Slang AI Agent

<h1 align="center">🗣️ SLANG</h1> <p align="center"> <img src="slang.png" alt="SLANG" width="200"/> </p> <p align="center"> <strong>Your AI workflow in a format anyone on the team can read, edit, and validate.</strong><br/> Write the workflow in SLANG, your favorite LLM runs it. No code needed.<br/> </p> <p align="center"> <a href="#two-ways-to-use-slang">How it works</a> • <a href="#quick-start">Quick Start</a> • <a href="#why-slang">Why SLANG</a> • <a href="docs/IDE.md">IDE Support</a> • <a href="docs/PLAYGROUND.md">Playground</a> • <a href="docs/CLI.md">CLI</a> • <a href="docs/API.md">API</a> • <a href="docs/MCP.md">MCP</a> • <a href="SPEC.md">Spec</a> • <a href="GRAMMAR.md">Grammar</a> </p>

---

The entire language in 30 seconds.

flow "research" {
  agent Researcher {
    tools: [web_search]
    stake gather(topic: "quantum computing") -> @Analyst
  }
  agent Analyst {
    await data <- @Researcher
    stake analyze(data) -> @out
    commit
  }
  converge when: all_committed
}

That's it. Three primitives. Your PM can read it. Your analyst can edit it. Your LLM can run it. No Python, no TypeScript — just intent.

Everything else follows from these:

Primitive	What it does
stake	Produce content and send it to another agent (or execute locally)
await	Block until another agent sends you data
commit	Accept the result and stop

Plus control flow: when/else conditionals, let/set variables, repeat until loops.

---

Two ways to use SLANG

Here's the thing: the same .slang file runs two different ways. Paste it into ChatGPT? Works. CLI? Works. API? Works. You pick.

<table> <tr> <td width="50%" valign="top">

🧠 Zero-Setup Mode

No install, no API key, no runtime.

1. Copy the system prompt
2. Paste into ChatGPT, Claude, Gemini (pick any LLM)
3. Paste your .slang flow
4. Done

The LLM becomes your runtime. Perfect for non-developers, quick prototyping, or when you just want it to work — no install, no code, no friction.

</td> <td width="50%" valign="top">

⚡ CLI / API / MCP Mode

Full runtime. Real tools. 300+ models via OpenRouter. Parallel execution.

npm install -g @riktar/slang
slang init my-project && cd my-project
slang run hello.slang

Checkpoint and resume. Deadlock detection. Structured output. Everything you need for real workflows.

</td> </tr> </table>

Feature	Zero-Setup	CLI / API / MCP
Parse & execute flows	✅	✅
All primitives (stake, await, commit, escalate)	✅	✅
Conditionals (when / if / else)	✅	✅
Variables (let / set)	✅	✅
Loops (repeat until)	✅	✅
deliver: post-convergence hooks	❌	✅ real handlers
model: multi-provider routing	❌ single LLM	✅ 300+ models
tools: functional execution	❌ simulated	✅ real handlers
Parallel agents	❌ sequential	✅ Promise.all
retry: with exponential backoff	❌	✅
output: structured contracts	✅ best-effort	✅ enforced
Checkpoint & resume	❌	✅
Static analysis & deadlock detection	❌	✅
IDE support (LSP, syntax highlighting)	❌	✅
Web playground	❌	✅

Start with zero-setup to prototype. Move to CLI or API when you're ready to ship. Same file both ways.

---

Quick Start

1. Install

npm install -g @riktar/slang

2. Scaffold a project

slang init my-project
cd my-project

This creates hello.slang, research.slang, tools.js, and .env.example. Everything you need.

3. Configure your API key

cp .env.example .env    # edit with your API key (SLANG loads it automatically)

4. Run your first flow

slang run hello.slang                        # echo adapter (no API key needed)
slang run hello.slang --adapter openrouter   # uses OPENROUTER_API_KEY from .env

5. Open the playground

slang playground
# Opens http://localhost:5174 - write, inspect, and visualize flows in the browser

---

Why SLANG?

It's not a framework. It's a shared language for your team.

SLANG is Super Language for Agent Negotiation & Governance

LangChain, CrewAI, AutoGen — they're SDKs. Python/TypeScript libraries. Only developers can use them. Everyone else has to wait, ask, or guess what the workflow actually does.

SLANG is a language anyone can read and write. Your PM defines the workflow. Your analyst tweaks the logic. Your developer hooks it up to real tools. Everyone works on the same .slang file.

SQL	SLANG
Doesn't replace C/Java	Doesn't replace Python/TypeScript
Non-devs write queries	Non-devs write workflows
Readable by the whole team	Same — anyone reads and edits it
LLMs generate it	LLMs generate it
Not complete. That's the point	Not general-purpose. That's the point

No code needed.

Describe what your agents should do. SLANG reads like plain English:

flow "hybrid-analysis" {
  agent Researcher {
    tools: [web_search]
    stake gather(topic: "quantum computing") -> @Analyst
    commit
  }
  agent Analyst {
    await data <- @Researcher
    stake analyze(data) -> @out
    commit
  }
  converge when: all_committed
}

Paste it into ChatGPT and it runs. Use the CLI for production with 300+ models via OpenRouter. Same file, zero vendor lock-in.

The .slang file is the documentation.

Read this flow out loud:

"The Researcher stakes gather on "quantum computer" topic and sends it to the Analyst. The Analyst awaits the data, analyzes it, and sends the output to the user. The flow stop when the Researcher and the Analyst have committed their job"

No diagrams, no comments, no docs needed. Show the .slang file in a meeting and everyone understands what the AI workflow does.

Who is SLANG for?

Audience	Why
PMs & business people	Write AI workflows without code. Describe what agents should do, paste into ChatGPT, and it runs. Your automation, your way.
Analysts & ops	Edit and validate workflows yourself. No waiting for engineering. Review the logic, tweak parameters, run it.
Developers	10 lines, 60 seconds, it runs. Skip the boilerplate, hook up real tools when you need them.
Teams	One .slang file everyone can read. The PM writes it, the dev ships it, the analyst audits it. Same source of truth.

SDK comparison

	SDK (LangChain, CrewAI)	SLANG
Who can use it	Developers only	Anyone on the team
Time to first workflow	Hours	60 seconds
Who reads / reviews it	Developers only	PMs, analysts, developers, LLMs
LLMs can generate it	No (boilerplate is messy)	Yes (text-to-SLANG like text-to-SQL)
Runtime needed	Yes, always	Optional — paste into ChatGPT and it works
Docs	Separate files	The .slang file is the documentation

SLANG isn't trying to replace SDKs. Like SQL didn't replace Java. It's a different category: workflows that everyone on the team can own.

---

IDE Support

See docs/IDE.md for VS Code, Neovim, Vim, Sublime, JetBrains, and other LSP-compatible editors.

---

Playground

See docs/PLAYGROUND.md for web editor features and usage.

---

CLI

See docs/CLI.md for all commands, options, and environment variables.

---

API

See docs/API.md for programmatic usage, adapters, tools, and checkpointing.

---

MCP Server

See docs/MCP.md for Model Context Protocol integration with Claude Desktop.

---

How it works

Your .slang file goes through these stages:

Source → Lexer → Parser → AST → Resolver → Graph → Runtime → Result

Stage	What happens
Lexer	Breaks source into tokens (with line/column info)
Parser	Recursive-descent, builds typed AST, recovers from errors
Resolver	Builds dependency graph, checks for deadlocks
Runtime	Schedules agents, mailbox, parallel dispatch, tools
Adapters	Connect to LLMs (OpenAI, Anthropic, OpenRouter, etc)
LSP	Language Server: diagnostics, completion, go-to-def, hover
Playground	Web editor (React + Vite) with visualization, tests

Examples

Examples are in the examples/ folder. Each demonstrates different patterns and features:

slang run examples/hello.slang                                     # Minimal flow
slang run examples/research.slang --adapter openrouter --tools examples/tools.js
slang check examples/broadcast.slang                               # Dependency analysis

See each .slang file for inline documentation.

Contributing

See CONTRIBUTING.md for guidelines.