orch core

Let coding agents do the work. Keep every decision yours.

orch core runs agent work through tickets on a fixed path. Agents write the requirements, the plan and the evidence. You sign the requirements, the plan and the verdict in a physical way, with your fingerprint or your phone. An agent can type any command, but it can’t put a finger on the sensor. It works with Claude Code, Codex and any agent that has a shell.

View on GitHub
DEMO-0043 Load tariff tables as dbt seeds Agent at work
  1. agent
    Ticket, in the requester’s words
  2. you sign
    Requirements and acceptance criteria
  3. agent
    Plan as a task list
  4. you sign
    Plan approved
  5. agent
    Code, tests, evidence per criterion
  6. you sign
    Verdict

Agents work between the gates and stop at each one. Only a person can open a gate.

Mission Control, Today: connections to fix, a blocking question with a recommended answer, three approvals and a verdict that wait for you, and which agents are working
Mission Control. Everything that waits for you, blocking items first.

Four problems I kept running into

Different laptops per client. I’d come back after a week or two and have forgotten where things stood, what was next and who still needed to be asked, because the agents had done the work, not me.

I lost control.

Too much at once. What did I do last week?

Open tasks, questions, last statusToday shows what waits for you. Every ticket has a current-state note: where it stands, what’s next.

Not what we agreed.

We plan together, the agent builds something else.

Signed requirements and planTwo gates before any code. You sign the exact text, and a change voids the signature.

“Done” isn’t done.

“Finished and perfect”, and half of it is missing.

A checklist with evidenceEvery acceptance criterion needs proof before the ticket reaches you. Your verdict signs one commit.

Same discussion again.

Already decided, asked again.

Ticket history and contextEvery answer and decision is a signed event on the ticket, and agents read it before they ask.

Everything flows into the core

A request comes in, in the requester’s own words. The agent gets the keys and knowledge it needs, follows your skills and rules, and writes every step into the ticket. People answer questions and sign the gates. The results go back to your tracker, docs, code host and data platform, and every finished ticket adds to what the next one can draw on.

orch core: agents work, people sign Request, keys and knowledge feed orch core, together with skills and rules. The workflow is Understand, You sign, Plan, You sign, Build, Prove, You sign. Every step is written into the ticket. Results return to the tracker, docs, code host and data platform. Every finished ticket feeds the knowledge base. People answer questions and physically sign every gate; no agent can sign for them. Skills mine · the team’s · the company’s Rules conventions · locks · the fence Request in the requester’s own words, from a tracker, chat or email Tools & Keys access per client, least privilege Knowledge wiki, code, earlier tickets orch core Every step is written into the ticket. Understand You sign Plan You sign Build Prove You sign The agent: Claude Code · Codex · any CLI agent WRITTEN BACK Tracker status and summary Docs documentation updated Code host change ready for review Data platform new table, checked People answer questions and sign every gate in a physical way, so no agent can sign for them. Fingerprint or phone. Every finished ticket feeds the knowledge base
a person signswork and contextknowledge feedback

A gate before every next step, signed by a person

Every ticket follows the same path. Agents work between the gates and stop at each one. A gate opens only with a person’s signature, and that signature takes a physical act, so it can’t come from an agent.

Ticket, then gate 1 requirements and acceptance criteria (you sign), plan, gate 2 plan approved (you sign), build and prove, gate 3 verdict (you sign).

Your fingerprint

Sign at your desk with the fingerprint sensor on your laptop.

Your phone

Sign from anywhere with Face ID. The key never leaves the phone’s secure chip.

Next: a hardware key

The same idea works with any signer that needs a physical act, such as a security key or the TPM chip in a company laptop.

What you sign is bound to the exact text you saw. If the plan changes after you signed it, the signature no longer counts.

A checklist, and evidence for every item

Agents like to report success. orch core doesn’t take a summary for an answer: the acceptance criteria are agreed before the work starts, and the agent ticks each one off only with evidence. Your verdict unlocks when every criterion is proven.

Left: an agent says 'Done, all tests green, ready to ship'. Right: the ticket's checklist of acceptance criteria, each with its evidence: row counts, a dbt test run, a reconciliation with zero difference, and one criterion still without evidence. The verdict unlocks only when every criterion is proven.

At your desk

Mission Control is a local dashboard. Every prompt that signs something shows you exactly what you sign, and only core can show it, never an addon.

A ticket in progress: requirements and plan approved, verify pending, the current-state handoff note and a question widget that compares DATE with TIMESTAMP
The ticket. Gates at a glance, the handoff note, the open question with the agent’s measurement, and evidence and history one tab away.
Approve plan: the prompt shows the task list exactly as signed
Approve the plan. The task list exactly as you sign it, with the command each task must pass.
Give verdict: acceptance criteria evidenced, the exact commit the verdict signs, pass or send back
Give the verdict. It signs one commit. A new commit on the branch voids it and sends the ticket back to testing.
The board: tickets grouped by epic across backlog, open, in progress, waiting and testing, with GitHub issues alongside
The board. Epics as swimlanes. Addons such as GitHub issues add their own lane.

On your phone

orch mobile answers questions and signs gates from an iPhone. One QR scan joins a workspace, and logging out wipes it. Every signature needs Face ID, and the device key stays in the Secure Enclave.

Home: your workspaces, what needs you, and Drop
Your workspaces
A workspace's tickets with their status
Tickets and epics
A question from an agent with its answer options
An agent’s question
The sign sheet: what you are signing, then Sign and send
Sign and send

The screens show example data.

AI Factory and Dark AI Factory

For bigger requests, sign a charter instead of every single gate. You set the limits once: how many tickets, how big each one may be, the budget, and whether the result stays a preview or gets delivered. Inside those limits the agents split the work into tickets and build it, and every approval the charter gives is marked as such.

You sign the charter with its limits. Inside the charter the agents split the request into tickets whose gates the charter approves, and produce a checked preview. Before delivery there is a hold with a Stop button, then delivery to the destinations named in the charter. AI Factory runs with the lights on, Dark AI Factory with the lights off.

How the pieces fit together

A workspace is one team’s or person’s space for a project or client. Agents work through the orch CLI, the workspace host holds the keys, and everything lives in plain, signed files. For a company, one orch hub connects all its workspaces and your phone. The relay and publish at the edge only ever see ciphertext.

orch technical overviewAgents call the CLI, which reaches the key-holding Workspace host through a socket. Mission Control and trusted addons attach to the workspace host. The workspace host records signed events in plain files. Workspaces dial out to the company hub; personal workspaces share selected items. The hub holds user keys and full-mode workspace content. It coordinates roles, routing, approvals and shared spaces. The iPhone supplies fresh channel proof through a ciphertext-only relay, which also connects hubs for shared objects. The hub forwards workspace-signed, human-approved releases to publish. Standalone core needs no hub. WORKSPACES A WORKSPACE · one per project, client or team COMPANY NETWORK EDGE socket sealed envelopes hub ↔ hub HTTPS, signed Agents Claude Code, Codex or any agent with a shell orch CLI Agents' commands, terse output You Mission Control Channel signing Workspace host Runs the workspace, holds its keys, enforces gates, roles and policies orch serve · Keychain / Secure Enclave Addons Trusted per version Tickets & files Plain files; every change is a signed event MORE WORKSPACES Marketing full mode Legal full mode Personal selective join orch hub ONE PER COMPANY · LINUX Trusted operator · holds user keys Signs with fresh channel proof Full mode: hub can read everything Selective: only what you share Users + groups read / edit / admin Directory Find people and workspaces Router ws → ws handoffs Release gate Human approval Shared spaces Filtered workspace slices Browser home needs-you list + admin console Signed events · hash-chained audit orch mobile iPhone · pair once by QR Key in Secure Enclave Fresh tap to sign orch-relay Ciphertext only Phone push · hub ↔ hub No workspace keys Another hub Shared objects only No workspace access orch-publish Human-approved release · hub forwards, workspace signs Solo workspace: works on its own, no hub needed
holds a key / can signaddonend-to-end encryptedplain link
Workspace
A team’s or person’s space for one project or client: tickets, agents, skills, memory and its own keys. Plain files per ticket (ticket.json, body.md, events.jsonl, artifacts/); every change is an appended, signed event.
orch CLI
Agent-first commands with terse output: new, claim, task, ask, artifact, submit. Any agent with a shell can use it.
Workspace host
orch serve is the only process that holds the workspace keys. It enforces gates, roles and policies, and checks every signature.
Mission Control
The desktop dashboard, itself an addon. Other addons: GitHub reviews and issues, Databricks, wiki, terminals, schedules, quick tasks, or your own, each trusted per version.
orch hub
One per company, next to its workspaces. People sign in once, and admins give groups read, edit or admin access per workspace. The hub routes handoffs between workspaces, keeps shared spaces and a signed audit log, and serves the browser home. In full mode it is a trusted operator, like an identity provider: it holds each user’s signing key, and every signature still needs a fresh tap on the phone or a passkey.
Selective join
A personal workspace can join without giving the hub its key. It shares only what its owner picks, and only the owner can switch it to full mode.
Standalone
A workspace without a hub works on its own: tickets, agents, Mission Control and local signing. The hub is an addition, never a dependency.
orch-relay
Directory, relay, Drop and push. It stores ciphertext only, carries phone push and hub-to-hub traffic, and runs as a public instance or inside your company. Workspaces dial out to the hub; nothing dials into a workspace. orch-relay
orch mobile
The iPhone app for questions, approvals and verdicts, with push notifications and Drop for files. It pairs once with the hub by QR. orch-mobile
orch-publish
Shares and small apps over HTTPS. The hub forwards the request, the workspace signs it, and nothing is released without human approval.
Linked workspaces
Hand a ticket to another workspace (yours, a colleague’s or one on a client’s VM) and continue with the result. Between companies, hubs link through a relay with no direct access to each other’s workspaces. Incoming tickets are untrusted until you accept them.

Works with the agent you already use

Every command is defined once, and gates and signatures are enforced by orch itself, not by a hook in one particular agent.

Claude CodeCodex CLI and others
InstructionsAGENTS.orch.md and skillsAGENTS.orch.md, skills where supported
Gates and signaturesEnforced by orchEnforced by orch
Hand edits to ticket filesRefused early by a guard hookDetected by orch check, or prevented by running agents as their own OS user