Forge Agent is a separate Forge product that remains available for existing customers. New inquiries here route to /contact.html.
Open-source. Type-safe tool use. Plug-and-play for any LLM. Ship a production agent in under 50 lines of code without Python circus tricks or 14-layer abstractions. Forge Agent is the runtime we wish we had when we started building voice receptionists, trading bots, and autonomous research agents.
No hidden chain-of-chain-of-agents. One class, typed tools, your LLM of choice. Run it locally, deploy it anywhere a Node or Python process runs.
// npm install @forge/agent @forge/tools import { Agent } from "@forge/agent"; import { webSearch, sendEmail } from "@forge/tools"; const researcher = new Agent({ model: "claude-opus-4-7", tools: [webSearch, sendEmail], systemPrompt: "You are a research analyst. Investigate leads, then email the summary.", }); await researcher.run({ input: "Research Forge Dev Studio and email findings to corey@forgedev.studio", maxSteps: 10, });
Every feature earns its place. No graph-of-graphs. No config file gymnastics. Just the primitives you need to build an agent that runs in production without babysitting.
Declare a tool's input and output schemas once in TypeScript or Zod. The runtime enforces them, catches hallucinated arguments, and returns typed results your code can trust.
Swap between every major hosted model and local-runtime models with one line. Prompt caching, tool use, and streaming work the same across providers.
Drop in first-party plugins for MCP servers, vector stores, CRM integrations, voice, browser automation. Or write your own in ten lines and publish to npm.
Server-sent events out of the box. Users can see tokens land in real time and interrupt the agent mid-stream without losing state.
Force JSON schemas, enum choices, or typed object trees. The runtime retries on shape mismatch and surfaces typed objects to your caller.
Every tool call, every reasoning step, every retry is logged with OpenTelemetry-compatible traces. Plug into Grafana, Datadog, or Forge Vault and watch agents work in real time.
The production bilingual voice agent that handles real calls at (706) 290-0000 is built on Forge Agent. One instance per call, typed CRM and calendar tools, low-latency streaming voice plugin.
Market-regime classifiers feed signals to execution agents that submit orders to Kraken and Alpaca. Runs 24/7 with budget tracking and human override.
Inbound web forms feed a triage agent that enriches with firmographic data, scores urgency, and routes to the right partner or calendar. Handles 2,000+ leads per day for our staffing partner.
Multi-step web research with source citations. Plugin chain: webSearch → fetchUrl → summarize → synthesize. Replaces a junior analyst for early-stage diligence memos.
First-party plugins for every major model provider, vector store, payments processor, and messaging platform. Everything else is a ten-line plugin.
Yes. It runs our production bilingual voice agent on a real phone number, our quant trading bots, and our lead triage system. Every major release is smoke-tested against those workloads before shipping.
Two things. One: no fake abstractions. We do not wrap every LLM call in a Chain of a Chain of a Graph of a Runnable. The runtime is the model, your tools, and a loop. Two: typed tool use is first-class. Tool arguments and results are fully typed, schema-validated, and retried on shape mismatch without config.
No. Forge Agent runs on any Node or Python host. Bring your own LLM API key and you are done. Forge Vault (separate product) is our cloud option if you want hosted multi-tenant agent management, but the runtime itself never phones home.
MIT. Use it commercially, modify it, redistribute it. No strings, no telemetry.
First-class. Forge Agent can consume MCP servers as plugins or expose its own tools as an MCP server for any MCP-aware client.
Read the quickstart on GitHub, then pick one of the example agents in /examples (voice receptionist, lead-scoring agent, research agent). If you can read 200 lines of TypeScript you can ship your first agent today.
Star the repo. Read the quickstart. By Monday you have a production-ready agent with typed tools and your LLM of choice. No abstractions hunt, no plumbing tax.