Mesh coordination
One Task can outlive one agent session.
A Mesh connects Tasks, people, computers, and repositories. One Mesh can link several repositories; a Task keeps its identity when an agent or computer changes. A provider chat is one execution conversation inside a Task.
AvailabilityAvailable with provider-specific limits
What works
Compact shared state
Tasks carry status, dependencies, resource claims, explicit decisions, and remaining work—not hidden reasoning.
Human attention
Product, security, destructive, unresolved conflict, and failed-handoff events reach Needs You on iPhone and Apple Watch.
Authenticated handoff
Claude Code ↔ Codex handoffs use a bounded capsule, phone authorization, a separate target worktree, and a receipt.
Current ownership
The Task route follows the current owner and never reopens a previous native execution after transfer.
Mesh Governance
Skills, MCP servers, and shell are allowed, asked, or denied in a Mesh—for a kind or one named capability. Each computer reports its applied revision and enforcement coverage.
Observed usage
Task and computer observations retain their source and coverage. Account quota, token usage, and sampled machine resources have different meanings and are not added into one cost figure.
Shared with people
A Mesh is shared by a one-time invite with a role—Viewer, Member, or Admin. The current controller checks messages, pauses, handoffs, and releases before they reach a computer.
Company accounts and repositories
A company account receives selected repository grants. A device also needs a separate Mesh invite and role before it can receive a Mesh's data or actions.
Computers of their own
A member adds an allowed Mac or PC to a Mesh. Its claims are visible to other permitted computers and a Task can be handed to it.
Linked repositories
A Mesh can bind several repositories. Separate Mesh spaces can share a repository without merging their access; Weavatrix relations provide code-dependency evidence.
Mesh memory
A person can record a Task-bound decision through Engine Memory, then correct it while the old record remains in audit history.
Full-screen code views
An available architecture report opens as a searchable graph. Health can open its observed repository code map as searchable code towers.
Pinned execution
A Mesh can pin new Tasks to one confirmed computer. The display name is not a routing key. Models come from that host's catalog. Offline, Mesh refuses or queues until a deadline.
Claims released by the person
A claim left behind by an agent that is gone is released from the Task screen. The computer answers, a refusal puts the claim back with its reason, and the release is written down so a late snapshot cannot undo it.
Reported at the end
A Task can be exported as a PDF or CSV with available observations and their source. Missing history and sampled resource values remain identified as such.
Honest limits
- A capsule carries committed facts, not files; uncommitted work blocks departure.
- GrantTap never pushes, fetches, merges, or resolves resource conflicts by itself.
- Cursor has scoped author attribution but no phase-one remote start. Grok Build has no trusted caller hook yet.
- A pin is not a secure VM. Isolation from a toggle is not claimed.
- Company repository grants mediate the owner's GrantTap forwarding; they do not change Git provider or filesystem ACLs. Previously delivered data cannot be recalled.
- MCP and skill requests do not yet carry a complete bundle to every host or prove native installation.
- Usage is observational. A strict spending cap and use-only secret broker are not available yet; agent process environment secrets are readable by that process.