AgentForge

Value

Problem. Builders rewrite the same tool-calling loop, sandbox, and memory glue for every local agent demo.

How it helps. A runnable TypeScript agent: system prompt → tool-calling loop → Zod tools → memory, with a Repo Helper demo CLI and optional /chat server. Bring your own OpenAI-compatible key.

Why buy. A small, honest agent starter you can fork in an afternoon — not AGI, not LangChain soup.

• Tool-calling loop with max steps and stop conditions
• Sandboxed read_file/write_file; shell disabled by default
• CLI + optional Express /chat; calculator, http_fetch, memory JSONL

How it works

flowchart LR
  User["User / CLI /chat"] --> Agent["Agent"]
  Agent --> Loop["Planner / Loop"]
  Loop --> LLM["OpenAI-compatible"]
  Loop --> Tools["Tool Registry"]
  Tools --> FS["read/write sandbox"]
  Loop --> Memory["Memory"]

Mermaid flowchart (render in GitHub / VS Code / mermaid.live).

Use case

Ask the Repo Helper to write workspace/hello.md and summarize it — tools run inside ./workspace only.

What you get

• Agent façade + loop
• Default tool registry + Repo Helper personality
• CLI and optional server
• 27 Vitest tests (loop, memory, registry, sandbox)
• MIT license

Project structure

agentforge/
  src/
    agent/   agent.ts, loop.ts
    tools/   calculator, FS, httpFetch, shell, registry
    memory/  safety/  llm/  demo/  cli/  server/
  tests/ workspace/

Quick start

cp .env.example .env  # OPENAI_API_KEY
npm install && npm test && npm run build
npm run agent -- "Create workspace/hello.md saying AgentForge works, then summarize it."

Example

From src/tools/registry.ts

const tools = createDefaultRegistry();
// register custom ToolDefinition with Zod schema + execute()
// Agent loop: LLM → tool_calls → execute → append → until final text or max steps

Design decisions

• Fetch-based OpenAI-compatible client — works with proxies/Azure/local gateways
• Shell off by default; secret-looking paths soft-blocked
• In-process rate-limit stub to reduce runaway provider spam

Limitations

• Not AGI / fully autonomous agents
• Not a multi-tenant secure runtime — do not expose /chat publicly without auth
• Anthropic not wired natively (use OpenAI-compatible proxy)
• Shell tool is dangerous if enabled

Who it’s for

Indie builders shipping local copilots / agents who want a readable TypeScript spine.

Not for: Teams needing a hosted agent cloud or enterprise multi-tenant sandbox.

What you own

Full MIT-licensed TypeScript source via Gumroad ZIP. Fork it, ship it commercially, keep the license notice. You are responsible for secrets, hardening, and production ops.

Related

McpForge · RagPipe · PromptKit · EvalHarness

Demo