Agents Core Web

English | 简体中文

An open-source web console for creating Agents and running durable conversations on Parsar Agent Core or another compatible Agents API Core.

What it is

Agents Core Web gives an independently deployed Agent Core a focused browser UI and a reusable TypeScript client. The Core remains responsible for authentication, persistence, scheduling, and execution; this project does not embed or reimplement it.

What you can do

• Create, view, edit, and delete reusable Agent configurations.
• Start durable Sessions and inspect their saved Items.
• Follow live progress over SSE and recover persisted output after reconnecting.
• Cancel active work and return function results or errors.
• Use the same Web client with Parsar Core or another proven-compatible Core.

Compatible Agent Cores

Core or interface	Can Web connect?	Notes
Parsar Agents API Core	Yes	Primary tested integration
Another Core implementing the tested /v1/agents/** HTTP/SSE subset	Yes	It must match the resources and behavior in protocol coverage
OpenAI's hosted Agents API	Not claimed	This project does not promise complete hosted-API compatibility
OpenAI Agents SDK, Responses API, or Parsar daemon WebSocket	No	They are an SDK interface, a model API, and an internal execution interface—not directly connectable Core protocols

Web speaks a tested subset of the OpenAI Agents API beta HTTP resource shape: JSON requests and authenticated SSE under /v1/agents/**, with OpenAI-Beta: agents=v1. Compatibility means this documented and tested subset, not merely accepting the header or sharing similar names.

Start Web with an existing Core

You need Node.js 22.12+, pnpm 10.30.3, a running compatible Agent Core, and a caller bearer issued or configured by that Core's operator. For working chat, the Core must also have a configured execution path.

Clone and start the Web:

git clone https://github.com/MiniMax-AI-Dev/agents-core-web.git
cd agents-core-web
pnpm install
pnpm dev

The default local setup expects:

• Agent Core at http://127.0.0.1:8091;
• browser requests through the same-origin /v1 proxy;
• the plaintext caller bearer at ~/.parsar/agents-api/web-token, read only by the local Vite server.

Open the loopback URL printed by Vite. Keep the connection base at /v1; when Server-managed Core key active appears, leave the browser token field empty.

To use another server target or private key file, copy .env.example to .env.local and set these optional server-side variables:

AGENTS_API_PROXY_TARGET=http://127.0.0.1:8091
AGENTS_API_PROXY_TOKEN_FILE=/absolute/private/path/to/web-token

Restart pnpm dev after changing them. Keep credentials server-side. A direct Core URL in the connection dialog is only for a compatible Core that explicitly allows the Web origin, methods, and headers through CORS.

Do not have a Core running yet? Follow Connecting Agent Core. That guide contains the complete PostgreSQL, caller key, device, daemon, native harness, CODEX_HOME, security, verification, and shutdown procedures.

First use

1. Open Agents and create an Agent with a name, instructions, and model ID.
2. Review, edit, or delete the saved Agent, or start a Session from it.
3. Open Sessions, select the Session, and send a message.
4. Follow live Items, cancel active work, or return a requested function result.

The model ID must be supported by the connected execution runtime. Core currently has no model-catalog endpoint, so Web suggestions are editable hints rather than availability guarantees. Successfully saving an Agent proves configuration storage, not that a daemon, model, or provider credential can execute it.

Relationship to Parsar

flowchart LR
  user["User"] --> web["Agents Core Web"]
  web -->|"HTTP JSON + SSE"| core["Compatible Agent Core"]
  core --> runtime["Execution runtime / daemon"]
  runtime --> tools["Models / MCP / tools"]

Loading

Parsar owns its standalone Agents API Core, parsar-daemon, and native Codex/Claude execution adapters. This repository owns only the open Web experience and @agents-core-web/agents-client. The browser connects to the Core protocol; it never uses the daemon WebSocket as its API URL.

See Architecture for the full component and trust boundaries.

Common problems

Symptom	What to check
Web cannot reach Core	Confirm the Core address and AGENTS_API_PROXY_TARGET, then restart Vite
401 invalid_api_key	The plaintext caller bearer must match the current Core key binding
503 execution_unavailable / Execution is not enabled	Core is reachable but has no enabled execution path; connect its configured executor/daemon
Agent saves but its model fails	Use a model ID and provider credential supported by the connected runtime

/healthz proves HTTP liveness only, not chat readiness. See the full troubleshooting guide before retrying an uncertain request.

Documentation

• Connect Agent Core — full local and deployment setup
• Protocol coverage — exact supported API surface
• Architecture — ownership, runtime, and trust boundaries
• Roadmap — planned Web and Core integrations
• Issue-driven iteration — contributor automation contract
• Contributor policy — repository scope, safety, and quality requirements

Agents Core Web is available under the MIT License.