Every machine in one bus
A note lands on the other machine at once, with a receipt and presence. orch machines. The bus
Orchemax
Preview — early access by invitation.
Govern several AI agents on one repo, from the CLI you already opened
One local binary under the agent CLIs you already pay for. Your source stays on your machine.
Free tier, no credit card. See plans.
Not another agent TUI: the seat is your own Claude Code, Codex, Cursor, Copilot, OpenCode, Gemini or agy, run as a subprocess or over ACP. Capabilities follow your active plan. What ships under the seat.
New in Orchemax 0.8.0
One coordinator, and a successor
If the machine holding it goes quiet, the next takes over, supervised. The bus
Three ways to pay for models
Your vendor account, your keys, or the Orchemax gateway — per seat, with failover and a ceiling. The gateway
The rules reach any agent
Rules and lessons reach any seat at turn one, sealed; an action a lesson refused never runs. The stack
Nothing of yours is overwritten
Scratch stays in the workshop sandbox; the manifest cannot be deleted by a seat. Guards
Numbers you can read
Failovers, counters per area, what learning cost and saved: orch stats. Stats
The pain
“It wrote formatMoney again.”
Third copy in the repo, different name, same body.
“It read six files to answer one question.”
The context window is spent before the edit starts.
“I killed a worker and it took the repo with it.”
Its lock is still held, and the next agent waits.
“It said it was done.”
Sealed, and nothing ran to prove it.
“Two agents, one file. One won.”
The other one’s work is in the reflog, if that.
“Half my identifiers are in Spanish.”
The chat language leaked into the code.
Also: a stray .py or .sh anywhere, tests in the wrong folder, node_modules committed, a skill nobody applied, no one in the chair.
How Orchemax cures it
Know when a worker finishes or asks
A desktop alert with sound, Telegram on your phone, and the worker's question answered from there. Docs
A budget that holds
A daily token cap stops a dispatch before it starts, and a busy machine queues work instead of failing it. Docs
Bring your own agents
One agent plans, another writes, a third from another vendor reviews. No new IDE, no proxy. Docs
One writer per file, one copy per helper
A lock is taken before a shared edit, and orch guard dup refuses a clone before it merges. Docs
Read one function, not six files
code_snippet returns one body by line range, and trace_path walks its callers and callees. Docs
Agents coordinate instead of colliding
task_claim gives each ledger line one owner, and ask with to_project gets the answer from the project next door. Docs
Your workshop stays yours
orch setup binds the folder and seeds thin law + memory. Your apps and shared modules keep their shape.
northforge/ ← workspace (bound; not a billable project)
├── orch.yaml ← register: projects + shared by lang
├── CLAUDE.md · AGENTS.md ← workshop law (seed / you own it)
├── GATES.md · doctrine/ ← optional kit + your doctrine
│
├── .orch/ ← Orchemax runtime (not your product)
│ ├── orch.db · mcp.json · LAST_SESSION.md · MEMORY_BRIEF.md
│ ├── gates/ · agents/worktrees/ · logs/ · telemetry/
│ └── …
├── .workflow/ ← BOARD · RESUME · LEDGER (disk memory)
├── .agents/skills/ ← orch-chair / worker / recall / clarify
│
├── shared/ ← SHARED by language (locks + symbols)
│ ├── go/ ← Go modules every app may use
│ │ ├── core/ · features/ · libs/
│ │ └── SHARED.md ← layer law (+ orch:export)
│ ├── ts/ ← TypeScript bucket (same idea)
│ │ └── …
│ └── css/ · js/ ← asset langs when every app assumes them
│ ↑ billing + portal edit here with orch lock — not forked per app
│
└── apps/
├── billing/ ← PROJECT (own git; own Chair)
│ ├── CLAUDE.md · AGENTS.md · docs/TRACKER.md · .sandbox/
└── portal/ ← another PROJECT; same shared above
├── CLAUDE.md · AGENTS.md · docs/TRACKER.md · .sandbox/
Order: 1 workspace → 2 register projects → 3 shared by language (+ assets)
billing's agent can ask portal's Chair; the answer carries project + path + line. Several apps, shared by language. Folder names are yours. Seeds are a kit — your process wins. Docs
Day one
orch setupregisters the MCP server in every agent CLI it finds.- Open your usual seat and ask the Chair for the work in plain words.
- At the end of the week,
orch usageshows what each worker spent. - A way out:
orch workspace clean --dry-runlists what is not yours, andorch uninstallremoves Orchemax in two confirmed steps.
Only here
Checked October 2026 against ten orchestrators’ own docs (the matrix, Herdr). The only one we found that:
- fails the task and refuses the merge on a duplicate or rule violation, across five CLIs it does not own.
- runs the rules you write once at sealing, at your git hooks and in your IDE.
- keeps agent scratch files out of your repo, shell redirects included.
- has a different vendor review every write task and refuses the merge on a rejection.
- enforces a read-only planning run, then executes that same plan.
- gives each worker a short-lived key, revoked when its task seals.
One coordinator, every machine
You talk to one Chair. It assigns, asks and answers the Chairs on your other machines through the Orchemax relay, so no machine waits for you.
Three operating systems, one plan
Our own build runs on four machines: a Windows coordinator, a second Windows machine, a Linux server and a macOS machine. Each runs its own agents, gates and tests for its platform.
A different machine audits every merge
A change built on one machine is audited by another before it lands. In our own build the Linux machine refused the coordinator’s own change because it claimed results not yet measured on macOS.
Measured, not promised
Night of October 14, 2026 in our own build: 5 merge trains, 19 branches landed, 7 core items closed, each with an independent audit and a test on main.
Preview. Messages between machines need a linked account.
Measure it yourself
No savings percentage. Four read-outs ship with the binary; run them on your repo.
orch usage: tokens by session, worker and model, plan-billed seats included.orch usage --leaks: why tokens went, such as re-reads and repeated calls.orch usage --governance: shared-file conflicts per pair of agents.verify_reportand the benchmark: a verdict per session and a duplication test in CI.
Go deeper
Compare
Every cell sourced, and where each peer beats us. Read the comparison.
Product
Daily flow, the overlook and the long FAQ. See the product.
FREE → STARTER → PROFESSIONAL → TEAM
Free installs locally — no cloud account. Paid plans meter managed agents and registered projects. LLM spend is BYO: you pay Orchemax for governance only. Gateway vs seat login.
Free
$0/mo · capped
- 2 projects · 2 open Chairs · no workers (workers from Starter)
- The full Chair: gates, memory, lessons, start and close
- No SaaS overlook
Starter
$10/mo
- 5 projects · 2 open Chairs · 1 concurrent worker · 300 dispatches/month
- BYO providers via gateway
- Solo policy
Professional
$25/mo
- 15 projects · 3 open Chairs · 3 concurrent workers · 1,500 dispatches/month
- 2 workshops · 2 devices · personal overlook of your spend
- Telegram soft notify · webhooks
Team
$35/seat/mo · 3-seat minimum
- CTO spend deck: tokens, actors, governance
- 100 projects · 1 open Chair per seat · 10 concurrent workers pooled · 6,000 dispatches/month
- A company: 5 workshops · seats are people · allowlist · schedules · Assist approvals · machines follow seats · above 100 seats, contact us
A dispatch is one managed agent start (orch task run / gateway worker) — not every chat turn. Seats meter humans; devices are linked machines (Team: at least one per seat). What Team includes.
Install it under the CLI you already opened
Run orch setup, open your usual seat, and decide with your own numbers at the end of the week.