Build agents that can act on context.

Build a specialist AI agent with models, tools and persistent context. An Oracle can read evidence, check it against agreed requirements and explain its recommendation. QiForge provides the runtime and plugin framework.

published runtime
npm
package  @ixo/oracle-runtime@1.3.3
entry    createOracleApp(options)
plugins  BUNDLED_PLUGINS
source   github.com/ixoworld/qiforge

Setup instructions: docs.ixo.world/build-an-oracle

Runtime published Follow the setup guide

Start with the plugins your agent needs.

QiForge is the framework behind IXO Oracles. The public package exposes createOracleApp, plugin classes, manifests, and the bundled plugin set. The framework assembles the web server, guarded model runtime, storage, streaming chat, API docs, channel adapters, and selected capabilities around a validated configuration.

Startup is explicit. Missing settings name the plugin and the fix. Tool-name collisions stop the boot. Unconfigured plugins are skipped cleanly.

main.ts · entry
TypeScript
import {
  BUNDLED_PLUGINS,
  createOracleApp,
  defineOraclePlugin,
} from "@ixo/oracle-runtime";

// createOracleApp requires validated config and environment.
// Follow the build guide before starting a networked runtime.
Give this to your coding agent Build your first Oracle capability.

Ask your coding agent to build and test one useful capability in your project. Review the result before connecting it to live services.

Review runtime source
You are making one safe, reviewable change to an existing IXO Oracle codebase.

Before editing, read the repository instructions, package.json, lockfile, current git status, runtime entry point, and existing tests. Inspect the official build guide at https://docs.ixo.world/build-an-oracle, package metadata at https://www.npmjs.com/package/@ixo/oracle-runtime, and matching source at https://github.com/ixoworld/qiforge/tree/main/packages/oracle-runtime. Confirm the installed version and exact exported APIs for createOracleApp, defineOraclePlugin, plugin manifests, and BUNDLED_PLUGINS. Use only APIs confirmed in the installed version or its matching source.

Do not run qiforge-cli or another scaffolding CLI. Do not create or modify a wallet, generate keys, connect to Matrix, register an Oracle, publish, deploy, bill, write secrets, or change chain state.

Implement the smallest useful local change: one deterministic, read-only plugin capability that reports application health and configuration presence without returning secret values. Follow the repository's existing plugin and test conventions. Add focused tests. Do not enable network calls or register the plugin in a production runtime unless existing configuration and tests explicitly support that step. If this is not an Oracle project or the required API is unavailable, stop after an evidence-backed implementation plan.

Report the files changed, official APIs confirmed, validation commands and results, configuration still required, and human review required before any networked or irreversible action.

Load the tools the task needs.

An Oracle finds the relevant plugin or skill when a task needs it. It loads the instructions and tools, ready to use throughout the conversation.

Plugins
Add tools, sub-agents, middleware, settings, hooks, web routes, shared data, and browser events.
Skills
Packaged abilities discovered on demand, then executed in the Sandbox.
Catalog
Start with the bundled plugins or add your own. Explore the plugin guide.

Give your agent memory, tools and skills.

Memory Engine

Private per-user memory

Give your agent context it can recall and update, with links to the conversations behind it. Memory activity is recorded in the user's encrypted Matrix room. Control retention, reuse and model access for your deployment.

AI Sandbox

Guarded execution workspace

Shell, Node, Python, API calls, and file generation. Commands pass rules screening and review before execution. Secrets inject only for verified skills and redact from output.

Skills System

Publish once, run immediately

Public and private skills appear in search right after publish and run without rebuilding the Oracle.

Set access and permissions for each agent.

Each Oracle has a registered identity. Requests carry tightly scoped UCAN permissions. Delegations are short-lived and replay-checked. Connected credentials are securely used by tools without being shown to the model.

identity
Oracle DID registered and authenticated on-chain, with verifiable credentials
storage
Per-user encrypted data, events, and files within the configured deployment boundary
durability
Encrypted snapshots and durable objects stored in Matrix
secrets
Secrets are encrypted, used by reference, and never shown to the model
metering
Model tokens and metered service usage, with spend caps

One Oracle. Web, Matrix, Slack.

The same runtime serves browser chat, Matrix Rooms, and Slack. It streams replies, reasoning state, and tool progress. In the browser, it can request UCAN-authenticated actions from the user's app and continue with the result.

Web Streaming chat and UCAN handoff from the user's app
Matrix Rooms with encrypted timeline and durable state
Slack Workspace adapters on the same runtime

Available components
and planned additions.

In production
  • Runtime and plugin framework
  • Capability discovery
  • Model runtime with tool permissions
  • Streaming multimodal chat
  • Web, Matrix, and Slack channels
  • UCAN security
  • Durable per-user storage
  • Memory Engine
  • Sandbox hosting
  • Skills registry
  • Reference Oracle and published runtime packages
  • 15 concrete bundled plugin implementations
In development
  • Side-effect-free public CLI quickstart
  • Production Calls plugin
  • Billing plugin
  • Verification plugin
  • Write your own plugin