One platform · every AI provider

Build the work.
Swap the model.

Build your agents, workflows, skills, and memory once — then point them at whatever model is best this week, no rewrites. One quiet workspace where chat, tools, data, and every model share the same context and the same keystrokes. Cloud or local.

Get started See it move Free · Google sign‑in · No credit card
memory · persistent
workflow · 3 steps
mcp · brave, fs, github
OpenAI· Anthropic / Claude· Google / Gemini· Local & self‑hosted· Mistral· DeepSeek· Groq· Together· vLLM· Self‑hosted· Bring your own endpoint· OpenAI· Anthropic / Claude· Google / Gemini· Local & self‑hosted· Mistral· DeepSeek· Groq· Together· vLLM· Self‑hosted· Bring your own endpoint·
01 — promote

Promote any chat
to a workflow.

Explore probabilistically. Ship deterministically. The path your model figured out in chat becomes a repeatable workflow — same answer, every time, at a fraction of the cost.

probabilistic · in chat
"Pull Q3 revenue, compare to forecast, draft the memo."
1sql.queryfinance_db · 14 rows
2python.transformgroupby segment
3llm.summarizeclaude · 2.1k tok
4doc.rendermemo.md
convert to workflow $0.84 · 18.4k tokens
Promote
one click
deterministic · as workflow
SQL
finance_db
PYTHON
transform
AGENT
summarize
frozen prompt
OUTPUT
memo.md
cost per re-run $0.04 · 0.8k tokens
21×
cheaper to re-run · 0.8k vs 18.4k tokens
0%
model variability · same answer, every time
schedulable · auditable · shareable with the team
02 — workflows

Workflows from the chats that worked.

The workflows you ship come out of the chats that worked. Go back and forth in chat — running code, calling tools, trying models — until you have the answer, then say "save this and run it every morning." Catalyst turns that working sequence into a workflow it can repeat exactly.

  • Written by AI, from chat. No blank canvas — the workflow falls out of the path your model already found.
  • Edit by asking. Reopen any saved workflow in chat to revise it; it updates in place when you save.
  • Schedule & email. Pick a time and time zone and it runs on its own — then emails you the result with files attached.
  • See every run. Watch each step as it runs, with full history after — click any step to see what went in, what came out, and any files.
See how workflows work →
workflow · cio stock analysis · multi‑model
INPUT
input
AGENT
research
AGENT
gpt5
AGENT
kimi
AGENT
gemini
AGENT
supermodel
PYTHON
build_pdf
TOOL
send_report
OUTPUT
output
cio stock analysis · multi‑model · completed · 12.4s · $0.08
03 — mini apps

Ask for an app. Get one.

Workflows automate the work; mini apps give that work a face. Describe the little tool you want — "an app to log my food and chart it" — and Catalyst builds the workflows behind it, writes the app, and opens it live in your chat. A real, working tool, generated on request, with no front-end to build.

  • You ask, it builds. No canvas, no template — describe it and the assistant writes the whole thing.
  • Live, not a mockup. The preview is the real app running on your real data. Use it on the spot.
  • A face over your work. Buttons run your workflows; tiles and charts read your data collections.
  • Walled in by design. Each app runs in a sealed sandbox and can only touch the workflows and data you grant it.
  • Publish & fork. Share an app to a public catalog anyone can browse, or fork one you like into your own account — your models, your data.
See how mini apps work →
catalyst · apps · food tracker
🍽 Food Tracker reads food_log
1,840calories today
2,0107-day average
this week
Chicken bowl850edit
Greek yogurt180edit
Apple95edit
What's inside

Many surfaces.
One loop.

Catalyst is a real workspace, not a thin shell over someone else's chatbot. Every surface — chat, workflows, mini apps, collections, agents, MCP, memory, data — shares the same models, the same context, the same keystrokes.

future‑proof

The model is
a dropdown.

A new state‑of‑the‑art model ships every few weeks. The price floor drops every few months. Build your agents, workflows, skills, and memory in Catalyst once — then point them at whatever's cheapest and smartest today, tomorrow, and the month after that. Build the work; swap the model.

2024 gpt‑4‑turbo $3.24/run
2025 claude sonnet 4 $1.40/run
2026 Q1 sonnet 4.6 $0.84/run
today open‑source 70B · local $0.06/run
next month whatever ships ?
same stock_research workflow · same output · 54× cheaper in 24 months · zero rewrites
04 — chat

A chat that can do things.

Switch between any model, watch it reason step by step, read images and PDFs, render math and diagrams, run code, generate images — and a toolbar that grows as you connect more tools.

  • Switch models mid‑chat. Start with Claude, finish with a local model. The conversation comes with you.
  • Dial up the thinking. Off → low → medium → high, whenever you want it.
  • Compare side by side. Ask several models the same thing and see the answers next to each other.
  • Share any chat. A read‑only link with a clean preview when you post it.
See how chat works →
catalyst · multi‑model · same prompt
claude‑sonnet‑4.71.2s
A concise summary of the Q3 numbers, framed around enterprise expansion and net retention.
gpt‑50.9s
Detail‑rich answer with a small table and footnotes; highlights deferred revenue drift.
open-source 70B · local3.4s
Conservative narrative with an actionable next‑step list for the CRO.
web.search sql.run py.exec img.generate memory.recall
05 — mcp tools

Plug in anything that speaks MCP.

Connect servers running locally or in the cloud — secure sign‑in is handled for you. Catalyst finds each server's tools automatically, and they work everywhere: chat, workflows, and agents.

Brave Searchweb · 14 tools
GitHubrepos · 32 tools
Linearissues · 11 tools
Stripepayments · 18 tools
Filesystemlocal · 6 tools
Notiondocs · 9 tools
Add a serverlocal · remote · hosted
06 — data

Your data, in every prompt.

Postgres, MySQL, Oracle, or anything on your private network. Drop live values straight into your queries, with results cached for speed. Never copy‑paste a spreadsheet again.

finance_db · postgres prod · 14ms · healthy
analytics · mysql prod · 22ms · healthy
crm_replica · postgres replica · 8ms · healthy
warehouse · oracle tailscale · 140ms · slow
vector store · local local · 2ms · healthy
SELECT * FROM revenue WHERE q = {{ quarter }} AND plan IN {{ plans }}
07 — collections

Your data, in one place.

A collection is a named bucket of your own data — no schema to design, no tables to set up. Workflows fill it, mini apps surface it, chat queries it. It's the shared layer that makes "log my food, then chart it" work end to end — one source of truth across a workflow, an app, and a conversation.

  • No setup. Name a bucket and start storing — the shape can grow with what you track.
  • Yours, not a workflow's. One workflow logs, another reports, an app shows it — same bucket.
  • Add or correct. Append as you go, update running totals in place, fix a wrong entry in chat.
  • Pull it any way. Read by recency, by date range, or by what's in the records.
See how collections work →
collection · food_log
food_log 214 documents
date2026-06-12 mealChicken bowl calories850
date2026-06-12 mealGreek yogurt calories180
date2026-06-11 mealPasta & salad calories720
workflow writes app reads chat edits
08 — vector memory

It remembers.

It captures facts, session summaries, and your preferences — then brings the relevant ones back exactly when they matter. It searches by meaning, runs locally or in the cloud, and nothing leaves your control.

storagelocal · nothing leaves
searchby meaning
recallinstant · most relevant
remembersfacts · sessions · preferences
memory.recall("Q4 priorities")
fact · global0.94

CRO prioritizes net retention over new logo growth through FY26.

summary · session #2140.88

Board agreed to defer mid‑market hiring until pipeline coverage exceeds 3.2×.

fact · session0.81

Enterprise expansion ARR concentrated in top 14 accounts; renewals Q1–Q2 FY26.

09 — agents

Autonomous, on a leash.

Set an agent loose to plan, call tools, and carry out the work. Step through every decision, pause at any point, or stop it instantly.

12:01:04plandecomposed into 7 steps
12:01:09tool · sql.run14 rows · 22ms
12:01:14tool · web.search5 results
12:01:21tool · py.execrunning…
draftawaiting py output
deliver · slackqueued
10 — skills

Package what works once.

Bundle instructions, tools, and prompts into named skills your AI can invoke on demand — across chat, workflows, and agents.

board-summary
2 tools · 1 data source · template
daily-standup
linear · github · slack · scheduled
code-review
filesystem · github diff · style guide
earnings-prep
postgres · web search · memo.md
+ new skill
describe it, run it, save it
The shape of the workspace

Everything connected,
nothing in your way.

workspace CATALYST · CORE Chat N MODELS Workflows VISUAL Agents AUTONOMOUS MCP TOOLS Memory VECTOR STORE Data SQL · FILES
12+
model providers, one API
200ms
to the first words on screen
0
data leaves your account by default
MCP tools you can plug in
01

Promote any chat to a workflow.

The chat that worked becomes a repeatable, scheduled workflow — same answer every time, at a fraction of the cost.

02

No model lock‑in.

Cheaper, smarter models drop monthly. Swap them like a config — your work stays.

03

Memory across sessions.

It remembers facts about you and your work, so your AI keeps getting sharper.

04

Extend with MCP.

Plug in tools, apps, and files — anything that speaks MCP, the open standard for connecting AI.

Ready when you are

One workspace.
Every model.

Models, agents, workflows, data, tools, and memory — in a single, quiet interface.

No credit card · Free forever for personal use · Self‑host docs →