One config. Eight specialists. Any harness.
Define your agent workflow once in .weave files. Run it across OpenCode, Pi, or Claude Code without changing a line.
Six guarantees, not six features.
Every one of these is enforced by the workflow rather than left to the model to remember.
Plan-then-execute workflow
Explicit phases separate planning from implementation. Pattern researches, Weft validates, then Shuttle executes.
Mandatory review gates
Weft reviews code quality and Warp audits security before work is considered complete. No bypassing.
Intelligent token management
Agents delegate to sub-agents for focused work. Token tracking warns at 80% and recovers at 95%.
Interrupt and resume
Stop anytime. Progress persists as checkboxes. Resume and pick up from the first unchecked task.
DSL-first configuration
Override models per agent. Inject skills from markdown. Disable what you don't need. Full control, zero lock-in.
Same config, any harness
Write your .weave config once. Run it on OpenCode, OpenCode 2, or Claude Code. No vendor lock-in.
Ask a question, or run a plan.
Not every request needs every agent. Loom routes on context; an explicit plan follows a fixed lifecycle.
Conversational routing
- Thread / Spindle — Evidence-gathering: codebase exploration and external research
- Pattern — Planning: structured implementation plans with rationale
- Shuttle — Implementation: domain-specific code changes
- Weft — Review gate: code quality validation (conditional)
- Warp — Security gate: vulnerability auditing (conditional)
Explicit plan execution
- Pattern creates plan — Researches codebase, produces checkboxed tasks with rationale
- User starts execution — Explicit approval before any code is written
- Tapestry delegates tasks — Coordinates execution, hands tasks to Shuttle workers
- Optional review gates — Weft/Warp run if configured, not unconditionally
Explore → Plan → Implement → Review → Resume.
Eight specialists with one job each.
Narrow remits keep prompts short and context focused, which is what makes delegation worth the hop.
Loom
Orchestrator
Primary coordinator that routes requests to specialists based on context and complexity
Pattern
Strategic Planner
Creates structured execution plans from complex requirements
Tapestry
Plan Executor
Drives plan completion by sequencing and delegating tasks
Shuttle
Domain Specialist
Implements bounded work packages within a single domain
Thread
Codebase Explorer
Gathers evidence from the local codebase through targeted exploration
Spindle
External Researcher
Researches external documentation, APIs, and specifications
Weft
Code Reviewer
Reviews changes for quality, correctness, and standards compliance
Warp
Security Auditor
Audits code for security vulnerabilities and compliance issues
Pick your harness.
Then follow the setup steps. For the complete walkthrough, see thequickstart guide.
Add the Weave plugin to opencode.json
In your project or in ~/.config/opencode/. OpenCode downloads the plugin itself.
"plugin": ["@weaveio/weave-adapter-opencode@0.2.0"]Restart OpenCode
Loom, Shuttle and the other Weave agents are ready. No .weave config needed to start.
Add the Weave plugin to opencode.jsonc
Note the plural "plugins". OpenCode 2 downloads the plugin itself.
"plugins": ["@weaveio/weave-adapter-opencode2@0.2.0"]Start opencode2
New sessions start on Loom. No .weave config needed to start.
Install the Weave CLI
Requires Bun.
bun add --global @weaveio/weave-cli@0.2.0Generate the Claude Code plugins
weave compose --adapter claude-code --initStart Claude Code with them
claude --plugin-dir ./weave-bootstrap-plugin --plugin-dir ./.weave/plugins/claude-code
Ready to orchestrate?
MIT licensed. Free to use. You pay only for model API calls.