Observability + improvement for AI agents
Eden is the observability + improvement platform for AI agents and LLM apps. Capture every LLM call through an OpenAI-compatible gateway, analyse traces in a workspace with a persistent AI assistant, and let /eden-improve turn regressions into shipped PRs.
Free up to 100K trace events / month. No credit card.
Trusted by engineering teams shipping AI agents
The three pillars
Capture. Analyse. Improve.
Comparable to Helicone or Langfuse for the observability layer, with a closed-loop improvement layer they don't have. Adopt any pillar independently; the loop closes when you use all three.
Observability
See every LLM call. With your models.
Drop-in SDKs (Python, TypeScript) and an OpenAI-compatible HTTP gateway that proxies every request to whichever upstream model you configure — OpenAI, Anthropic, Bedrock, self-hosted. Eden is a proxy, not a model vendor.
- OpenAI-compatible gateway — change one line, capture everything
- 7+ source decoders (OpenAI, Anthropic, Gemini, Bedrock, LangChain, …)
- PII pre-masked at ingest; structured evals (LLM-as-judge, code-based, third-party)
Web platform
Workspace, visualisations, and a persistent AI assistant.
Org-aware workspace for teams of any size. Rich visualisations of traces, cost, latency, and eval scores. Data export, webhooks, and a persistent AI assistant sidebar in every view that answers questions against your real data — cited to trace IDs and eval records.
- Multi-tenant orgs, RBAC, SSO/SAML, SCIM, audit log
- Time-series + drill-down: traces, tokens, cost, eval regressions
- AI assistant sidebar — every answer is cited to a trace or eval
CLI agent
Turn failing traces into shipped PRs.
`eden` is a CLI coding agent (Aider fork) with a flagship `/eden-improve` command. It reads failing traces and eval results, hypothesises a root cause, proposes a patch, runs your eval suite, and opens a PR. Closed loop, end-to-end.
- /eden-improve — read → hypothesise → patch → eval → PR
- /eden-trace, /eden-eval, /eden-compare for targeted workflows
- Faster cycle every iteration — KB-grounded, never re-debugs a known regression
5-minute integration
One line. Your models. Full observability.
Point your existing OpenAI SDK at Eden's gateway. Every call is captured and queryable in the portal — no other code changes.
from openai import OpenAI
client = OpenAI(
base_url="https://gateway.eden.example/v1", # Eden
api_key=os.environ["EDEN_API_KEY"],
)
resp = client.chat.completions.create(
model="gpt-4o", # any upstream: OpenAI, Anthropic, Bedrock…
messages=[{"role": "user", "content": "Hello"}],
)The differentiator
From "this regression" to "this PR"
Helicone and Langfuse show you traces. Eden ships the fix. The/eden-improvecommand reads failing traces and eval results from your workspace, proposes a patch to your prompt or code, runs your eval suite, and opens a PR.
- Hypothesises root causes — cited to trace IDs and eval records
- Patches prompt or code, then runs your evals locally
- Opens a PR only if evals pass; iterates with the failure log otherwise
- Gets faster every cycle — KB-grounded, never re-debugs a known regression
Why did latency spike on Tuesday for project copilot-prod?
Latency p95 went from 1.2s → 2.4s on Tuesday 14:00 UTC, driven by a regression in planner.v3 triggered 2,400 traces with multi-step tool loops.
Web platform
Ask your data. Cited answers.
A persistent AI assistant sidebar lives in every view of the Eden portal. It queries your real traces and eval records — not generic web results — and cites the trace ID or eval record behind every claim.
- Org management with RBAC, SSO/SAML, SCIM, audit log
- Trace + eval visualisations: time-series, drill-down, regression alerts
- Export to JSONL/CSV, scheduled S3/GCS, webhook delivery
"We swapped our base URL to Eden's gateway on a Friday. By Monday we had full per-trace cost and eval data we had been hand-rolling for two quarters."
"/eden-improve caught a prompt regression on a tool-use branch we didn't even have alerting on. The PR merged the same day."
Ship better agents, faster.
Capture, analyse, improve — start free in under five minutes.