$threadOS$

Built for agentic engineering teams

The Control Plane
for Agent Work

Six thread types. Structured sequences. Runtime surfaces. Gate approvals. Full audit. Local-first. One workbench for all of it.

Get Started See it work ↓

● Next.js workbench● TypeScript + Bun● Local-first

Without a control plane

Multi-agent work degrades into chaos

✗Terminal sprawl — five agent sessions, no shared state

✗Unclear dependencies — which step waits for which?

✗No approval gates — agents merge unchecked code

✗No runtime truth — what's running, finished, or failed?

✗No lineage — who changed what, and why?

ThreadOS

An explicit, inspectable operating model for agent work

Steps and dependencies. Gates and approvals. Runtime surfaces and runs. Six coordination patterns from simple sequential to long-autonomy with watchdogs. SAFE and POWER policy modes. A workbench and CLI that operate on the same underlying state.

Thread Types

Six coordination patterns

Every multi-agent workflow maps to one of these types. Pick a pattern, define steps, and ThreadOS handles execution, dependencies, and runtime state.

base

Base

Single sequential agent. One agent, one task. The simplest pattern.

Use cases

Code generation, document writing, one-off analysis

- id: write-docs
  type: base
  model: claude-code
  prompt_file: prompts/docs.md

The Workbench

One place for everything

The left rail exposes Sequence, Hierarchy, Lanes, Gates, Agents, Builders, and Packs. The canvas renders your thread surfaces. The inspector lets you deep-dive any node.

⊞

Sequence

Author threads, steps, gates, dependencies. Apply templates. Set thread types.

⊟

Hierarchy

Parent-child surface relationships. Spawn chains. Depth inspection.

≡

Lanes

Run-scoped execution order. Merge markers. Completion state per lane.

◈

Gates

Condition-aware approvals. Block, approve, cascade. Full audit trail.

◎

Agents

⊡

Inspector

Deep-dive any surface: run events, merge history, artifacts, annotations.

CLI

Same state, deterministic control

thread — init, step, dep, gate, run, stop, restart, group, fusion, mprocs, template, status.

thread — chained pipeline

▌

Under the hood

Built for structured agentic engineering

Surfaces & Runs

Threads materialise into runtime surfaces. Each holds runs with full event history — spawns, merges, gate decisions.

Chat Orchestration

Structured tool_use for sequence editing. Propose, validate, and apply actions without leaving the workbench.

SAFE & POWER Modes

SAFE mode requires confirmation for every action. POWER mode removes prompts but enforces allowlists, forbidden patterns, and concurrency limits.

AutoResearch

LLM-driven workflow optimisation. Analyses sequences. Suggests improvements with confidence + impact.

Policy & Audit

Local policy.yaml enforces command allowlists, cwd patterns, max fanout, forbidden regex. Every mutation is auditable.

CLI + API + UI

Same underlying state. thread CLI for scripting. Local API for integrations. Workbench for humans.

Who it's for

Founders & technical operators — who need visibility and control over agent execution

Staff & principal engineers — designing multi-step workflows instead of one-off prompts

Agentic engineers — who want a local-first workbench instead of terminal improvisation