Seeing what happened
Their tracing, dataset curation and eval UX are mature in a way ours is not. If the question you need answered is what did my agent do and did it get better, that is what they are for, and train.cloud does not replace them.
Compare
Most of the field improves the prompt.
Nearly everything train.cloud gets compared to operates on context: a better prompt, a bigger retrieval window, a smarter scaffold. All of it is thrown away when the session ends. train.cloud operates on weights — which is a different kind of thing, and the reason the matrix below mostly explains itself.
| Capability | train.cloudthe loop | Capture & eval cloudsLangSmith · Braintrust · Weave | Tuning platformsTogether · Fireworks · managed FT | RL toolkitsprime-rl · verifiers · verl | Agent harnessesClaude Code · Codex · Cursor |
|---|---|---|---|---|---|
| Kind of thing | A loop: capture → mine → train → deploy → switch | A dashboard over runs you already made | A service you hand a finished dataset to | Libraries you assemble and operate yourself | The agent doing the work — not a competitor |
| What your agent's work becomes | Weights. Mined into tasks, trained on, served back. | A trace you can search. It never becomes a model. | Whatever you turned it into first — they train what you hand them. | Weights, if you build and run the pipeline that gets it there. | Context, discarded when the session ends. |
| Who builds the training set | task-mine does — dedup, provenance, verifiers | — | You, before you arrive | You, plus an environment per task | — |
| What you own at the end | The weights. Export them, self-host, air-gap. | Your traces, and the dashboard over them | Varies — some export the adapter, some only serve it | Everything. It is your cluster. | Nothing to own — same base model tomorrow |
| Improves without you starting it | Yes — retrains as you ship | — | No — each run is a job you launch | No — you operate the loop | No — better prompts, same weights |
| Where raw traces live | Your machine. PII redacted at source, before upload. | Their cloud by default; several offer self-hosting | Wherever you put the dataset | Your infrastructure | The vendor's, per their terms |
| Runs offline | The capture half does — Border Collie needs no account | — | — | Yes | — |
The deciding row
The row that sorts the field is the second one: what your agent's work becomes once the task is over. Everything else on this page follows from that answer.
Honest boundaries
What each one is better at.
We are in private beta and Border Collie is not public yet. Against shipped tools, that matters — so here is where they win.
A dataset you already have
If your training data is already clean and labelled, a managed fine-tune is fewer moving parts than a loop, and it is available today. The loop earns its keep only when building the dataset is the expensive part — which, for agent traces, it usually is.
Track record and ecosystem
PrimeIntellect has trained frontier-scale models — INTELLECT-1 at 10B, INTELLECT-2 at 32B — and their Environments Hub is live with hundreds of community environments. Ours are per-tenant and unbuilt. Those are real advantages of a shipped project over a young one, and they stand until our own milestones retire them.
Doing the actual work
Claude Code, Codex, Cursor and Gemini are not on the other side of this table. Border Collie drives them through one adapter contract, and the model you train is served back to whichever one you already use. Nothing about your setup changes except the base URL.
Straight answers
One-liners, for deciding fast.
- train.cloud vs LangSmith / Braintrust / Weave
- They show you what your agents did. train.cloud turns what they did into a model. If your problem is visibility, use them — their eval and tracing UX is years ahead of ours, and we do not try to replace it.
- train.cloud vs Together / Fireworks fine-tuning
- They train the dataset you bring. The hard part was never the training job — it was building the dataset and knowing which examples were worth learning from. That is the part train.cloud automates. If you already have a clean dataset, they are simpler and cheaper than a loop.
- train.cloud vs prime-rl / verifiers
- The most complete open RL stack there is, and we adopt its structure openly. They have trained frontier-scale models; we are greenfield. The difference is that they are a toolkit you operate and train.cloud is a loop that runs — and that we target trusted central pooling rather than untrusted compute over the open internet.
- train.cloud vs verl
- verl is the reference RL implementation and our training tier is a Rust port of its algorithm registry, verified against it with golden files. If you want to run RLVR yourself on your own cluster, use verl. train.cloud is the thing that decides what to train on.
- train.cloud vs just using a bigger model
- A bigger model is the right answer until the cost per token or the latency stops working. A tuned small model gets cheaper and faster the more specific your work is; a frontier model does not learn your codebase between sessions.
- train.cloud vs RAG and context engineering
- Retrieval puts the right text in the window. It works, and it resets every session. Weights carry the behaviour across sessions without spending context on it. The two compose — this is not either/or.
- train.cloud vs Claude Code / Codex / Cursor
- Not a comparison. Those are the harnesses train.cloud drives and improves; you keep using the one you already use, and switch it onto your own model with one base-URL flip.