The web companion to the CLI. Your team's traces, projects, benches, datasets, and sealed replays become browsable, with the evidence behind every claim attached. This is the real prototype, running in your browser. Drive the whole thing below, or scroll for a page-by-page tour.
interactive prototype
The live Hub preview is built for desktop. Come back on a larger screen to drive it — for now, here's what each view does.
Get early access for you and your team, a 15-minute call with the founder to get you set up, plus unlimited trace intelligence credits while the Hub is in preview.
A live look at what we're building into the Hub, from the trace ledger to the bench that gates your pull requests. Every panel below is the real view, booted on its own, not a screenshot.
The Hub opens on activity, spend, and the ledger it all stands on: a year of runs as a contribution heatmap, token distribution across projects and models, and the traces underneath. Click a day in the heatmap and the ledger filters to it.
The full ledger is built for recall. Scrub the day timeline to pin what happened on a given day, narrow by project, dataset or harness, then page through the whole history.
Replay any run as a conversation: prompts, responses, reasoning, and every tool call inline, with git commits surfaced as checkpoints in the stream. A privacy panel shows what was filtered out before you ever saw it.
The same run, read differently: lanes for user, plan, think, read, exec and write, beside working, local and remote git columns and per-step tokens, duration and context fill. Context-waste and run-health signals sit in the gutter, and the panel says whether the numbers came off the wire or off the record.
A long session is the sum of several trajectories. The Hub cuts it into usable units with a small library of slicers: two deterministic grains cut at capture, two labeled by a small model running locally. Every cut is a gap-free tiling of the same steps.
Every project opens on its own pulse: sessions this week, tokens and cost for the month, the session in flight, and a four-week session map with one lane per harness. A flow diagram then follows those sessions to where they ended up, in datasets, in pull request trails, or retained in the bucket.
Open a pull request and the bench's change gate sits above the review: the evidence pack behind it, the guarantees the change touches, and any touched surface that maps to no guarantee, named as surface-drift instead of passing quietly. Below it, the agent trail that built the branch, commit by commit, with intents satisfied, cost, steps and wall time.
The atlas is one row per product guarantee, written in product language: the scenario bound to it, the verifier that judges it, the last verdict, and how fresh that verdict is. Every hole is a named state, unbound, stale-run, stale-verifier, no-red-proof, rather than a blank cell.
Each run pairs a scenario with a specific verifier version, executes it on a leased box, and freezes one record. Skips are listed by name, and a green whose verifier was never shown failing carries red_proof_ref: null instead of being counted as proof.
Every frozen run record renders twice: a page for people, with a scrubbable terminal cast, the red proof, the scorecard and the named skips, and the same record as a JSON feed for agents. When no cast exists the record says rewatchable: false rather than faking one.
One table for everything projected out of an application and its agentic work. Datasets are static, rows materialized by a workflow; arenas are stateful, where a bench plays a scenario, a capsule plays a trace, and a gym plays a task family. A bench lives in its project, so its row links back home.
A dataset is local first. Its workflow drops candidate rows into an inbox, you promote the ones worth keeping to staging, and the batch goes to a Hugging Face remote when you say so. Each candidate carries the slice of the source trace it was cut from.
A capsule is one failing episode, sealed. Anyone can witness the film straight from the URL with no account, or step through the sealed conversation and trail. The badge is the honest part: today every capsule refuses the reproducible claim and reports floor trust instead of implying a live verdict.
Ask anything in plain english. Spotlight reads bodies, reasoning and tool calls with an LLM grader, then returns ranked traces with a score, a snippet, and the run's fingerprint.
Boxes bill by the hour per environment and AI credits are spent per model call. Two meters track the month against its allowance, with every lease and model call itemized underneath and daily lease hours stacked by environment, model spend drawn over them.
One page for the whole setup: the Hugging Face identity everything runs under, every machine mirroring its local bucket into a single private remote with its own digest and sync state, and the scrubbing pipeline that has to pass before anything syncs.
Get early access for you and your team, a 15-minute call with the founder to get you set up, plus unlimited trace intelligence credits while the Hub is in preview.