Request access
How Decisive does AI

Decisive AI

Startups run on details — everything matters, and it all moves fast. Decisive puts AI on that problem: it reads your whole workspace, every turn, so speed doesn’t cost you the details.

Request early access See the architecture
Claude · Whole workspace as context · 21 typed tools · Runs in your browser · Open source
01 context

The model is rented.
The workspace is yours.

A model without context is a brilliant stranger — it can reason, but it doesn’t know your work. In Decisive everything lives on one surface, so the AI reads all of it.

01

Complete by design

No DMs, no channels, no seventh tab. Everything the team knows lives on one open surface.

02

Read whole, not retrieved

Every turn, the live workspace becomes one cached block Claude reads before it answers. No ranking, no guessing.

03

Gets better with use

The AI writes into the same workspace it reads. Tomorrow’s context includes everything it did today.

Most tools search for context or fetch it a piece at a time. Decisive does the simple thing: it reads everything.

Everything, read before it answers.

02 the loop

Read. Decide. Act.

Every turn, @AI gets a live snapshot of the whole workspace. Then it doesn’t just talk — it acts.

Your workspace

chat · tasks · docs · decisions · members

Context resolver

assembles a live snapshot — cached per turn

Claude

reads it all in a 1M-token window

Acts through tools

creates, triages, replies, decides, ships

Every action writes straight back into the workspace — in the open, where the next turn will read it.
An agent loop, not autocomplete

It doesn’t answer. It acts.

@AI works through real workspace tools — triaging tasks, drafting decisions, replying in chat, building pages. The loop runs in your browser; the server is just a relay.

  • Full workspace in context on every turn
  • Typed, non-destructive tools — no hallucinated or dangerous actions
  • Your workspace’s own Claude key, encrypted — the relay stores nothing
@AI · one turn

› @AI triage the new login reports

↳ read  list_recent_chat_messages · read_discussion · list_tasks

↳ act   create_task “Fix OAuth redirect loop” · P1

↳ act   send_chat_message “Grouped 3 reports into T-142.”

✓ done — full workspace in context

The 21 tools it actually has

straight from the registry — one tool per capability
Read
list_recent_chat_messageslist_tasksread_taskfind_tasklist_discussionsread_discussionfind_discussionlist_pagesread_pagelist_workspace_memberslist_sagaslist_labelslist_statuses
Act
create_taskupdate_taskadd_commentadd_reactionsend_chat_messagecreate_discussionbuild_pagenavigate

build_page is generative UI — a live, interactive page composed in your sidebar.

Web
web_search

Runs on Anthropic’s side, capped at five lookups a turn.

Custom slash commands run the same loop — each with its own model tier and context level.

03 everywhere

One intelligence.
Every surface.

The same teammate, with the same complete context, wherever the work happens.

Chat & threads

Mention @AI in chat or any thread. It answers in place — or just does the thing.

@AI · chat · comments · threads

Voice — ⌘O

Press ⌘O and talk. It listens, thinks and answers out loud — same context, same tools.

Deepgram → Claude → ElevenLabs

Huddles

Call it into a live huddle. It hears the room and can act on what was said.

on-demand · in-call · acts live

Task agents

Launch a coding agent from a task. It works in its own cloud branch and opens the PR.

Claude Code · 1 task = 1 branch · → PR

Build

Your dev server, live in the cloud. Claude Code edits while you watch — then ships a PR.

live cloud preview · then a PR

Your editor

Decisive is an MCP server. Connect Claude Code, Cursor or Codex — your context follows.

MCP · OAuth 2.1 · 13 tools

And some of it is ambient Long discussions summarize themselves, comments get fact-checked, conflicting decisions get flagged — nothing to invoke.
04 memory

Every action becomes context.

Everything the AI does — tasks it opens, replies it sends, pages it builds — lands in the open and becomes part of the next turn’s context. A teammate that remembers gets more useful every week.

read→decide→act→land in the open↺

05 headroom

Built for where
this is going.

Claude holds a million tokens at once — a lean team’s whole workspace fits many times over. No summarizing, no sampling; as the windows grow, Decisive just reads more.

Claude · context window
1,000,000
tokens — about 750,000 words held in mind at once.
~200K
tokens of your live workspace, assembled every turn
Your workspace Headroom to grow into →
Decisive live context (~200K) Typical retrieval (~a few K) 1M-token window

A lean team's entire workspace is a fraction of the window — so it all fits, with room to spare. Retrieval systems pass a sliver and hope it's the right one. Decisive loads the whole thing, then leaves 5× headroom for the day your workspace is bigger and the windows are larger still.

10K100K1M10M~4K100K200K1M10M+?20222023202420262030
Claude's context window, over time · log scale the bet, not a promise
How we made these comparisons ›
  • These compare architectures, not vendors. We’re contrasting how each approach gets context to the model — not running a head-to-head benchmark with win-rates we’d have to invent.
  • The hard numbers are real. 1,000,000 tokens is Claude’s published context window. Today Decisive assembles up to ~200K tokens of live workspace context per turn, cached between calls. The tool names on this page are the app’s actual registry.
  • The growth chart shows Claude’s window over time on a log scale; the dashed line is a projection, labelled as one — a bet, not a promise.
  • “No bullshit” is the rule. If we can’t stand behind a number, it isn’t on this page.

See it read your workspace.

Decisive is in alpha. Bring your GitHub repo and a Claude key — the AI shows up already knowing everything.

Request early access See pricing
Home · Pricing · Blog · Open app Built on GitHub + Claude · Decisive 2026