Repository navigation
v0.0.3
WARP v0.0.3 Release Notes
We are excited to announce the release of WARP v0.0.3. This is a major update that transitions WARP from a basic declarative format to a full-fledged agent engineering platform.
This release introduces a robust Plugin Ecosystem, native Model Context Protocol (MCP) integration, fine-grained Security & Tool Policies, hierarchical Subagents Delegation, and a highly advanced Context-Aware Templating Engine.
🚀 Key Highlights
1. Plugin Ecosystem & Package Management
WARP now supports modular packaging and distribution of your agents, skills, commands, and tools.
- CLI Package Installer: Use the new
warp getcommand to fetch, configure, and install plugins directly from remote repositories. - Interactive Configuration: Powered by
huh, the CLI prompts you to select exactly which resources to import. - Integrity Lockfile (
warp.lock): Computes cryptographic checksums of the plugin tree to guarantee workspace supply-chain security. - Post-Install Setup Hooks: Supports lifecycle execution hooks (
hooks.postInstall) to automatically download assets or compile plugin binaries upon installation. - Developer DX (Live Linking): Specifying a relative or absolute local directory as the plugin source bypasses the global cache and enables real-time hot-reloading during development.
2. Model Context Protocol (MCP) Integration
Native integration of the Model Context Protocol allows agents to dynamically interact with external development servers.
- New
Kind: MCPResource: Declare MCP server configurations using a simple schema. - SSE & Stdio Transports: Connect to servers via shell execution (stdio) or HTTP streams (SSE) with support for custom environment variables.
- Resource Filters: Use
policies.tools,policies.prompts, andpolicies.resourceswithinclude/excludeglob lists to filter what the MCP server exposes to agents. - Annotations & Safety Profiles: Assign safety labels (like
isDangerousorisOpenWorld) to all exposed tools, with key overrides for individual tools.
---
apiVersion: warp/v1alpha1
kind: MCP
metadata:
name: github-mcp
spec:
type: stdio
command: ["npx", "-y", "@modelcontextprotocol/server-github"]
env:
GITHUB_PERSONAL_ACCESS_TOKEN: "${GITHUB_TOKEN}"
policies:
tools:
include: ["get_issue", "search_code"]
exclude: ["create_or_update_file"]
---3. Fine-Grained Security & Tool Policies
Administrators and users can now enforce strict permission boundaries on tool execution.
- Agent & Workspace Policies: Restrict tool usage at the workspace level (
WORKSPACE.md) or override them per-agent (spec.policies). - Dangerous Tool Prevention: Setting
allowDangerous: falseblocks execution of any tool annotated as dangerous. - Open-World Access Limits: Restrict tool executions that make uncontrolled network requests or touch the internet.
- Inheritance & Merging: Child agents automatically inherit, combine, and resolve parent policies during agent resolution.
4. Hierarchical Subagents Delegation
Agents can now collaborate and spawn subagents under clean control flow structures.
- Delegation Reference Lists: List allowable subagents (
spec.subagents) that an agent is permitted to delegate tasks to. - Trigger Filtering: Use
spec.triggersto define what architectural components (e.g.human,agent,system) are allowed to invoke a particular agent. - Wildcard Imports: Support for wildcards (
*) to import all commands, skills, or subagents.
5. Rich Model Provider Configurations
LLM Provider specifications have been expanded to capture modern production capabilities and cost tracking.
- Custom HTTP Headers: Configure custom authentication or tracking headers for all outgoing API requests.
- Pricing Matrices: Track token usage costs (
cost), cache read/write discounts, and conditional pricing tiers based on context window size. - Thinking Token Settings: Define sampling bounds and strategy controls (like
toggle,effort,budget) for reasoning-based models. - Modality Support: Declare accepted input and output modalities (e.g., text, images, video) per-model.
🛠️ Developer Experience & Tooling
Advanced Context-Aware Templating
The templating engine now injects highly structured hierarchical view objects into your instructions.
- Workspace & Project Directives: Access workspace roots, project directories, and local rules files (
AGENT.md). - Shorthand Arguments Mappings: Commands map user args to hints automatically (e.g.
$ticketmaps to the first argument if configured). - Self-Aware Metadata: Agents and skills can inspect their own metadata (e.g.
{{.Agent.Name}},{{.Skill.Dir}}).
New Formatting Tooling
- Format API: Serialize parsed WARP structures back into standard WARP Markdown.
- Resilient Directory Loader: The filesystem loader now skips files lacking YAML front-matter (like READMEs) inside
.agents/rather than throwing errors. - Reflected JSON Schema: Dynamically updates
schema/warp.jsonbased on Go struct reflection for VS Code validation and autocompletion.
📈 Commits Summary
- Plugin Architecture: Commits
6d02fbc(Plugin management),f0f2a88(Local files & deduplication),11ea062(Plugin hooks and triggers for skills),d02ee3d(Validation rules). - MCP Integration: Commits
12acdcf(MCP support),6128916(MCP policies). - Security Policies: Commits
c079fac(Tool policies),90b0178(Agent tool policies). - Subagents: Commits
7c89ee6(Subagents support),58a5cd1(Filter on trigger type),44410db(Wildcard support). - Model Provider: Commits
19b6cc7(Headers/config),0b2bf1d(Thinking config and cost),e564f85(Output schema),0af54d9(Advance auth). - Templating & Context: Commits
1d6863a(Workspace templating),83c8236(Template paths/names),eb09684(Global lists rendering). - CLI & Loader: Commits
dc32745(CLI get update),f28a1c7(Custom resource path),81126c7(Skip files without front matter). - Formatting: Commits
37c73b5(Format API),d014c37(Formatter improvements).