weft-core 0.1.1

OpenAI-compatible AI agent runtime with WASM capability plugin system
Documentation

Weft Core

OpenAI-compatible API + WASM capability plugin OS

Weft Core is a standalone Rust runtime that exposes an OpenAI-compatible HTTP API and loads capability packages (WASM via Extism, native, or embedded) to extend behavior — agent loops, tools, memory, workflows, and more. Any client that speaks the OpenAI protocol (or plain HTTP) can drive it; no UI required.

[!WARNING] Early preview. Weft Core is under active development. HTTP endpoints, capability contracts, and package manifests may change without notice. Treat this as an exploration preview, not a production dependency.

Quick start

Requires a stable Rust toolchain.

# From the repository root
mkdir -p config
cp config.example.toml config/config.toml
# Edit config/config.toml and add at least one [[providers]] entry with API keys

cargo build --release -p weft-core
cargo run --release -p weft-core --bin weft-core

The server binds to 127.0.0.1:17830 by default (loopback only).

Health check

curl -s http://127.0.0.1:17830/api/health
# {"status":"ok","version":"0.1.0"}

Chat completion

Most endpoints require a loopback bearer token. On first start, the core writes one to data/runtime-token:

TOKEN=$(cat data/runtime-token)

curl -s http://127.0.0.1:17830/v1/chat/completions \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-chat",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

Configure providers and default routing in config/config.toml. See config.example.toml for the full template.

What you get

  • OpenAI-compatible surface — /v1/chat/completions (SSE streaming), /v1/models
  • Multi-provider router — OpenAI, Anthropic, DeepSeek, OpenRouter, and any compatible endpoint, with API-key failover/rotation
  • Capability registry — stable capability ids (agent.runtime, tool.shell, memory.store, …) resolved to packages at startup
  • Pluggable pipeline — request → transform → provider → response; each stage overridable by WASM packages
  • Management API — /api/apps, /api/capabilities, /api/packages, /api/providers, and more

How it compares to Pi (pi-agent-core)

Pi Agent Core (@mariozechner/pi-agent-core) is a TypeScript in-process library: you embed an Agent class, wire tools in code, and subscribe to streaming events. It sits on top of @mariozechner/pi-ai for unified LLM calls.

Pi Agent Core Weft Core
Form factor npm library (in-process) Standalone HTTP server
Language TypeScript Rust
Extension model Tools & hooks in application code WASM / native capability packages loaded at runtime
API surface Programmatic Agent class OpenAI-compatible REST + management API
Agent loop Built into the library Provided by packages (e.g. agent-core) on the capability bus
Best for Embedding agents inside Node/Browser apps Running a local agent platform any HTTP client can talk to

Pi excels at embedding a minimal agent loop in your app. Weft Core excels at being a local capability OS — swap packages without recompiling, serve multiple clients (scripts, desktop UI, OpenAI SDK) over one loopback endpoint.

Official desktop client (optional)

Weft Desktop is the optional official UI — chat, multi-agent teams, package manager, and product surfaces. It connects to the same loopback API (127.0.0.1:17830). You can run Weft Core alone and bring your own client.

Prebuilt desktop bundles: Weft Releases.

Documentation & references

Resource Description
examples/ Minimal integration examples (sidecar HTTP, custom WASM packages)
openapi.yaml OpenAPI 3 spec for the HTTP surface
docs/SEMVER.md Versioning policy
docs/CRATES_IO.md Publishing to crates.io
Weft monorepo architecture Full platform architecture (client + packages)
Weft package index Full capability → package source authority

Configuration

[core]
host = "127.0.0.1"
port = 17830

[[providers]]
name = "deepseek"
base_url = "https://api.deepseek.com"
format = "openai"
models = ["deepseek-chat"]

[[providers.keys]]
value = "sk-your-key"

config/config.toml is gitignored; only the example template is committed.

License

Licensed under the Apache License 2.0.

Install from crates.io

[dependencies]
weft-core = "0.1"
weft-package-sdk = "0.1"

Binary: download from weft-core Releases or build with cargo install weft-core (after first crates.io publish).