# Run the trace store yourself and keep it healthy

> Run the trace store yourself and keep it healthy: 0 companies fully do it, 8 with a limit, 22 carry the tag. Every listing in Agent observability & evals with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/agent-observability.self-host-retain/
- Markdown: https://www.anchorterminal.com/niches/agent-observability.self-host-retain/index.md (~1,750 tokens)
- Slim: https://www.anchorterminal.com/niches/agent-observability.self-host-retain/index.min.md (~330 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/agent-observability.self-host-retain/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

No company's docs show a full fit yet. The listings below do the job with a limit, or not at all.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 0 | 8 | 22 | 10 | 11 |

## What counts as done

The stack runs from the documented compose or Helm files, with disk growth and retention controls documented.

Measured, for our own tests: it ingests at the team's volume, and after 30 days the UI still loads a trace list in under 5 seconds.

How agents ask for it: self-host it with docker compose; keep traces in our region; disk is full after three months; upgrade broke the UI

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Arize Phoenix (Arize AI, Inc.)**: Docker, Compose and Helm with SQLite or Postgres; PHOENIX_DEFAULT_RETENTION_POLICY_DAYS and per-project retention policies since 9.0, infinite by default. Before that a user hit 44 GB and an unusable UI (issue 8420).
- **Helicone AI Gateway + MCP (Helicone, Inc.)**: An all-in-one Docker image and Kubernetes guide exist, with S3, Better Auth and the ClickHouse and Postgres stack from the listing. No retention or disk-growth setting appears on the Docker page.
- **Laminar API + MCP (LMNR AI, Inc.)**: Compose runs frontend, app server, Postgres, ClickHouse and Quickwit; Helm adds RabbitMQ, Redis and a consumer. Signals and alerts need a licence key, no retention setting is documented, and ClickHouse system logs grew without bound on a small instance (issue 2176).
- **Langfuse API + MCP (Langfuse GmbH (a ClickHouse, Inc. company))**: Compose and Helm with Postgres, ClickHouse, Redis and S3; retention is per project with nightly deletes and data is kept indefinitely by default. One user grew a terabyte ClickHouse on a laptop and another lost the UI after an image upgrade (issue 12172).
- **LangWatch (Reasoning Engine B.V. (doing business as LangWatch))**: Apache 2.0 self-hosting with Docker or Helm is claimed and retention is configurable per organisation, team or project on paid and self-hosted plans, with 49 days on Free. The component list and disk guidance were not in the pages read.
- **MLflow Tracing (MLflow Project, a Series of LF Projects, LLC)**: SQLite by default, a Compose bundle with Postgres and RustFS, and a Helm chart; trace archival moves old span payloads to artifact storage under a retention policy. Authentication is off unless basic-auth is enabled. An HN user calls it well established.
- **Opik API + MCP (Comet ML, Inc.)**: Local Compose or Kubernetes with ClickHouse, Zookeeper, MySQL, Redis and MinIO; a tiered-storage TTL moves trace data to cold S3 after a set number of weeks. The open-source build has no authentication.
- **Overmind (Overmind Technology Ltd (England and Wales, company 16585996))**: Compose with six Celery worker profiles under AGPL, but startup validation requires Modal, OpenRouter, AWS checkpoint and sandbox credentials even to run; no retention or disk guidance.

- not this job: Agent Tree (No legal entity found. The LICENSE, plugin.json and marketplace.json name "Mike Pollow, Digitalroots"): Nothing to host.
- not this job: agents-tree (No legal entity found. The LICENSE and pyproject.toml name Karel Blavka): Nothing to host.
- not this job: BotGauge (BotGauge AI Inc. (Delaware, United States)): Hosted only; no self-hosting pages.
- not this job: Braintrust API + MCP (Braintrust Data, Inc.): Only SaaS or BYOC, and BYOC needs the Enterprise plan with Braintrust operating the data plane; there is no documented Compose or Helm stack to run yourself.
- not this job: DeepEval (Confident AI, Inc.): Nothing to host; runs are written to local JSON or SQLite stores.
- not this job: Galileo API + MCP (Galileo Technologies, Inc.): On-prem and VPC deployment are Enterprise only per the pricing notes; no Compose or Helm files are documented.
- not this job: Gauge CLI (Gauge): Local mode runs in Docker with results in .gauge/gauge.sqlite; there is no trace store to host.
- not this job: HoneyHive (HoneyHive Inc.): Self-hosting is BYOC or BYOVPC on AWS, deployed by HoneyHive with Terraform and Helm for Enterprise; there is no stack to run from documented files.
- not this job: LangSmith API + MCP (LangChain, Inc.): Self-hosted LangSmith is an Enterprise add-on with a licence key; the ClickHouse or SmithDB, Postgres, Redis and blob stack is documented but not available to run otherwise. One builder made a local tool because LangSmith needs a cloud account.
- not this job: Navige (Navige Ltd): Hosted only.
- not this job: Prefactor (Prefactor Pty Ltd): Hosted only.
- not this job: Pydantic Logfire (Pydantic Services Inc.): Self-hosted Logfire is an Enterprise deployment; the Personal, Team and Growth plans are cloud only.
- not this job: Respan API + MCP (Keywords AI Inc. (d/b/a Respan)): The trace platform is hosted only; the open-source gateway is a proxy with a request log, not a trace store.
- not this job: W&B Weave (Weights and Biases, LLC): Multi-tenant or Dedicated Cloud only on the plans read; the trace server left the public repository in October 2026.

## What stops agents finishing

- self-hosting is Enterprise only
- ClickHouse or S3 growth with no retention setting
- upgrade path breaks the trace tables
- auth off by default on the self-hosted build
- cloud account required to see one's own traces

## Around it

Before: agent-observability.instrument-app. After: agent-observability.debug-failed-run.

## Where this comes from

Weight 10 of 100 in Agent observability & evals: 11 of 86 accounts: Langfuse self-hosters (ClickHouse stack, terabyte disk, upgrade breakage), Phoenix at 44 GB, Laminar ClickHouse logs, MLflow as the established self-hosted option, the Ask HN comparison thread, LangSmith lock-in and cloud-account complaints that led to a local-first alternative

As data: https://www.anchorterminal.com/api/v1/use-cases/agent-observability.self-host-retain.json, or the MCP tool get_use_case with id agent-observability.self-host-retain. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
