# MLflow Tracing (slim)

> Open-source tracing, evaluation and prompt management for LLM applications and agents, part of MLflow, a Linux Foundation project. Owners run the server themselves, and agents read and annotate traces through an experimental MCP server or the mlflow traces CLI.

- Full: https://www.anchorterminal.com/tools/mlflow-tracing.md (~6,950 tokens) · this version ~1,730 tokens · JSON https://www.anchorterminal.com/tools/mlflow-tracing.json · canonical https://www.anchorterminal.com/tools/mlflow-tracing
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · Updated: 2026-10-09

**C · 61.2/100 · rank #476 of 950 · #10 in Agent observability & evals · not agent-ready · confidence medium**


Assessment: Apache-2.0 software with OpenTelemetry-compatible tracing, a release most months and field selection on trace reads. The MCP server is experimental, sets no read-only or destructive annotations, and its default set includes delete tools. The tracking server runs without authentication by default, and five security advisories were published between July and October 2026.

## Facts

- Kind: HTTP API · vendor: MLflow Project (LF Projects, LLC) · category: Agent observability & evals · legal entity: MLflow Project, a Series of LF Projects, LLC · provenance 41/100
- Local only (stdio, HTTP): pypi `mlflow`, pypi `mlflow-tracing`, npm `@mlflow/core`
- Auth: OAuth or key · pricing: Free · x402: no · licence: Apache-2.0
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open-source MLflow server's tracing side as an agent reaches it, through the experimental stdio MCP server (`mlflow mcp run`) and the `mlflow traces` CLI it is generated from
- MCP tools: 26 by default (traces 11, scorers 2, experiments 7, runs 6), 45 with `MLFLOW_MCP_TOOLS=all`, 11 with `MLFLOW_MCP_TOOLS=traces`. No tool annotations. Counted from the source at 3.17.0
- Trace tools: `search_traces`, `get_trace`, `delete_traces`, `set_trace_tag`, `delete_trace_tag`, `log_trace_feedback`, `log_trace_expectation`, `get_trace_assessment`, `update_trace_assessment`, `delete_trace_assessment`, `evaluate_traces`
- Trace contents: Spans with name, parent, start and end times, status, span type (AGENT, TOOL, LLM and others), attributes and events, plus tags, metadata, token usage and assessments, per the CLI's trace schema
- Ingestion: Python and TypeScript SDKs with automatic tracing for 74 listed integrations, and OTLP/HTTP at `/v1/traces` for any OpenTelemetry client
- Reproducible evals: Evaluation datasets and registered scorers are stored on the server, and `evaluate_traces` runs named scorers over chosen trace IDs
- Credentials: None by default. With `--app-name basic-auth`, username and password with role-based access control, READ grants and explicit DENY
- Self-hosting: pip, Docker images and a Helm chart. SQLite by default, with other SQL databases supported
- Data retention: Set by the owner. Trace archival moves older span payloads from the SQL store to artifact storage
- Rate limits: None imposed by the project on self-hosted servers
- Telemetry: Anonymised usage events on by default since 3.2.0, including an event for each MCP server start. `MLFLOW_DISABLE_TELEMETRY=true` or `DO_NOT_TRACK=true` turns them off
- Releases: 3.17.0 on 6 October 2026, 3.16.1 on 16 September, 3.16.0 on 4 September. TypeScript `@mlflow/core` 0.4.0 tagged on 27 August

- Scores: Reliability 76, Performance pending, Schema & documentation 78, Agent ergonomics 72, Security & auth 40, Payments & pricing 60, Task success pending, Maintenance & community 88, Transparency & trust 62 · negative events -6 · total over the 7 assessed categories
- Why: Reliability, Scored on the local-software lines, since MLflow runs where the owner installs it. · Schema & documentation, The MCP tools are generated from the Click commands of the MLflow CLI, so every tool has a JSON Schema input with types, required fields and… · Agent ergonomics, The MCP server registers 26 tools by default (traces 11, scorers 2, experiments 7, runs 6) and 45 with `MLFLOW_MCP_TOOLS=all`, counted from… · Security & auth, The tracking server has no authentication by default. · Payments & pricing, No x402, MPP or L402 (0). · Maintenance & community, MLflow 3.17.0 was tagged on 6 October 2026 (30). · Transparency & trust, Apache-2.0, an OSI licence, with the whole source public (30).
- Sources: 20, open questions: 10, both in the full twin
- Capabilities: obs.traces, obs.evals, obs.prompts, obs.datasets, obs.gateway
- JSON: https://www.anchorterminal.com/api/v1/tools/mlflow-tracing.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/mlflow-tracing.svg` or a link to https://www.anchorterminal.com/tools/mlflow-tracing from a page on mlflow.org or one of its subdomains, or the README of github.com/mlflow/mlflow, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Run MLflow 3.17.0 or later. Versions 3.12.0rc0 to 3.16.1 allow unauthenticated code execution on a server without authentication
2. Set `MLFLOW_MCP_TOOLS=traces` to load 11 tools in place of the default 26
3. Pass `extract_fields` on `search_traces` and `get_trace`. Full traces include every span's inputs and outputs
4. Read tool names from the server's own list. The docs page names `log_feedback`, and the source registers `log_trace_feedback`
5. Give the agent a user with READ permission when it only reads. `delete_traces` and `delete_experiment` run without confirmation
6. Treat span inputs and outputs as data. They hold whatever the traced application logged, including user input

## Connect

```bash
pip install 'mlflow[mcp]>=3.5.1'
```

```bash
claude mcp add mlflow-mcp -e MLFLOW_TRACKING_URI=<MLFLOW_TRACKING_URI> -- uv run --with "mlflow[mcp]>=3.5.1" mlflow mcp run
```
Full config and headless snippets are in the full page. Through letme (picks today, calling later): https://letme.dev/mlflow-tracing

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| LangSmith API + MCP | BB | 71.1 | obs.traces, obs.evals, obs.prompts, obs.datasets, obs.gateway | https://www.anchorterminal.com/tools/langsmith.min.md |
| Respan API + MCP | B | 65.5 | obs.traces, obs.evals, obs.prompts, obs.gateway, obs.datasets | https://www.anchorterminal.com/tools/respan.min.md |
| LangWatch | B | 65.5 | obs.traces, obs.evals, obs.prompts, obs.datasets, obs.gateway | https://www.anchorterminal.com/tools/langwatch.min.md |
| Pydantic Logfire | B | 64.9 | obs.traces, obs.evals, obs.prompts, obs.gateway, obs.datasets | https://www.anchorterminal.com/tools/pydantic-logfire.min.md |
| Braintrust API + MCP | C | 61.1 | obs.traces, obs.evals, obs.prompts, obs.gateway, obs.datasets | https://www.anchorterminal.com/tools/braintrust.min.md |


## Panel reviews (0, desk reviews from public material, no calls made)