It evaporates
The request lived in a transcript. Tomorrow there is no task, no owner, and no next step.
lang_mark stack · briefing
A local coordinator for agent work that has to last: queued, reviewed, and resumable on the machine that owns the files.
The language makes agency explicit. The operator lands the work and verifies when it knows the work has been done.
The drift
A useful answer is the easy part. Production work crosses files, tools, other people, and a later morning. That is where agent applications can lose the thread.
The request lived in a transcript. Tomorrow there is no task, no owner, and no next step.
Three things are “important.” None has a wake time, a queue slot, or a place to look.
Mail, money, research, and code share a single unbounded turn. Nothing is narrow enough to review.
Prepared work and performed work collapse. The operator cannot tell which one happened.
What you are looking at
It takes a person’s intent and turns it into durable work: a queue entry, a project folder, a specialist when one is required, and a review item before anything leaves the machine.
It is a Docker kit you run locally. You bring the model key. Artifacts and settings stay in folders on that host. One local address covers conversation, tasks, and the queue.
The whole stack
Language and runtime. Agents, skills, molds, HTTP serve, fleet. Apache 2.0.
Local coordinator. Queue, wake, review, hub-and-spoke, bridges. Proprietary.
Editor, language service, and agent modes aimed at the operator or a local runtime.
Hosted chat. Identities and five roles: office, supply, people, sales, creative.
A mold is a branded agent configuration. The primitive ships. Distribution is in flight.
How the surfaces meet
Write and run agents against the operator URL or a local backend.
A role produces work. The operator is where that step lands.
Optional inbound channels. Only allowed senders reach the agent.
Other tools can call the same HTTP contract, including through MCP.
How a task moves
A person, a role, or a scheduled job becomes a concrete task.
The run narrows to the smallest safe sequence and the current files.
Approved capabilities read, write, run, search, schedule, or prepare a send.
A file, a log, a review item, or an API result remains after the turn.
How the operator runs
The hub stays the front door. Specialists stay narrow. A separate process starts only when persistent state or isolation requires it.
Scheduled jobs, deferred commitments, and items waiting on a person. Wake order is explicit.
A folder, a status, a queue, and a next step. Idle time proposes an advance. It does not invent a new project.
A specialist gets a question, an input, an output path, and a definition of done. The hub joins the results.
Sends, publishes, bookings, and destructive changes wait. A draft is a draft until a person releases it.
The boundary
The workspace, operator settings, and runtime state live in host folders you can back up. The model key is yours. Optional connectors stay off until you turn them on.
A draft, a queued action, and a completed external effect are different states. The console is there to show which one you are in, and the next safe move.
Ambiguity routes to a question or to review. A failed call stays a failure. Partial work reports the exact state and the next action.
What a later person can do. Open the artifact, the queue, and the review item. Reconstruct what happened without replaying the conversation. The hosted team surface is optional. The record remains on the operator.
Use · individuals
A solo builder points the IDE at the operator, keeps agents and skills beside the project, and lets the queue hold the next implementation step when they step away.
An artifact lands in a known drafts folder. Editing it does not publish it. Sharing it is a separate, reviewable step the next wake can resume.
Messages become jobs: new, in progress, draft ready, needs approval, sent, or blocked. “Draft ready” never looks like “sent.”
A briefing, a travel shortlist, or a comparison is saved with what was checked. The recommendation can wait. The booking cannot slip out with it.
Use · teams
TeamLead AI is the chat surface — the hosted front end you talk to. PIN pairing, seating, and identity are features of that surface: who is signed in, and which operator they are paired to.
You delegate orders in chat. The five roles — office, supply, people, sales, creative — are AI experts: each takes an order in its lane and produces. The operator is the one place that work lands, queues, and waits for review.
In our planning, Org Policy is the organization layer over this: one entitlement for hosted chat across seats, with member add-ons and budget presets — the upgrade path from a single Operator bundle to a multi-seat deployment, each seat still paired to its own local operator.
Scheduling, follow-ups, and internal packets that stay drafts until someone sends them.
Orders, outreach, and customer threads with a status a reviewer can trust.
Campaign drafts and assets with a handoff into review, not straight onto a channel.
Three pictures
Someone asks for a piece and leaves. The operator writes the draft, verifies the file, and leaves publish and share as later gated steps. The next wake resumes from the path.
“Refund the blue jacket” matches two orders. The run reads both, refuses to guess, and either asks or opens a review. It records what it checked.
Source, docs, and rollback each return a status, evidence, and a next action. One blocked result holds the package. A missing result is not a pass.
Where the product is going
Where we are going
Standing, plainly
Public runtime, tests, and the 1.1 line. This is the open core the proprietary surfaces sit on.
The coordinator: install path, queue, review, projects, and the bridges a buyer actually starts.
The control station and the hosted team surface are usable. They are younger than the runtime.
Pack catalog, on-chain publishing, and an external security review are still ahead of us.
Held on purpose. Unsupervised spend, irreversible public posts, and high-value payment claims stay behind review. The 1.1 kit is for operators who want the loop now and can tolerate a preview.
The invitation
Bring a model key. Keep the files. Let the queue and the review gate carry the week.
TeamLead pairs to one operator with a PIN. Identity and seating come with the surface; external effects still wait for a person.
Team, audit, design partners, and the distribution layer for reusable agent shapes.
The Operator bundle is where that work lands, waits, and leaves a record.