Blocks – whole agent and AI surfaces, built from the components.
Pro Blocks are part of Oxyn Pro. Every preview here is free to open and try; Pro opens their code.
Style
Agents – Whole agent surfaces: a live session with a team of agents, mission control and its timeline, a run inspector, background coding agents, routines, MCP servers, a knowledge base, an eval suite, an agent builder and an agent gallery.
Sit in on a crew of agents mid-run: the chat between them fills the centre, their shared plan checks itself off on the right and a bar tracks the context window used so far. Choose one agent to spotlight its lines and steps; when one asks you something, the plan holds until you reply.
One overview for a whole fleet: a stacked event chart of the last eight hours, lead agents with their helpers, and a side queue of asks and failures, all narrowed by a single status switch.
The fleet overview laid out against time: one row per agent across the last half hour, the same event chart above and the same queue of asks beside it.
Debug a single run of several agents: a growing tree of who spawned whom, each card listing its model and tool calls as timing bars, totals across the top, and any call expandable in place to show its input and output or failure.
Hand coding jobs to agents that run off-screen: describe the change, choose repo, branch and model, then follow each job through queue, setup and its steps until a diff is waiting for you to turn into a pull request and merge.
Agents on a timer, listed like plain sentences: when each fires and how long until the next time, which apps it reads, which channel gets the result and which model writes it, plus a strip of recent outcomes. Trigger one early or expand it to change how often it runs.
Look after the Model Context Protocol connections behind an agent: a token budget bar split by server, health per connection with sign in and retry, per-tool switches with their cost, and a catalogue to plug in new ones.
Build and probe a search index for an assistant: documents go through four ingest stages with live counts, then a test query returns scored passages, highlights what matched and shows what a reranker would change.
Score a release candidate of a model-backed feature against the version in production: a grid of test cases fills as the judge works, the pass rate climbs beside last release's, and anything that got worse or better links to the exact case.
Shape a single assistant before shipping it: rename it in place, choose its model, write its brief, switch integrations on, attach reference files, try it in a sandbox chat, and push the result live.
Browse ready-made assistants for each department, search by job or integration, watch a short example of one at work and add it to a workspace without leaving the card while each integration connects in turn.
AI – AI surfaces for products: a KPI card that explains its own change, a blind model arena, a prompt studio, assistant memory, a catch-up digest, a feature launch card, a setup wizard, sales call insights and lead scoring.
A metric tile that can tell you why it moved: ask it and a short write-up appears word by word, the main causes line up with signed bars, numbered pins drop onto the trend line as each cause is mentioned, and suggested next questions swap the answer in place.
A blind taste test for language models: one prompt goes to two unnamed models, both answers arrive side by side at each model's own speed with first token time, speed and cost underneath, and only after you vote are the names and marks revealed; a leaderboard keeps the win rates.
You are a careful support assistant for {{product}}. Answer in a {{tone}} voice, keep it under 80 words, link one help article when it helps and never promise a refund without asking a teammate.
Edit, test and ship a prompt template in one card: variables show as coloured chips, fields fill them from saved test cases, Run streams a reply with latency, tokens and cost, and versions can be compared word by word and promoted to production with Undo.
A card listing the facts an assistant keeps about its user, one short line each: filter by group or search, edit or forget a memory in place with Undo, file the memory this chat just made, add your own and switch saved memories or chat history off.
Back from time off? Press one button and the card skims Slack, Linear, GitHub and Gmail over the chosen period, then lists what matters in four ranked groups; each line has its source, people and link, and marking lines as read clears them until you are caught up.
A what's-new card for an AI capability: a tiny animated preview sits over a three part walkthrough that autoplays until you take over, three benefits per step, a workspace switch only admins can turn on, and Try it now.
Switch on a support assistant in three stages: link your knowledge tools and see them import, set the model, voice and safety rules, then try a sample question and check the numbered sources behind the reply before launching.
A sales call in one card: talk time split by speaker, pricing, competitor and objection moments on a rail that play their quote with a seekable waveform, an AI summary that streams once, and next steps whose checking lowers the deal risk.
New sign-ups sorted by how well they fit, as judged by a model: every line carries a ring with the number, a Hot, Warm or Cold label and a short why; open one for its signals and weighted breakdown, and send it to HubSpot or Airtable.
Oxyn · 74 free components · 264 Pro components · 96 screens · 20 blocks · Dioxus and GPUI