Features
Everything in the TANK.
The foundation, today
Built and running. This is what you get on day one.
Treads
Streams organized around an objective, with a pinned Goal Header: who owns it, its status, its decision log. You always know what a conversation is for and whether it is done.
@tank does the work
Ask for the work where it was agreed. The agent reads the context, posts a plan, waits for approval, then opens a real pull request and reports the green build back — or builds the deck, the video, the doc.
Answer mode
Ask @tank a question and get an answer in seconds, grounded in your team's decisions and documents — no sandbox, no plan, no cards.
Neural Knowledge — agents that learn your code
Every repository TANK works in carries what its agents learned there, refreshed on its own as the code moves. Runs get cheaper and faster the longer TANK works in a codebase: 30% cheaper and twice as fast on our own web app, measured.
Neural Vault — search that gets it
Search that understands meaning, not just words, across messages, files and decisions. The Neural Sidebar brings the relevant past into the live discussion as it moves.
Home — what today asks of you
Open the app and know what today asks of you: what is waiting on you, which agent needs a decision, who is around, which threads moved.
Waiting on — the list that never scrolls away
Flag what you need from someone; it stops scrolling away. Everyone has one list of what is owed to them and what they owe. Answered or dismissed, in place.
Decisions that stay decided
Confirmed decisions become a ledger the whole team can argue with: what was decided, from which messages, by whom, and what replaced it. Survey proposes; people confirm.
Topo — the map of the conversation
The map of a conversation. Beside every Tread, a strip shows its shape — your mentions, what is waiting on you, decisions, files, deploys — so you jump to what matters instead of scrolling for it.
Thermal
Topo one level up: where the heat is across the whole workspace right now — which Treads are moving, what is stuck, what needs a decision.
TANK Blueprint — the design is the running app
A Figma and Lucid alternative inside your workspace. A frame on the board holds the real application — web or mobile, running, at whichever commit you choose — not a picture of one. Click an element in it, ask for a change to the design, the copy or the behaviour, and the agent finds that element in the source, writes the change and puts the result in a second live frame beside the first, so you compare two running apps rather than a comp and a hope. The board records the commit it was drawn from and says when the product moved past it. Clicking through a flow writes the test. Built in the open in thirty hours.
TANK Neural — TANK's own model, inside your workspace
Neural 1.0 is the model TANK trains and serves itself. In a Premium workspace it has its own section: chat with it the way you would anywhere, and it answers from the Treads and files you can read, plus TANK's own index of the web for anything that changes. Your conversations are private to you; the workspace sees only what they cost, on Spend and in Books.
TANK Guardrails — rules agents cannot talk their way around
Write a rule as a sentence, or give it a shape TANK enforces itself: an order (this lands before that starts), a decision a person makes first, a check that must pass, a pull request's shape, the paths a run may touch, the commands it may run. Shaped rules are refused at the action from TANK's own records, before a commit reaches GitHub; sentences are judged at every checkpoint by a separate model. Each guardrail is rated as you type it, a document can be read for its rules, and daily drills prove the shapes hold. Included in Premium.
TANK AgentOps — a swarm of agents, run like a team
An Ops team is a coordinator and specialists, each with a role, a model ladder and a tool policy, attached to any Tread. Address the team by name and the coordinator plans, hands pieces to specialists inside an envelope of dollars and turns, judges what comes back before it can call the work done, and reports in the thread. Hand-offs are threads people read and steer; a person can hold a seat with a deadline; two specialists can be asked to agree before a risky step; what the team learns reaches its next run. A formation board shows who is doing what, a replay attributes cost down the hand-off tree, a shift works a backlog overnight and briefs you in the morning, and Shield drills the team daily. Built in the open in a day.
TANK Atlas — pages that say when they stopped being true
A Notion alternative inside your workspace. Ask in a thread and the page exists, written from what was actually said and citing it. Every page records what it rests on and tells you when that moved, so a wiki nobody trusts stops being the normal outcome. Decisions are blocks rather than paragraphs, collected into a ledger you can ask; a block can carry this month's real cash instead of a screenshot of it; and the Neural Vault answers questions with the passage. Built in the open in an afternoon.
TANK Books — the money side, kept by the agent
A QuickBooks alternative inside your workspace. Say "invoice Acme for the September work" in a thread and the invoice exists, is sent, carries a card link and is chased until it is paid. Photograph a receipt and the expense books itself. Bank statements reconcile with a reason for every match, reports answer in a sentence, and an accountant gets a clean export. Built in the open in an afternoon.
Armor Mode — quiet, not absent
Do not disturb, done properly: ambient chatter collapses into short summaries per Tread. Only the critical breaks through.
Send when things are calm
Schedule a message for a time — or for when the Tread has gone quiet, so it lands when people can read it rather than mid-conversation.
…and the table stakes, done right
Threads, reactions, files with virus scanning, pins and bookmarks, huddles, custom emoji, drafts, notifications on every device, several accounts on one phone.
Where this goes.
Flare — the workspace's voice
Every workspace owns its accounts on X and Instagram; its agent runs them. Connect a handle or make one, say how it may post and in what voice, and ask in any thread: "@tank flare our last feature on X." Daily or weekly on its own if you like, with a go from you first if you like. The audience a product earns is the traffic it can call on later.
Auto-Pilot
Turn a thread into a workflow: triggers, steps, approvals. Conversation in, pipeline out.
Profiles that tell you something
What someone is working on, what is between you and them, when they are usually around, the decisions they made.
Agents that carry work across teams
A decision in one Tread starts the work in another, with the same approval and the same record. The organization moves as one thread.
Enterprise
Single sign-on, provisioning and retention policies.
Neural Knowledge
Agents that learn your code, and get cheaper every time they work in it.
Every repository TANK works in carries a layer of what its agents have learned there: where things live, which commands prove a change, which rules are enforced, what went wrong last time. It is refreshed on its own as the code moves and as TANK learns, so the hundredth run is cheaper than the tenth, and a repository TANK has never seen starts from what it learned everywhere else.
- 30%cheaper per changewith the layer against the same model without it
- 57%faster to a pull request67 minutes became 29
- ½the exploration before the first edit103 tool calls became 48
- 6 / 6pull requests deliveredevery task, every run
Measured 4 October 2026 on TANK's own repositories: the same tasks, from the same commits, once with the layer and once with the same model and no layer at all, costs read from the metering ledger. Across six re-run tasks the saving was 24%. Early numbers; the full tables are published with the method.
Since then: Measuring.
- Nothing to maintain. The layer is generated and refreshed by TANK. Hand edits are overwritten; the one file people write stays theirs.
- Verified before it ships. Every change is checked with the repository's own commands before a pull request opens, and the pull request says what was run.
- Yours to keep. The layer in your repository belongs to you and is readable by any tool. The service that writes it is TANK's.
- Free inside TANK. Included with agents in every TANK workspace. Available for repositories outside TANK soon, metered on refreshes.
@tank
Say it once. Watch it ship.
Mention @tank where the work was agreed. It plans in the open, waits for your approval, does the work in a sandbox, opens a real pull request and reports the green build back to the same thread. Everyone watching can steer. Nobody has to re-explain.
- sam
@tank the retry loop in
worker/dispatch.godrops jobs when NATS reconnects. Fix it and add a regression test. - tank agentPlan fix-nats-redelivery
- Reproduce with a reconnect test around
Subscribe - Ack after redelivery instead of before dispatch
- Add
TestDispatch_Reconnect - Open a PR against
main
- Reproduce with a reconnect test around
- tank gateApproval push branch + open PR?Approved by sam · 14:02
- tank runPull request #482 opened
fix(worker): ack after redelivery on NATS reconnect
+64 −12 · 2 files · sandbox run 0x1f3a - tank ciCI green
- lint ✓
- test ✓
- verify-service ✓
Plan first
The agent reads the repo and posts a plan card before it touches anything.
Approval gate
Nothing leaves the sandbox until a human approves in the thread. Edits and stops are one click.
Real pull request
Work happens in an isolated sandbox with a short-lived, least-privilege repo token. The PR is a normal PR.
CI reported back
Status flows into the same thread and the Tread's decision log, so the record of why lives with the code.
The workspace
One screen. One objective. No tab pile.
Navigation on the left, the Tread you are working in the middle, and a Neural Sidebar that keeps the right context and the right automations one glance away.
- Treads
- # platform
- # mobile
- # incident-4127 critical
- Automation
- Runs · 3 live
- Armor Mode
- On until 15:00
passkey_credentials, backfill nothing, gate withauth.passkeys. Ready for approval.api/auth/passkeys.goADR-014 · self-owned authauth/?Navigation
Treads, Automation and Armor Mode. Nothing else competes for attention.
Treads
A pinned Goal Header, a filtered focus feed, threaded action cards and one-click task triggers.
Neural Sidebar
Live context from the Neural Vault, Auto-Pilot suggestions and instant answers, tied to the conversation.
The math
Three numbers move. Everything else follows.
Whatever the team does, it is paying for the same three things — and paying in time. Take the seams out of the tools and the time comes back.
- 01 · Execution speed
Days between “yes” and “done” become minutes
“Agreement to execution takes too long.”
TANK connects the conversation directly to a proposed plan, an approval and the execution — fewer handoffs, nothing repeated.
- 02 · Accountability
Nobody asks “where are we?” again
“Who owns it? What was approved? What is holding things up?”
TANK keeps responsibility and status visible throughout the project, in the place the work is discussed.
- 03 · Retained knowledge
Your team’s memory becomes an asset
“We solved this before. Why are we starting over — or waiting for the one person who remembers?”
TANK makes previous decisions and relevant information available where the next decision is being made.
Deploys everywhere
On your desk. In your pocket. In a tab.
Same workspace everywhere. The desktop app adds what a tab can't: native notifications, a badge, links that open the app — and it keeps itself up to date.
macOS
Universal (Apple silicon + Intel)
Download for macOSWindows
64-bit installer
Download for WindowsLinux
AppImage · also .deb
Download for Linux.deb packageWeb
Nothing to install
app.tank.chatiPhone & Android
Every Tread, thread and approval in your pocket
Coming to the app stores
Built on
Open protocols, boring infrastructure.
No proprietary bus and no lock-in. The contracts and the SDK are public.
- Connect RPCtyped protobuf contracts, public
- NATS JetStreamevents and agent intake
- PostgreSQLpartitioned by workspace
- Temporaldurable agent runs and gates
- the agents' brain
FAQ
Questions engineers ask first.
What is a Tread?
A Tread is what TANK has instead of a channel: a stream organized around an objective. It carries a pinned Goal Header with assignees, a decision log and pipeline status, a focus feed that filters out noise by default, and action cards for the work happening inside it.
Is my code safe?
Agent runs execute in sandboxed pods that are torn down after each run. Repo access uses short-lived, least-privilege tokens scoped to the run through a GitHub App. Anything that changes the world, such as pushing a branch, opening a pull request or merging, waits at an approval gate in the thread.
How do I summon an agent?
Mention @tank or use /tank in any Tread or thread. There is no ticket queue: the conversation is the brief, and the people in it steer the run.
Does TANK replace GitHub or my CI?
No. Agents work in your existing repositories and open ordinary pull requests. Your CI stays where it is; TANK reports its result back into the thread.
Which model powers the agents?
TANK agents run on frontier models through a provider-agnostic gateway, so the model can change without the product changing. The runtime is TANK's own, built on Temporal and open contracts.
Do we still need our chat tool?
Yes — and the AI assistant and the tracker you bolted onto it. Channels, threads, DMs, files, huddles and notifications are all here; the difference is that the conversation connects straight to a plan, an approval and the finished work, and every decision stays where the next one is made.
Is there a desktop app?
Yes, for macOS, Windows and Linux. It is the same TANK with native notifications, a badge, links that open the app, and self-update. Download it above.
Is there a mobile app?
Yes, in testing now and coming to the app stores. iOS and Android apps share the same workspace, Treads and agent runs, so you can approve a plan from your phone.