Jinn Code

Multi-agent AI coding assistant for VS Code — powered by AWS Bedrock and local models.

Built by Jinn Systems · Sharjah, UAE

What is Jinn Code?

Jinn Code is a VS Code extension that brings a four-agent AI pipeline directly into your sidebar. Unlike single-model copilots, Jinn Code routes each part of a coding task to the model best suited for it — a Planner to break down the problem, a Thinker to reason through complexity, a Coder to write the implementation, and a Reviewer to catch issues before they hit your files.

It works with AWS Bedrock (Claude, Llama, Nova, DeepSeek, Mistral, AI21, Cohere) and any local model server (Ollama, LM Studio, llama.cpp, or any OpenAI-compatible endpoint) — and you can mix providers per agent.

Features

• Three modes — Ask (Q&A), Plan (step-by-step breakdown), Agent (full pipeline with file writes)
• Four-agent pipeline — Thinker · Planner · Coder · Reviewer, each independently configurable
• Dynamic Model Interpreter — automatically learns how to talk to any unknown model; stores results locally so calibration only runs once per model
• Multi-model support — correct request format and streaming parser for every Bedrock model family (Claude, Nova, Titan, Llama, DeepSeek, AI21, Cohere, Mistral)
• Reasoning model UI — DeepSeek R1 <think> blocks render as a collapsible Reasoning section
• Inline diff — AI-generated code appears as Keep/Discard hunks in your editor, Cursor-style
• Kernel Scheduler — route tasks to models by preset (Performance / Balanced / Custom / Named)
• Live model fetch — pulls both foundation models and inference profiles from your AWS account; shows purpose badges (💬 Chat, 🖼 Image, 📊 Embed) and filters non-text models automatically
• Manual model entry — add any model ID that isn't returned by the API
• Pinned models — pin frequently used models to the top of every agent dropdown with ★
• Codebase indexing — keyword + semantic search over your workspace for relevant context injection
• MCP support — connect any Model Context Protocol server for extended tool use
• Export — save chat sessions as .txt or .json; copy debug logs for sharing
• Cross-platform — Windows, macOS, Linux

Quick Start

Requirements

• VS Code 1.85 or later
• Node.js 18 or later
• An AWS account with Bedrock access and/or a local model server (Ollama, LM Studio)

Install

Search "Jinn Code" in the VS Code Extensions marketplace and click Install.

code --install-extension jinn-systems.jinn-code

Open the panel

Press Ctrl+Shift+J (Windows/Linux) or Cmd+Shift+J (macOS), or click the ✦ icon in the Activity Bar.

Setup

AWS Bedrock

1. Click the ⚙ gear icon → AWS Bedrock tab
2. Enter your AWS Access Key ID and Secret Access Key — stored securely in VS Code SecretStorage
3. Set your region (e.g. us-east-1) and Auth Mode
4. Click Fetch Models — pulls both foundation models and inference profiles from your account
5. Search or scroll the model list; check any models you use frequently to pin them (★)
6. Go to General tab → assign pinned or any fetched model to each agent

Can't find a model? Use the Add Model Manually input at the bottom of the Available Models card to add any model ID directly (e.g. us.anthropic.claude-sonnet-4-6).

Auth modes: | Mode | When to use | |------|-------------| | IAM / AWS CLI | Local dev with ~/.aws/credentials or instance role | | Access Key + Secret | Explicit credentials entered via the UI | | Endpoint URL | Bedrock Gateway or custom proxy |

Local Models (Ollama / LM Studio)

ollama pull llama3.1:8b && ollama serve

Settings → Local Models → enter endpoint → Fetch Models → Save Local Config.

Modes

Ask

Single-agent Q&A. The Thinker reasons through your question with full workspace context. No file writes.

Plan

The Planner produces a numbered implementation plan. Review it, then Execute Plan, Edit, or Discard.

Agent

Full pipeline: Planner → Thinker → Coder → Reviewer. Code appears as inline diffs — accept or discard each hunk individually or use Keep All / Discard All.

Agent Roles

Each agent is independently assigned any provider and model via Settings → General.

Kernel Scheduler

Routes each task to the right model. Switch presets in Settings → Kernel:

Presets derive model IDs from your General tab configuration — changing your agents automatically updates what Performance and Balanced use.

Dynamic Model Interpreter

When you select a model Jinn Code hasn't seen before, it automatically calibrates: sends a probe prompt in each supported request format, captures the raw streaming events, and uses Claude Haiku to identify the correct format and response structure. The result is stored at ~/.jinn-code/interpreters.json and reused on all future calls — calibration only happens once per model.

View learned models and trigger manual re-calibration in Settings → Interpreter DB.

Reasoning Models

DeepSeek R1 and other reasoning models emit a <think>...</think> chain-of-thought block before the answer. Jinn Code renders this as a collapsible Reasoning section (collapsed by default) above the main response so you can inspect the model's thinking without it cluttering the answer.

Inline Diff

When the Coder modifies an existing file:

• Green lines — AI's new content
• Red ghost text — what was there before
• ✔ Keep / ✖ Discard CodeLens above each hunk
• Keep All / Discard All bar in the chat UI

Clicking Keep or Discard auto-saves the file.

Export & Logs

Chat export (📤): Copy to Clipboard / Save as .txt / Save as .json

Log panel (📄): Filter by level, search by keyword, copy debug summary, export as .txt/.json. Red dot badge appears when an error occurs.

MCP Servers

"jinnCode.mcp": [
  {
    "name": "filesystem",
    "command": "npx",
    "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/project"]
  }
]

Privacy

Credentials are stored in VS Code's SecretStorage. No telemetry is collected. Prompts go directly to AWS Bedrock or your local model server — nothing passes through Jinn Studios servers.

License

MIT © 2026 Jinn Systems