Four dials. Two interfaces. One artifact.
Taste gives companies four dials — task, codebase, models, budget — and gives candidates exactly two interfaces: an OpenAI- and Anthropic-compatible model gateway and a built-in git remote. Everything the platform knows comes from those two channels.
Everything a company configures per assessment.
Task
A real slice of work plus a definition of done. From the template library or your own repo and acceptance criteria.
Codebase
A repo snapshot cloned into every session. Scrubbed and scoped, sized for a single session of work.
Models
An allowlist of models behind simple aliases — eco, standard, premium — served by your configured model providers.
Budget
A token budget and a time window. Metered exactly, enforced in real time, surfaced to the candidate as a live meter.
The candidate's side is their own machine.
No sandbox to fight, no browser extension, no screen recording. The candidate's whole footprint is model traffic and git pushes — which is exactly what real work produces.
The model gateway
/v1/chat/completions and /v1/messages,
streamed, with tool use. Any agent works with a one-time session key. Anthropic
requests are converted for OpenAI-format backends and converted back — transparently.
Every request is metered to the exact token and written into the session transcript.
ANTHROPIC_BASE_URL=https://tasted.dev/s/{session}/v1
ANTHROPIC_AUTH_TOKEN=sk-taste-…
ANTHROPIC_MODEL=standard
The git remote
A push-only, per-session remote. History is immutable — force-pushes are rejected — and every pushed commit triggers a CI run on the platform. The commit stream is the session timeline; nothing a candidate ships can bypass it.
git clone https://tasted.dev/git/{session_id}
# …work, commit, push — like any normal day
git push origin main
The session record.
The commit timeline merged with the agent transcript and every CI result — immutable, evidence-linked, and replayable. This is what reviewers open, and what every score points back into.
Commit time-lapse
Watch the repo evolve at N× speed. Pause at any commit: its diff, its CI result, and what the agent was doing at that moment.
The red-to-green journey
CI history shows how they handled failure — and how long they were willing to live in red.
Budget-per-commit
Tokens spent per unit of shipped work. Where the budget went, and what it bought.
Transcript vs. commits
Tool activity from the transcript compared automatically against what was committed — drift between the two is a signal in itself.
Engineered like the infrastructure it is.
Protocol conversion, both directions
Anthropic-format requests are converted for OpenAI-format backends and back — tool calls, streaming, and usage accounting included. One session key speaks both dialects.
Exact metering, enforced live
Token usage is counted per request — streamed or not — and enforced against the budget in real time. GET /v1/budget tells the candidate exactly where they stand.
Secure by default
Ephemeral per-session credentials, encrypted provider keys, immutable git history, full audit trail. Built in Rust — no unsafe, pedantic-lint clean.
“The session replay shows the plan, the tradeoffs, the debugging, the recovery from dead ends. The PR shows the code. The budget curve shows the judgment. Together they give you what a two-hour conversation never will: what this person is like at a keyboard, solving your problem, under real constraints.” — The Taste team
Run a sample task against your own repo.
Set the four dials, send one invite, and open your first session record — in under a day.