Skip to content
AI Software Factory @orkastery/cli 0.6.10MIT

Conduct 20 roadmaps with the calm of conducting one.

Orkastery is the platform that conducts your coding agents in parallel threads, proves every delivery against the repository and only calls you when the decision is truly yours.

Talk to us

No server and no built-in LLM. It runs on your Claude Code and Codex subscriptions.

now · 20 roadmaps · 3 machines · 2 runtimes
0 running 0 waiting on you 0 already verified
Conduct from where you already work
Claude CodeCodexHermesOpenClaw
01 · Parallelism

One layer conducts all your runtimes, on all your machines.

You conduct from a single conversation. The ork CLI reserves the roadmap items, spreads the threads across machines, runtimes and accounts, isolates each one in its own worktree and brings everything back through a serialized merge.

You, from one conversation
#TAG in the request · hourly summary · batch a–d
Claude CodeCodexHermesOpenClaw
orchestration layer
ork
deterministic · no server · no LLM
reserves the roadmapspreads threadsisolates in worktreesrotates accountsverifies claimsserializes the merge
mac-studiolocal · macOS
claude-bgprofile personalactive
checkout-v2
onboarding
pricing
codexprofile teamactive
design-tokens
4 threads · 4 worktrees · leases ok
vps-01cloud · Ubuntu
claude-bgprofile teamactive
search-index
analytics
codexprofile personalactive
billing-api
crm-sync
4 threads · 4 worktrees · leases ok
vps-02cloud · Ubuntu
claude-bgprofile studioactive
docs-modes
i18n-en
status-page
codexprofile teamactive
auth-keys
4 threads · 4 worktrees · leases ok
main serialized merge, one at a time billing-api ✓ ls-remote2. pricing3. docs-modes
One roadmap, several machines, no item done twice.
$ ork roadmap pegar · each machine reserves the item before it starts
mac-studiovps-01vps-02 free
RM-031RM-032RM-033RM-034RM-035RM-036RM-037RM-038RM-039RM-040RM-041RM-042RM-043RM-044RM-045RM-046RM-047RM-048RM-049RM-050RM-051RM-052RM-053RM-054
One worktree per thread

Each thread writes only to its own worktree. Writing anywhere else is blocked by a lease, with a queue.

ork thread new
Serialized merge

One merge at a time. A push only counts when the remote confirms the SHA.

ork ship
Accounts that take turns

Quota exhausted or login lost: the same prompt carries on in the next profile. Subscriptions only; a paid key in the environment is cost.violation.

ork accounts add
Several machines, one roadmap

Before starting, each machine reserves the roadmap item. Two never take the same one.

ork roadmap pegar
02 · The problem

Agents got good. Conducting agents, not yet.

Point three agents at the same repository and you get three agents tripping over each other, plus a report saying everything went fine.

20terminal tabs open, seven products half done and none shipped.
02hthe runtime account runs out and the roadmap stops until you wake up.
∞times a day you turn into the status report of your own agents.
Without orkWith ork
The agent says the tests passork verify reruns the command on the real HEAD, and CI reruns it again on a clean runner
You ask for status several times a dayAn hourly summary says what is pending, urgent, blocking or critical
Every obvious decision turns into a questionThe obvious one arrives decided and recorded; you change it if you want
The account runs out and the roadmap stopsThe same prompt carries on in the next profile or on the fallback runtime
Two sessions step on the same fileOne worktree per thread, and leases with a queue
The documentation lies about the codeork docs verificar fails the PR when a page diverges
03 · Who it is for
Builders, vibecoders, founders

Many products, one person, no burnout.

Run roadmaps in parallel on your own subscriptions. Stop switching between twenty terminals and being the status report of your agents.

CTOs, startups, companies

Governance that does not depend on trusting the agent.

Independent CHECK on the exact SHA, proven push, decisions recorded with the why and the cost of changing. Since 0.6.9, Ork Mesh brings every builder’s work into one company network and requires the factory receipt in Git before the merge. Governed memory with OrkMind when the team needs it.

04 · Modes

How much autonomy do you give the factory? Write a #TAG.

The factory was built to work with a lot of autonomy. The #TAG in the request sets the points where it calls you to decide or approve, the HITL (human-in-the-loop) points. Outside them, it carries on by itself. Delivery verification is the same in all four modes.

HITL point: the factory waits for your decision or approvalblock the factory runs by itself
#Maestro · when to use it
A clear, quick solution with no heavy tradeoff: one HITL point to align the premises.
at the HITL points, you decide or approve
Premises.

The mode sets the autonomy. Verification is always the same.

05 · Your attention

The simple decisions, Orkastery makes for you. You decide only what matters.

Every question from ork arrives short and organized, with alternatives and a recommendation, so you can decide in seconds. That lightness is a design goal, and it also has people with ADHD in mind.

$ ork board 7 threads · updates by itself
ThreadWhere it runsPhaseState
checkout-v2 · migrate sessions
loja/web
mac-studio · claude-bg
CHECK
running
billing-api · batch invoices
fin/billing
vps-01 · codex
SHIP
running
onboarding · invite flow
app/mobile
mac-studio · claude-bg
PLAN
waiting on you
docs · modes guide
orkastery/docs
vps-02 · claude-bg
MASTER
running
search · incremental index
core/search
vps-01 · claude-bg
GO
GO-FIX
auth · key rotation
infra/auth
vps-02 · codex
GOAL
running
pricing · new page
site/www
mac-studio · codex
GO
running
ork
09:00 summary · BRT
3
pending
1
urgent
1
blocks a thread
0
critical
Can I send the 3 questions now?
Yes, nowAt 11:00
question 1 of 3 · onboarding · PLANa–d
How should the active sessions move to the new invite flow?
After the objective ones, up to 2 open questions, one at a time. A missed deadline waits or escalates, never approves.
01
Hourly summary

How many items are pending, urgent, blocking a thread or critical, and whether it can send them now.

02
Batched questions

Up to 5 objective ones at a time, with alternatives a–d and one recommended. Then, up to 2 open ones.

03
The obvious one comes decided

Recorded with the why, how to change it and the cost of changing now or later. You are informed.

04
A deadline never approves

A missed deadline waits or escalates. An answer only counts with the proof of the authenticated channel.

06 · While you sleep

The factory does not stop when you sleep.

You close the laptop at 23:00. This is what the ledger records until coffee. Nothing here needed you.

scroll the page · an example night in your time zone
23:04GO

Six threads open on three machines, each in its own worktree. None touches the files of another.

00:37CHECK ✕

checkout-v2 failed two tests against the baseline. It goes back as GO-FIX with a typed spec, without waking you.

01:52rotation

The quota of the claude-bg/personal profile ran out. The same prompt carries on in the claude-bg/team profile.

03:18SHIP ✓

billing-api in a serialized merge. The push only counted when git ls-remote matched the local SHA.

05:40decision

An obvious decision made and recorded: the why, how to change it and how much changing costs now or later.

07:00 · night summary
Good morning. 11 verified deliveries, 2 questions waiting for you, nothing critical.
07 · How it works

Six phases no agent skips.

Each phase has a contract. The result does not depend on believing the agent’s report. Tap a phase.

CHECK delivers
Verification against the baseline, plus independent CI.
Required to pass
A failed claim becomes a typed failure, never an optimistic report.
claims.failed · verify.regression
CHECK ✕Failed? The thread goes back as GO-FIX with a typed spec. Nobody needs to wake you for that.
08 · Proof

Every number has the command that proves it.

Reproducible commands check the work on the real tree. The baseline separates regression from earlier debt.

6,425
core tests, zero failures in CI
$ npm --prefix core run test:ci
24
behavior scenarios, all green
$ ork eval --so-cenarios
247
assertions across 21 skills and 116 cases
$ ork eval --so-skills
0
parity errors between docs and code
$ ork docs verificar
ork demo30 s · no account · no model
$ork demo
claimsoma() handles negative numbers
verifynpm test -- soma.spec.ts
✕ failedclaims.failed · 2 of 5 cases on HEAD a41f9c2
→GO-FIX with a typed spec
verifynpm test -- soma.spec.ts
✓ accepted5 of 5 cases on HEAD 7d02e1b
the agent’s report did not count. The command did.
A gap published as a gap.
does not exist
Stable local verify on a VM whose CPU is stolen by the hypervisor. On that machine, the reliable proof is CI.
planned
Account state shared across projects, on the roadmap as RM-040.

Start conducting today.

Install in a minute. Run ork demo and watch a false claim get rejected in 30 seconds.

Talk to us