OpenAI is shutting down its v1/prompts endpoint on November 30, 2026, and its migration guide tells you to move prompt content back into application code and version it with git. Meanwhile the rest of the category has never moved faster: Langfuse got acquired by ClickHouse, Braintrust raised an $80M Series B, Humanloop shut down completely, Helicone got frozen inside Mintlify, Promptfoo went to OpenAI, and Vellum pivoted to a consumer assistant. A lot of the tool lists still ranking on Google recommend products you literally cannot adopt anymore.
I wrote a much longer version of this on DevToolLab, with every code sample run locally before publishing: Best Prompt Management Tools for LLM Apps in 2026. This is the condensed field guide.
What prompt management buys you
A prompt management platform stores prompts outside your code, gives every edit an immutable version with a commit message, and lets your app fetch by name and label (production, staging) instead of redeploying. Promoting a change means repointing a label. Rolling back means repointing it again.
The real reason teams adopt one is the UI: a support lead or PM can fix the wording of a refusal message without opening a pull request. The counter-argument, which OpenAI itself now makes, is that your repo already has versioning, review and CI, and a prompt edit that goes live without a deploy is a production change with no PR attached.
The deciding question is simple: does anyone outside engineering need to edit these prompts? If no, keep them in the repo. If yes, keep reading.
Two architectures, one big decision
Nearly every tool is one of two shapes. Runtime registries (Langfuse, PromptLayer, Braintrust, PromptHub) serve prompts from the vendor's database at request time, with SDK-side caching so it is not a network hop per request. Edits go live without a deploy, but you now have a runtime dependency and need a fallback path.
Git-native tools (Promptfoo, Microsoft's POML markup language) keep prompts as files in your repo and add a test harness on top. Nothing ships without a merge, but a PM who wants to soften one sentence needs an engineer. MLflow's Prompt Registry sits in the middle: immutable git-style versions, but on a server you run yourself.
Picking the wrong shape hurts more than picking the wrong vendor.
The six tools worth evaluating
| Tool | Type | Open source | Paid entry |
|---|---|---|---|
| PromptLayer | Hosted registry + CMS | No | $49/mo |
| Langfuse | Registry + tracing + evals | MIT | $29/mo |
| Braintrust | Eval-first platform | No | $249/mo |
| PromptHub | Hosted registry, git-style | No | $12/mo |
| MLflow Prompt Registry | Self-hosted registry | Apache-2.0 | Free |
| Opik | Observability + prompts | Apache-2.0 | Usage-based |
PromptLayer
PromptLayer is built first for the person who will never open your repo. Each save creates a version with a commit message, release labels decide what production receives, and critical labels can require approval before an edit lands. Free tier covers 5 users and 2,500 requests a month; Pro is $49/month, Team jumps to $500/month, and self-hosting plus RBAC are enterprise-only. If your priority is eval rigor rather than collaborative editing, look at Braintrust instead.
Langfuse
Langfuse is the default pick for most engineering teams, and the January 2026 ClickHouse acquisition made it safer: the company committed to keeping the MIT license, self-hosting and the existing cloud endpoints. Around 32,000 GitHub stars, and prompt management sits next to tracing and evals, so you can tie a quality drop to the exact prompt version that produced it.
The SDK caches prompts in memory and accepts a fallback, which is what makes a runtime registry tolerable in production. I verified this against an unreachable host on langfuse 4.14.2:
prompt = langfuse.get_prompt(
"support-reply",
label="production",
max_retries=0,
fetch_timeout_seconds=2,
fallback="You are {{brand}} support. Answer {{question}} in under 80 words.",
)
print(prompt.compile(brand="Acme", question="Where is my order?"))
With no server running it logs two warnings and prints the compiled fallback. Self-hosting is free under MIT; cloud starts free (50k units/month) and Core is $29/month. Weak spots: the UI targets engineers rather than PMs, and SSO enforcement costs a $300/month add-on on top of the $199/month Pro plan.
Braintrust
Braintrust comes at the problem from evals: prompts are versioned, deployable objects, but the point is that changes get scored against datasets and CI gates before they earn the production environment. It raised $80M in February 2026 and lists Notion, Replit and Cloudflare as customers. loadPrompt compiles the prompt locally so you keep the model call; invoke runs it server-side and logs the trace automatically. Every tier has unlimited seats, which is rare. The catch: the free-to-Pro jump is $0 to $249/month, and if you will not write evals you are paying for the half you skip.
PromptHub
PromptHub puts a UI on the git mental model: branch a prompt, commit with messages, open the equivalent of a PR, merge. At $12/month for Pro (private prompts, 10k requests) it is by far the cheapest paid tier here. You bring your own provider API keys and pay token costs directly, there is no self-hosted option, and observability is thin compared to Langfuse or Braintrust.
MLflow Prompt Registry
If you already run MLflow for classical ML, you already own a prompt registry. It ships in MLflow 3.x under Apache-2.0 with immutable commit-based versions and mutable aliases. This ran clean on 3.15.0 against a local SQLite store:
import mlflow
mlflow.set_tracking_uri("sqlite:///mlflow.db")
v2 = mlflow.genai.register_prompt(
name="support-reply",
template="You are {{brand}} support. Answer {{question}} in under 80 words.",
commit_message="add refund guardrail",
)
mlflow.genai.set_prompt_alias("support-reply", alias="production", version=v2.version)
prompt = mlflow.genai.load_prompt("prompts:/support-reply@production")
print(prompt.format(brand="Acme", question="Where is my order?"))
One migration note: mlflow.set_prompt_alias now emits a FutureWarning; use the mlflow.genai namespace. There is no collaborative editor a PM would enjoy, but as a free registry for an engineering team it is solid.
Opik
Comet's Opik is the most generous open-source option: the Apache-2.0 self-hosted build has no feature gates, so evals, guardrails and the agent optimizer are all free. Its prompt library versions automatically, with one gotcha worth knowing: content, metadata and type changes create a new version, but renames and tag edits do not. Prompt management is a small slice of what is mostly an agent observability platform, so if prompts are your only problem it is a lot of software to run.
Honorable mentions
LangSmith has arguably the best prompt editor of the hosted tools, but per-seat pricing ($39/seat/month plus compute and storage units) fights its own cross-functional pitch. Agenta is still self-hostable for free but has repositioned around agents; read the changelog before committing. Portkey bundles prompts into its AI gateway, and Palo Alto Networks closed its acquisition of Portkey in May 2026 with no roadmap statement yet for the open-source gateway.
The 2026 shakeout, in one list
The fastest way to get burned right now is adopting a product that is already off the board:
- Humanloop: shut down September 2025. Anthropic hired the founders but did not buy the platform. It is gone.
- Helicone: acquired by Mintlify in March 2026, now maintenance-only.
- Promptfoo: acquired by OpenAI in March 2026. Stays open source, but the roadmap now belongs to OpenAI Frontier.
- Vellum: pivoted in May 2026 to a consumer AI assistant. No platform pricing published anymore.
- Portkey: now inside Palo Alto Networks. Langfuse: now inside ClickHouse, with public open-source commitments.
Six well-known names changed ownership or identity in about eighteen months. The filter I would apply: prefer tools you could fork and run yourself (Langfuse, MLflow, Opik), or tools with an obvious standalone business and fresh funding (PromptLayer, Braintrust). Be skeptical of everything in between.
A 30-minute trial that answers the real question
Pricing pages cannot tell you whether your team will actually use one of these. Running Langfuse locally can: docker compose up from the repo, create a project and keys at http://localhost:3000, pip install langfuse, register one prompt with a production label, and swap exactly one hardcoded prompt in your app for get_prompt() with a fallback. Then edit the prompt in the UI, move the label, and watch your running app pick it up without a restart. If that moment does not feel valuable, you have your answer for the cost of an afternoon.
The full guide on DevToolLab walks through all eight steps, including the fallback test most teams skip and how to link traces to prompt versions so you can compare quality across versions later.
Two small utilities that help during a migration: a YAML diff checker for comparing exported prompt configs between versions, and a JSON to JSONL converter for turning a JSON array of test cases into the format most eval harnesses expect.
Recommendations
- Non-engineers edit prompts: PromptLayer.
- One tool for prompts, traces and evals, with self-hosting: Langfuse.
- Prompt changes must pass evals before shipping: Braintrust, budgeted at $249/month.
- Already running MLflow: use its Prompt Registry before buying anything.
- One or two prompts and an all-engineer team: keep them in the repo. OpenAI's migration guide is right about that case.
Top comments (0)