---

Why another agent-observability tool?

Because observability stops at the dashboard. You can see that your agent broke — then what?

Every eval tool asks you to hand-author a dataset: sit down, invent questions, write ideal
answers, keep them current as the product changes. That dataset is a guess about what might break.

Production already handed you the real thing: a trace of the exact run that failed, with the exact
input, the exact tool calls, the exact model responses.

The recorded run is the test. Tracely freezes that trace into a hermetic regression case and
replays it on every PR. Everything else — quality scores, failure clusters, suggested fixes, CI
verdicts, trends — is derived from the trace. There are no hand-authored datasets.

	Dataset-first tools	Tracely
Where tests come from	You write them	Promoted from real failing traces
Fidelity to production	A guess	The exact failing run, byte for byte
Cost to replay in CI	Live model calls	$0 — recorded tool/LLM fixtures
What happens on regression	A dashboard number moves	The PR is blocked

---

The spine

The product maps onto four steps. Each one is a page in the app.

1 · Observe — every run, hierarchically

Traces arrive over plain OTLP. Agent semantics (agent.id, conversation.id, turn, step) are
promoted to first-class indexed columns, so runs group into conversation threads instead of a flat
span soup. The waterfall shows agent → tool → thinking → generation, with the failing span in red.

Evaluators are columns on the trace table, not a separate tab — each one grades at conversation,
run or span level and writes its verdict into the grid. Scores stream in live over SSE as judges
finish, so you watch a run get graded in place.

2 · Detect & triage — failures group into issues

Online evaluators grade every run as it lands (LLM-as-judge at conversation / run / span level, plus
structural checks that need no model at all). Failures then cluster — structurally and semantically —
so 31 broken runs become one issue with a count, not 31 rows to read.

3 · Test — freeze the failure

One click promotes a failing trace into a hermetic case: recorded input, tool and LLM outputs bundled
as fixtures, and a fail-to-pass contract attached — the case must fail on the old code and pass on
the fix, or the promotion isn't trusted.

4 · Ship — block the PR

The suite replays in CI against recorded fixtures: deterministic, offline, no API keys and no model
spend. tracely gate exits non-zero, posts a commit status, and upserts a PR comment.

Plus — trends and cross-metric analysis

Daily failure and gate pass-rates, latency percentiles, token spend, and per-agent meta-analysis
(Spearman correlations + z-score outliers, LLM-synthesized).

---

Quickstart

Prerequisites: Docker + Docker Compose. (For local dev also uv and Node 20+ / pnpm.)

Everything in Docker — the whole product, populated

git clone https://github.com/Jwuthri/Tracely && cd Tracely
docker compose --profile demo up -d --build --wait
open http://localhost:3001

That brings up ClickHouse, Postgres, Redis and MinIO, runs every migration, seeds the default project
and ingest key (tracely_dev_key), then populates traces, clusters, cases and gates — so the app
opens with the screenshots above rather than an empty shell.

docker compose down            # stop  (add -v to wipe data)

Host ports default to web :3001 and backend :8000; remap with TRACELY_WEB_PORT / TRACELY_BACKEND_PORT.
backend/worker/frontend run off source volume-mounts, so most edits need only docker compose restart <svc> —
except the Celery worker, which doesn't hot-reload.

Local dev (hot reload)

cp .env.example .env
make infra-up      # clickhouse, postgres, redis, minio
make install       # uv sync + pnpm install
make migrate       # ClickHouse DDL + Alembic (Postgres)
make seed          # default project + ingest key → tracely_dev_key

make backend       # FastAPI  :8000  (OpenAPI at /docs)   ┐
make workers       # Celery ingestion/eval worker          ├ three terminals
make frontend      # Next.js  :3001                        ┘

make demo          # populate the WHOLE product: traces + clusters + cases + gates
make test          # backend unit tests (no infra, ~6s)

Deploy your own

One click provisions the whole stack on Railway — API, worker, UI, Postgres
(pgvector), ClickHouse, Redis and MinIO, wired together with volumes and private networking.
Migrations and seeding run on the first deploy; set SESSION_SECRET and SECRETS_ENCRYPTION_KEY
(openssl rand -hex 32 each) when prompted, then open the frontend's domain and create your
workspace.

Prefer to wire it yourself, or deploying somewhere else? The manual walkthrough is
deploy/railway/README.md (every variable pre-written in
.env.railway.example), and the production-hardening runbook
— auth guards, backups, worker pool, post-deploy verification — is
guides/DEPLOY.md.

---

Send your first trace

pip install "tracely-ai[openai]"     # or [anthropic], [langchain], [all]

Initialize once at startup, then wrap a run — your normal provider calls are captured automatically,
with no span code:

import tracely_sdk as tracely

tracely.init(
    endpoint="http://localhost:8000",   # your Tracely API
    api_key="tracely_dev_key",          # an ingest key
    service_name="support-agent",
    env="prod",                         # prod | staging | ci | dev — the gating axis
    instrument="auto",                  # auto-detect openai / anthropic / google / mistral / langchain
)

with tracely.trace(agent="support-agent", conversation="conv-1", user="u_42"):
    client.chat.completions.create(
        model="gpt-4o",
        messages=[{"role": "user", "content": "Where is order ORD-4471?"}],
    )

That produces a GENERATION span with model, messages, tokens, latency, tool calls and cost.
Need spans for your own logic? @observe and the manual agent / tool / llm / retriever /
guardrail context managers are all there.

Any OTLP/HTTP exporter works too — point it at POST {endpoint}/v1/traces with
Authorization: Bearer tracely_dev_key. Tracely reads standard gen_ai.* / OpenInference attributes
plus first-class hints: tracely.agent.id (auto-registered), tracely.agent.version,
tracely.conversation.id / turn.* / step.*, tracely.observation.type, and tracely.env
(prod|staging|ci|dev — the gating axis).

Declare your agents and record your state

Two optional lines make a conversation self-describing. The agent catalog tells Tracely which
agents, tools, prompts and models the conversation has (not just which ones fired) — it fills the
Conversation Agents panel and is readable from judge prompts as @LIST_AGENT. State deltas
record what each step wrote to your shared state, folded into the Conversation State drawer and
the per-message State Δ column:

AGENTS = [{
    "name": "support",
    "description": "front-line agent; routes billing questions",
    "system_prompt": "You are the support agent for Acme…",   # free-form keys kept verbatim
    "model": "gpt-5.2",
    "tools": {"lookup_order": {"name": "lookup_order", "description": "order by id",
                               "parameters": {"type": "object", "properties": {"order_id": {"type": "string"}}}}},
}]

with tracely.trace(agent="support", conversation="conv-1", agents=AGENTS):
    ...
    tracely.set_state({"cart": cart, "last_action": "add_to_cart"})   # inside any span/@observe

LangGraph users get state for free (node outputs are captured as deltas automatically), and
non-Python services can push the catalog with POST /api/sessions/{conversation_id}/config.

Full instrumentation guide → doc.tracely-studio.xyz · sdk/README.md

---

Gate your PRs

A promoted production failure becomes a regression test that blocks the PR that reintroduces it.
All you need is your ingest key (it identifies your workspace) and your Tracely API URL. Three ways to
wire it, depending on how your CI can reach your agent.

Option A — let Tracely call your agent (no agent code in CI, any language)

Register your agent's HTTP endpoint once, author scenarios —
multi-turn conversations, or an adversarial goal a red-team model improvises against — and Tracely
drives them itself. Nothing to install, import or shim, so a TypeScript or Go service gates exactly
like a Python one.

      - uses: Jwuthri/Tracely/.github/actions/tracely-gate@master
        with:
          api:   https://tracely.your-co.dev
          key:   ${{ secrets.TRACELY_KEY }}
          # agent: planner,support-agent   ← a subset; omit it to gate EVERY agent with scenarios

Scenarios belong to an agent, so leaving agent blank gates each one in its own run and fails the job
if any of them fails — a new agent is covered the day someone writes its first scenario.

Option B — gate the traces your CI already emits

If your pipeline already runs your agent instrumented with tracely.env=ci, the gate matches those
traces to your promoted cases (by input) and returns PASS/FAIL.

# .github/workflows/tracely.yml
name: Tracely gate
on: pull_request
permissions:
  contents: read
  statuses: write          # post the blocking commit status
  pull-requests: write     # upsert the results comment
jobs:
  gate:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      # → your existing step(s) that run the agent and emit env=ci traces go here ←
      - uses: Jwuthri/Tracely/.github/actions/tracely-gate@master
        with:
          mode:  gate                          # grade the ci traces this workflow emitted
          agent: planner                       # which agent's promoted suite to run
          api:   https://tracely.your-co.dev   # your Tracely backend (TRACELY_API)
          key:   ${{ secrets.TRACELY_KEY }}    # your ingest key = your workspace

Option C — replay recorded cases against your code (hermetic, $0)

Re-runs your agent on each promoted case's recorded input, serving the recorded tool/LLM outputs as
fixtures — deterministic, offline, no API keys, no cost — then gates. Guarantees the exact failing
inputs are tested.

      - run: pip install tracely-ai
      - run: tracely replay planner --entrypoint my_pkg.agent:run    # a Python agent
        # …or any language:  tracely replay planner --cmd "node run.js"  (reads $TRACELY_INPUT)
        env:
          TRACELY_API: https://tracely.your-co.dev
          TRACELY_KEY: ${{ secrets.TRACELY_KEY }}

Hermetic replay requires your agent to route tool/model calls through the SDK's call_tool /
call_llm seam (see the SDK guide); add --live to make real calls instead.

Both commands auto-detect the PR/commit from the Actions context; web-url / TRACELY_WEB_URL is
optional and only builds the "view gate run" link in the PR comment.

ℹ️ The repo's own .github/workflows/tracely-gate.yml is
Tracely dogfooding itself — it replays the bundled weather_agent example, which is why it uses
an in-repo pip install ./sdk. Your integration is one of the three options above, not that file.

---

Drive it from your editor

Every backend serves an MCP endpoint at /mcp, so a coding agent
reads your traces and writes your evaluators without any glue code:

claude mcp add --transport http tracely http://localhost:8000/mcp \
  --header "Authorization: Bearer tracely_dev_key"

Then: "look at the last 20 traces, find what's failing, and add an evaluation column that catches
it." Eleven tools over traces, failure clusters, evaluators and trends — scoped to the key's
workspace, same as every other call. On hosted Tracely the endpoint is
https://api.tracely-studio.xyz/mcp. Docs

---

Teach your coding agent Tracely

MCP gives your agent your data. The Tracely skill gives it the know-how — how to instrument,
what to evaluate, and how to wire the gate — so "add Tracely to this agent" is one sentence instead
of a docs tab.

npx skills add https://github.com/Jwuthri/Tracely --skill tracely

Works with Claude Code, Cursor, Copilot, Antigravity and anything else the
skills CLI supports — add -g for a global install,
--agent '*' for every agent on the machine.

Knows the whole surface Automatic tracing — init(instrument="auto"), provider + framework extras, @observe, drop-ins, LangGraph, LiteLLM, agent SDKs, redaction Manual spans — every observation type, handoffs, RAG, state deltas, multimodal I/O Non-Python — the OTLP conventions to emit from TypeScript, Go or Ruby Evaluators — structural vs judge, levels, @VARIABLE templates, advisory verdicts, sampling CI gate — scenarios, red-team runs, hermetic replay, the GitHub Action Troubleshooting — symptom → cause → fix for the failures that look like success

And the traps that silently produce a useless workspace a missing conversation id turns one thread into twelve orphan rows a swallowed tool error is invisible to detection, clustering and the gate no flush() and a script loses its last spans a dropped traceparent makes the gate blind to what your agent did an adversarial scenario is inverted — goal achieved means the attack won

Prefer to read it yourself? It's plain Markdown: skills/tracely/.

---

Architecture

The write path deliberately mirrors Langfuse's proven design — reimplemented in Python, with agent
semantics promoted to first-class indexed columns (Langfuse keeps them as read-time strings):

SDK/OTLP → POST /v1/traces → S3 blob (durable FIRST) → Redis/Celery
  → worker: otel mapping → registry upsert → ClickHouse events
  → evaluate_run_task → scores + structural clustering

Layer	Tech	Where
Backend (API + domain)	FastAPI + Pydantic v2	backend/
Workers	Celery + Redis	workers/
Traces + scores (OLAP)	ClickHouse (ReplacingMergeTree)	backend/tracely/infrastructure/clickhouse/ddl
Registry (OLTP)	Postgres + pgvector + SQLAlchemy 2.0 + Alembic	backend/migrations
Queue / Blobs	Redis / MinIO·S3 (blob-first, source of truth)	—
Frontend	Next.js 15 (App Router) + Tailwind	frontend/
SDK + CI gate CLI	tracely-ai (OTel wrapper + tracely CLI)	sdk/
Tooling	uv workspace (Python) · pnpm (web)	—

One deliberate adaptation: ClickHouse server-side async_insert instead of an in-process write buffer
(Celery tasks don't share memory). Why

Repo map — each folder has its own detailed README

Folder	What's inside
backend/	The tracely package: FastAPI API + shared domain (OTLP mapping, ClickHouse/Postgres/S3, registry, evaluators, failure intelligence, regression, gate, auth, Celery tasks).
workers/	The deployable Celery worker runtime.
frontend/	The Next.js web app — trace explorer, clusters, cases, gates, trends, settings, auth.
sdk/	The Python SDK (instrument agents over OTLP, hermetic record-replay) + the tracely CI gate CLI.
docs/	The published SDK docs site (Nextra). make docs → :3002.
skills/	The Tracely agent skill — npx skills add https://github.com/Jwuthri/Tracely --skill tracely.
scripts/	Dev/demo helpers (raw-OTLP sender, one-command seed_demo.py, gate shim).
design/	The full design dossier — reverse-engineered Langfuse + every Tracely design decision.

---

What's shipped

The core trace → detect → cluster → regression → gate loop is end-to-end:

• Ingest — any OTLP/HTTP source, first-class agent semantics, blob-first durability.
• Evaluate — DB-backed evaluators as table columns: CRUD from the UI, run on every ingest.
  Multi-output LLM-as-judge (score / number / boolean / text / JSON with custom schema) at
  conversation / run / span granularity, in basic (context auto-injected) or advanced
  (@VARIABLE template prompts with live preview + autocomplete) mode. Batch and sequential
  execution, per-evaluator targeting (agent/env) + deterministic sampling to scope judge spend;
  advisory evaluators record a verdict without flipping the roll-up.
• Triage — structural + semantic failure clustering, creatable suggested-evaluator drafts, promote-to-case.
• Regression — hermetic fixture bundles, fail-to-pass contracts, CI replay.
• Gate — PR blocking via tracely simulate / replay / gate, GitHub status + comment.
• Insights — daily traces/failures/gate pass-rate Trends + per-agent cross-metric meta-analysis.
• Conversation intelligence — real-time rolling summary (per-turn memory backing the judge's
  @HISTORY) + a conversation-agents panel.
• Judge calibration — label judge verdicts against human review, get per-evaluator agreement, and
  catch an over-flagging judge before you let it gate a release.
• MCP — the API doubles as an MCP server (/mcp): a coding agent reads traces, inspects failure
  clusters and creates evaluators itself, authenticated by an ordinary ingest key.
• Auth — three modes: dev (open), local (email/password + invites, self-host), clerk (hosted).
  Team management, API keys, invitations, account settings.

Near-term plan: design/part2-tracely/11-prd-next-steps.md ·
Long-term roadmap: 10-mvp-and-roadmap.md

Key environment variables

Variable	Default	Purpose
AUTH_MODE	dev	dev (open, no login) · local (email/password, self-host) · clerk (hosted SaaS).
SESSION_SECRET	—	Required when AUTH_MODE=local: HS256 signing key for JWTs (≥32 chars).
CLERK_ISSUER	—	Required when AUTH_MODE=clerk.
OPENROUTER_API_KEY	—	Enables LLM-as-judge evaluators (any model). Skipped gracefully if absent.
OPENAI_API_KEY	—	Alternative LLM backend for judges + failure-intelligence embeddings.
TRACELY_BACKEND_PORT	8000	Backend host port (Docker compose override).
TRACELY_WEB_PORT	3001	Frontend host port (Docker compose override).

With no LLM key at all the pipeline still runs — judges, failure intelligence and meta-analysis
degrade rather than crash.

---

Contributing

Issues and PRs welcome. Before pushing:

uv run pytest -q backend/tests sdk/tests    # what CI runs
uv run ruff check . && uv run ruff format .
cd frontend && pnpm test && pnpm build      # vitest + tsc typecheck + lint

License

MIT © Julien Wuthrich