents communicate (Phase 1 / Q3 2026).
• Java SDK — same drop-in for Spring AI / LangChain4j (Phase 1).
• Centralized policy enforcement — server-side policy decisions, audited delegation chains.
• Hosted memory — Engram bundled into Cloud, scoped per agent, shared across an agent fleet.
• Replay — re-run any trace from the dashboard with input recordings.
• OTel GenAI ingestion — point your existing Phoenix / OpenLLMetry / Langfuse-instrumented apps at JamJet without an SDK migration.

Troubleshooting