{weave}
Weave
v0.7.3 · MIT License

Same agent. Same context.
Same blind spots.

Different roles. Different perspectives. Better results.

Weave adds structure to your AI coding workflow. One conversation with Loom — it plans before it codes, reviews before it ships, and audits for security. Automatically.

Get Started View on GitHub npm ↗
bash
$ echo '{"plugin":["@opencode_weave/weave"]}' > opencode.json $ opencode ✓ Weave loaded. Ready.
Built for the OpenCode ecosystem
{why_weave}

Why Multi-Agent?

Manual Workflow

  • ✗ Context-switching between tasks
  • ✗ No structured planning
  • ✗ Lost work on interruption
  • ✗ Single-threaded execution
  • ✗ No security review

Single Agent

  • ✓ AI-assisted coding
  • ✗ No specialization
  • ✗ No review or validation
  • ✗ Context window limits
  • ✗ No security auditing
Recommended

Weave

  • ✓ Structured plan → review → execute → audit
  • ✓ Plan → Review → Execute
  • ✓ Checkpoint-based resumption
  • ✓ Per-agent tool permissions
  • ✓ Security auditing built in
{features}

Features

Ship better code, faster — with agents that know their roles

🤖

Plan, Review, Execute, Audit

Talk to Loom. It creates a structured plan, validates it before coding begins, executes step-by-step, then automatically runs code quality and security reviews when done. No single agent juggling everything.

📋

Every Change Is Planned and Reviewed

Pattern creates a structured plan. Weft validates it before a single line of code is written. Tapestry executes step-by-step. No work is ever lost.

🧠

Teach Agents Your Standards

Injectable domain expertise via markdown skill files. Specialize agent behavior — from custom coding standards to fleet-orchestration for multi-session workflows.

📊

Never Blow Your Context Window

Automatic token tracking with warnings at 80% and recovery strategies at 95%. Agents delegate to sub-agents to stay within limits.

⚡

Parallel Work, No Conflicts

Multiple agents work simultaneously with configurable concurrency. Independent tasks execute in parallel while the orchestrator keeps everything coordinated.

🔒

Security Review on Every Change

Warp automatically audits code that touches auth, crypto, or secrets. Per-agent tool permissions ensure read-only agents can't write. Safety by design.

⚙️

Works Instantly, Customizes Deeply

Zero config out of the box. When you want control, override models, temperatures, skills, and tool access per agent via JSONC configuration.

🔄

Never Lose Work Mid-Task

Interrupt anytime. Resume with /start-work and pick up from the first unchecked task. Progress persisted in .weave/state.json.

{agents}

How Weave Works

You talk to Loom. It handles planning, review, execution, and security auditing automatically.

🧵

Loom

your entry point

You talk to Loom. It handles everything.

Loom is the agent you interact with for every request. It assesses complexity, orchestrates the workflow, and delegates to specialist agents behind the scenes — you never need to address them directly. Simple tasks are handled immediately; complex work goes through a full Plan → Review → Execute → Audit cycle.

Behind the scenes, Loom delegates to:

Stage 1 · Plan
📐

Pattern

subagent

Researches the codebase and produces a structured, checkboxed implementation plan. Uses Thread for codebase exploration and Spindle for external docs.

🔍 Thread·🌐 Spindle
Stage 2 · Review
✅

Weft

subagent

Validates the plan before a single line of code is written. Approval-biased — rejects only for true blocking issues. If rejected, Pattern revises.

Stage 3 · Execute
🪡

Tapestry

primary

Executes every task in the plan sequentially — writing code, running commands, verifying output, marking checkboxes. Resumes seamlessly from interruptions.

🚀 Shuttle for quick tasks
Stage 4 · Automatic
🔐

Weft + Warp

mandatory

After execution completes, Weft reviews code quality and Warp audits for security vulnerabilities — automatically, in parallel. This is a mandatory gate. If either flags issues, they're surfaced before the workflow completes.

↓ View full agent roster
🧵

Loom

primary

Main Orchestrator

Plans tasks, coordinates work, delegates to specialized agents. The central team lead.

🪡

Tapestry

primary

Execution Engine

Executes plans step-by-step. Marks checkboxes. Resumes seamlessly from interruptions.

🚀

Shuttle

all

Category Specialist

Domain-specific worker with full tool access, dispatched via the category system.

📐

Pattern

subagent

Strategic Planner

Researches the codebase and produces structured implementation plans in markdown.

🔍

Thread

subagent

Codebase Explorer

Fast, read-only navigation — grep, glob, read. Precise answers with file paths.

🌐

Spindle

subagent

External Researcher

Fetches documentation, reads APIs, synthesizes findings with source citations.

✅

Weft

subagent

Reviewer / Auditor

Validates plans and reviews work. Approval-biased — rejects only for blocking issues.

🔐

Warp

subagent

Security Auditor

Audits for security vulnerabilities. Skeptical bias — rejects by default on security patterns.

Tool Permission Matrix

Agent Read Write Edit Task WebFetch Glob Grep Bash
🧵 Loom ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
🪡 Tapestry ✓ ✓ ✓ ✗ ✓ ✓ ✓ ✓
🚀 Shuttle ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
📐 Pattern ✓ .md* .md* ✗ ✓ ✓ ✓ ✓
🔍 Thread ✓ ✗ ✗ ✗ ✓ ✓ ✓ ✓
🌐 Spindle ✓ ✗ ✗ ✗ ✓ ✓ ✓ ✓
✅ Weft ✓ ✗ ✗ ✗ ✓ ✓ ✓ ✓
🔐 Warp ✓ ✗ ✗ ✗ ✓ ✓ ✓ ✓

* Pattern's Write/Edit restricted to .weave/*.md files only

{workflow}

How It Works

A structured workflow with built-in quality and security gates

1

Pattern

Plan

Loom assesses complexity and delegates to Pattern. Pattern researches the codebase via Thread, consults external docs via Spindle, and produces a structured plan with checkboxed tasks in .weave/plans/.

# Add Search Functionality
- [ ] Create search index service
- [ ] Build search API endpoint
- [ ] Add search UI component
↓
2

Weft

Review

Weft validates the plan — checks file references exist, tasks have sufficient context, no contradictions. Approval-biased: only rejects for true blocking issues. If rejected, Pattern revises.

[APPROVE]
Plan is solid. Tasks are well-scoped
and references resolve correctly.
↓
3

Tapestry

Execute

Run /start-work to begin. Tapestry reads the plan, executes tasks sequentially — writing code, running commands, verifying output — and marks each checkbox. Interrupted? Resume from exactly where you left off.

- [x] Create search index service
- [x] Build search API endpoint
- [x] Add search UI component
✓ 3/3 tasks complete
↓
4

Weft + Warp · automatic

Review & Audit

After all tasks complete, Weft reviews code quality and Warp audits for security vulnerabilities — automatically, in parallel. This is a mandatory gate. If either flags an issue, it's surfaced before the workflow completes.

Weft: Reviewing code quality...
Warp: Auditing security...
Weft: [APPROVE] No issues found
Warp: [APPROVE] No vulnerabilities
✓ All reviews passed
The Plan → Review → Execute → Audit workflow in action: Loom delegates to Pattern, Pattern creates a plan, Weft reviews, Tapestry executes, and Weft + Warp audit automatically

Simple tasks? Loom handles them directly — no plan needed.

{install}

Get Started in 30 Seconds

1

Add to opencode.json

{
  "plugin": ["@opencode_weave/weave"]
}
2

Restart OpenCode

✓

OpenCode automatically installs npm plugins at startup — no manual bun add or npm install required. All agents are immediately available.

Requires OpenCode · Read the full guide

{config}

Configure Everything. Or Nothing.

Zero config out of the box. JSONC configuration with comments when you need control.

  • Override models and temperature per agent

    Use Claude for planning, GPT-4o-mini for exploration, Gemini for UI work.

  • Define domain categories with specialized models

    Route visual-engineering tasks to vision-capable models automatically.

  • Inject custom skills from markdown files

    Prepend domain knowledge to agent prompts without touching source code.

  • Disable specific hooks, agents, or tools

    Turn off what you don't need. Full control over the agent roster.

  • Tune background agent concurrency

    Control how many agents run in parallel to manage cost and rate limits.

  • Project-level and user-level config merging

    Team defaults in the repo, personal overrides in ~/.config/.

Project: .opencode/weave-opencode.jsonc
User: ~/.config/opencode/weave-opencode.jsonc
{
  // Override agent models
  "agents": {
    "loom": {
      "model": "anthropic/claude-3-5-sonnet",
      "temperature": 0.1
    },
    "thread": {
      "model": "openai/gpt-4o-mini"
    }
  },
  // Category-based dispatch
  "categories": {
    "visual-engineering": {
      "model": "google/gemini-2-pro"
    }
  },
  // Feature toggling
  "disabled_hooks": [],
  "disabled_agents": []
}

weave-opencode.jsonc

{faq}

Frequently Asked Questions

What models does Weave support?

Any model supported by OpenCode — Claude, GPT-4o, Gemini, local models via Ollama. Override models per agent in your config. Mix and match based on cost and capability.

Does Weave work with my existing OpenCode setup?

Yes. Install the plugin, add it to opencode.json, and restart. Your existing config and workflows are untouched. Weave layers on top.

Can I customize which agents are available?

Yes. Disable any agent via disabled_agents in config, override models and temperatures per agent, or inject custom skills from markdown files to specialize behavior.

Is Weave free?

Yes. MIT licensed, open source, free forever. You pay only for the AI model API calls you make — Weave itself has no cost.

How is this different from just using OpenCode?

OpenCode gives you one powerful agent. Weave adds a structured workflow on top — you talk to Loom, and it orchestrates planning, review, execution, and security auditing automatically. After code is written, quality and security reviews run as mandatory gates. The agents are still there (Pattern, Weft, Tapestry, Warp, and more), but you interact with one.

{also_from_weave}

Your agents can spawn other agents

With Agent Fleet and the fleet-orchestration skill, Loom and Tapestry autonomously spawn child sessions, parallelize work across repos, and collect results — all from a single browser tab.

Explore Fleet