Skip to content
botiversePublic

About

oar (Open Agent Runtime): programming interface for all agent harnesses

Resources

Stars

681 stars

Watchers

0 watching

Forks

Repository files navigation

OAR - programming interface for all agent harnesses Built with Raft

OAR logo

OAR (Open Agent Runtime) is a provider-independent programming interface for coding-agent runtimes: a solid foundation for building agent workspaces and other applications.

Delete harness logic and focus on outcomes and UX.

OAR is built in the open: the agents and people working on it talk, review and ship on Raft, where you can follow along.

Supported runtimes


Antigravity

Claude Code

CodeBuddy Code

Codex

Cursor

DeepSeek Harness

Devin

Factory Droid

Grok Build

Kimi Code

OpenCode

Pi

Qoder

Qwen Code

Pi Durable also runs through a host-owned Harness, including in a browser (demo).

Library

One API drives every runtime: swap "claude" for "codex", "grok" or "pi" (with a model it has) and the code stays the same. An option a runtime cannot honor is refused before anything opens, never silently dropped.

import { defaultRuntimes, promptAndWait } from "@botiverse/oar";

const claude = defaultRuntimes.require("claude");
const installation = await claude.installation?.();
if (installation?.kind !== "available") throw new Error("claude is not installed");

const session = await claude.session(installation, { cwd: process.cwd(), model: "sonnet" });

session.events((event) => {
  // Every event says which agent it came from: [] is the root.
  if (event.kind === "text_delta" && event.agentPath.length === 0) process.stdout.write(event.text);
  if (event.kind === "tool_call_started") console.log(event.tool);
});

const run = promptAndWait(session, "Find and fix the flaky test", { timeoutMs: 600_000 });
// From your UI: steers the running turn, or queues where the runtime cannot steer.
await session.deliver("Leave the snapshots alone.");
console.log(await run);
await session.dispose();

// Later, even from another process: the same conversation.
const resumed = await claude.session(installation, { cwd: process.cwd(), resume: session.id });

A session gives you:

  • Every fact, attributed. One event per fact with seq and agentPath; rawEvents() keeps every native payload verbatim (record stream).
  • Control with honest answers. prompt, steer, queue, withdraw and abort answer accepted or rejected with a typed code (conversation). Runtimes with native compaction also offer compact (contract).
  • Your own tools and subagents. serveMcpTools attaches host tools to any runtime; @botiverse/oar/agents hands work to other runtimes (host tools).
  • Views and tests. @botiverse/oar/observe folds records for a UI, in a browser too; @botiverse/oar/testing is a scripted runtime with no binary or login.

What differs by runtime is declared on the Runtime, not guessed (matrix). Entry points and setup: package README.

Without a session

The same API answers the questions a dashboard or setup wizard asks, for every runtime:

for (const runtime of defaultRuntimes.list()) {
  const installation = await runtime.installation?.();
  if (installation?.kind === "available") {
    console.log(runtime.id, await runtime.listModels?.(installation));
  }
}
  • Installation: whether each runtime is installed, its version and where; install and update through the vendor's own installer and updater, and the minimum version OAR supports where a runtime has one.
  • Account: log in, log out and status through the runtime's own login, without a terminal (OAR never handles the tokens); usage: plan, quota windows and their resets, and Codex and Claude limit resets.
  • Models: what the installation can run now, with efforts and service tiers.
  • Inventories: configured skills, commands, MCP servers and tools.
  • Setup forms: settings as data a UI renders.
  • Orientation: all of the above for every runtime in one bounded call.

CLI

The same API on the command line: set up runtimes on a machine (install, sign in, check usage and models) and run a turn to see what a runtime does, for setup, debugging and verifying OAR itself. oar mcp also lets an agent hand tasks to other runtimes.

npx @botiverse/oar-cli installation
oar install codex --plan
oar account login claude
oar account usage
oar run claude "What does this repo do?" --record run.jsonl

Every command: packages/cli (published as @botiverse/oar-cli).

Projects using OAR

  • Raft: where people and agents work together; Raft Computers run agent runtimes through OAR.
  • Kaze: a calm, quick browser for macOS and Windows, with an agent that works in your tabs.
  • DeepSeek Harness: subagent-oar hands DeepSeek Harness subagents to other runtimes.
  • repo-harness: file-backed workflow harness for Claude Code and Codex sessions.
  • Eta: a self-contained multi-agent loop, runtime and app.
  • dsh-agent-team: agent teams for DeepSeek Harness.
  • dsh-engine-switch: a DeepSeek Harness plugin that hands the main session to another agent runtime.
  • ducter: a personal AI that hands work to coding agents on your machines.
  • tmux-agent-deck: control tmux sessions and coding agents from a web terminal.

Using OAR in a public project? Open a pull request to add it here.

Contributing

Issues and pull requests are welcome, from people and from agents, including the agents of projects that use OAR. Issues are preferred. AI-written contributions are fine without a person checking them by hand, as long as the AI verified them in a real environment: the real runtime, a real account, a real task.

  • Issues. Name the @botiverse/oar (or @botiverse/oar-cli) version, the runtime and its version, and the OS; say what you called, what you expected and what happened; include the session's records that show it (session.records(), or a log from oar run ... --record run.jsonl), redacted, and how you reproduced it in a real environment. One problem per issue.
  • Pull requests. Read docs/development.md first. One change per pull request, rebased on main, titled in Conventional Commits (fix(pi): ...); a fix comes with a test that fails without it; docs are updated in the same change; CI=1 pnpm run check passes (without CI=1, snapshot tests rewrite their snapshots instead of failing). Do not bump package versions or push tags: releases are cut by the maintainers. Say in the description how the change was verified in a real environment.
  • Never paste credentials, tokens, account identifiers or emails, or machine-local paths.
  • Projects that use both Raft and OAR can ask contact@raft.build for a joint channel with the OAR team, to take part more directly.

Docs

  • Design: why oar exists, its decisions and what comes next.
  • Spec: the record stream and every contract built on it.
  • Runtimes: what each runtime says natively and how oar maps it.
  • Prior art: related projects compared feature by feature.
  • Development: working in this repo.

ESM-only, Node entry requires Node.js 24+; the browser entry needs a modern browser with Web Crypto. Apache-2.0.

About

oar (Open Agent Runtime): programming interface for all agent harnesses

Resources

Stars

681 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages