Bring your coding agents into every step of building software. Thraft keeps every decision, and why, as you go.
Freegoals, plans, and memory on your MacWorkspaceyour team and its asks, by invite
The missing planning layer for teams that build with agents
Requests arrive in chat, email, and meetings. Nobody can say which were taken, which were dropped, or why.
People settle things in calls, and agents settle them in sessions that scroll away. The next person, or the next agent, starts from zero.
An agent makes a reasonable call the plan never covered. You find out when something breaks.
A year later the code is still there. Why it works that way is gone.
Asks from anyone land in one pile. Next, guests will send them from their own agent. A product manager opens a case for each ask: the evidence, the assumptions, and a decision with its reason. A no is recorded too, with what would change it.
A goal says where you are going: the target, what done looks like, and the questions nobody can answer yet. A plan says how to build it: requirements, each with criteria you can check. Your agent drafts, you leave notes on the blocks, and each revision shows what changed since you last read it.
Agents build and write back to the plan over MCP. When the code disagrees with the plan, the agent leaves a build note with a proposed edit, and a person decides.
Each criterion is checked against the code. Decisions go to Memory by domain, and when a later one reverses them, you see both. In a workspace, the person who asked gets the answer, linked to the decision behind it.
On your own Mac, you start at step 2, with a goal, a plan, or an idea. In a workspace, asks come in from the rest of the company, and the whole team reads the same record.
Every agent turn runs on your Mac, on your own Claude Code sign-in and the plan you already pay for. Each person on a team runs their own agent on their own Mac. Thraft never runs a turn on a server, never runs one for anyone else, and never meters AI. Your projects stay on your Mac until you share one.
Finance finds refunds issued twice by hand every Friday, and wants that to stop. Scroll, and the window follows the ask from the person who raised it to the answer they get back.
Marta runs Finance. She sends her ask from the agent she already uses, as a guest of the workspace. Lena, the product manager, opens a case: the evidence, the assumptions, and a decision with its reason. Once it is settled, the case is drafted into a plan, and each assumption becomes a criterion.
Write the target and what done looks like. What nobody can answer yet stays as open questions. Each answer becomes an item, each item drafts a plan, and the goal's progress is those plans, not a status someone wrote.
Your agent drafts the plan: requirements, their criteria, and the questions it could not answer. You leave notes on the blocks they are about, and the next revision answers every note and marks what changed. An audit reads the plan against the code before anyone builds.
Agents build and write back over MCP. When the code disagrees with the plan, the agent leaves a note on the block, with a proposed edit and how sure it is. A person accepts it, and the plan changes on purpose.
When the work completes, each criterion is checked against the code. Its decisions go to Memory by domain, with where each one came from. When a later plan reverses one, Memory shows both.
The answer goes back to whoever asked, linked to the decision behind each part of it. An ask that is parked or declined says why, and what would change the answer.
A workspace puts your team, and the people who ask things of it, on the same record. Your team's work speaks for itself upward. Every role reads the same goals and plans, each person lands where their work is, and a role never hides anything.
In a workspace today: shared projects, the pile of asks, cases, and roles. Next: guests who send asks from their own agent and hear back through it.
Thraft is built for the gap between the people who ask and the team that builds. A team with that gap shows whether the record holds up when people other than its authors read it. That is the test Thraft has to pass, and we want to pass it with you.
Say how big your team is and who sends you requests when you join the waitlist.
Goals, plans, and memory are free on one Mac, forever. Workspaces open by invite, so join the waitlist and say how big your team is.
Know why you built it, not just how.
0041Refund reasonsopen2 settled · 3 open20m ago0037Flag double refundsticketed3 done · 1 in review5h ago0031Refund totals by dayticketed4 done · 2 in progress1d ago0036Export to the ledgersettled9 settled · 2 findings2d agoBO-312· your agentlow confidenceF2· round 1standsscripts/refunds.py sums by order. So today a partial refund and a full one read as one order refunded twice.D1Hear about refunds from Stripe's webhookmediumround 1R3A missed event is retried, not droppedhighround 113 findings · revised 2 blocks21 finding · answered with no change3nothing that blocks · the audit holdsI2Refund totals by day0031building · 4 of 6I3Flag double refunds0037building · 4 of 5I4Reconcile by charge·to draft · from Q3DraftI1Refund reasons0034completedRFD6 decisions1 reversed since you readfrom 0037, 0052, 0061D2Billing5h agoField kit1d agoCASE-44d agoCASE-51w ago