jcode

 ◆ Found it — the goroutine in handleConnection() is never joined.
   I'll patch it now.

   ⚙ Tool  edit   path=server.go

   ╭─────────────────────────────────────────────────────╮
   │  - go handle(conn)                                  │
   │  + wg.Add(1)                                        │
   │  + go func() { defer wg.Done(); handle(conn) }()   │
   ╰─────────────────────────────────────────────────────╯
      ✓ Edit applied

──────────────────────────────────────────────────────────
 > _
──────────────────────────────────────────────────────────
  Agent │ Model: openai / gpt-4o │ Approve: Ask │ [████░░░░░░] 2% │ MCP: 2/5

Why jcode?

Install

Quick Install (recommended)

curl -fsSL https://raw.githubusercontent.com/cnjack/jcode/main/script/install.sh | sh

From Source

Requires Go 1.22+ and Node.js + pnpm.

git clone https://github.com/cnjack/jcode.git
cd jcode
make install

Update

jcode update

First launch creates ~/.jcode/config.json with a setup wizard. Run jcode doctor to verify model & MCP connectivity.

Features

Core Agent Loop

Describe a task in plain English. The agent reads your codebase, writes surgical edits, runs commands, and reports every step — no black boxes.

🤝 Agent Teams

Spawn multiple AI teammates that work in parallel, each with independent tools, conversation history, and environment. The lead agent coordinates; teammates idle until they receive an explicit message.

 You › Create a team and spawn a backend developer

   ⚙ Tool  team_create   team_name=dev-team
      ✓ Done

   ⚙ Tool  team_spawn   name=backend  prompt="Senior Go backend developer"
      ✓ Done

 You › Send backend a task

   ⚙ Tool  team_send_message   to=backend  message="Add pagination to /users"
      ✓ Message sent to @backend

Switch between agent views with Shift+↑/↓ and see live status in the team panel:

  ╭ Team: dev-team (2) ───────────────────────────╮
  │  ● Main (leader)                               │
  │  ○ ⟳ @architect 1m32s [3 tools]               │
  │  ○ ◇ @backend   0m45s                          │
  │                                                 │
  │  shift+↑/↓: switch agent | esc: back to leader │
  ╰─────────────────────────────────────────────────╯

• Persistent mailbox — session-scoped, file-based message passing between teammates
• Per-agent approval — mutating tool calls surface an approval dialog tagged with the teammate's name and color
• Independent conversations — each agent has its own full chat history, tool calls, and markdown-rendered output
• Agent types — explore (read-only), general (full tools), coordinator (can spawn sub-teams)

🌐 SSH — work on any machine

Type /ssh user@host and every tool runs transparently on the remote host. No agents, no tunnels, no extra setup.

 You › /ssh deploy@10.0.1.5:/var/www/app

   ✓ SSH  Connected · linux/amd64

 You › why is nginx restarting?

   ⚙ Tool  execute  [deploy@10.0.1.5]  docker logs app-nginx-1 --tail 20

   ╭─────────────────────────────────────────────────────╮
   │  nginx: [emerg] bind() to 0.0.0.0:80 failed        │
   │  (98: Address already in use)                       │
   ╰─────────────────────────────────────────────────────╯

 ◆ Port 80 is already taken. Let me find what's holding it.

Save connections as named aliases and jump between hosts with /ssh:

  ┌─────── /ssh ────────────────────────────────────┐
  │  > 🔗 prod        deploy@10.0.1.5:/var/www/app  │
  │    🔗 staging      ci@10.0.1.8:/srv/staging      │
  │    ➕ Connect New SSH                             │
  └─────────────────────────────────────────────────┘

📋 Plan Mode

Press Ctrl+P to enter Plan Mode. The agent explores your codebase read-only and presents a structured plan before touching any file. Review, approve or reject with feedback — then let it execute step by step.

  Plan │ Model: openai / gpt-4o │ Approve: Ask │ [██░░░░░░░░] 12%

🔌 MCP Integration

Connect any MCP-compatible server — stdio, HTTP, or SSE — and its tools merge with the built-ins. Auto-reconnect with exponential backoff. Status shown live in the status bar.

{
  "mcp_servers": {
    "github": { "type": "stdio", "command": "gh-mcp" },
    "db":     { "type": "http",  "url": "http://localhost:3001/mcp" }
  }
}

  Agent │ Model: openai / gpt-4o │ Approve: Ask │ [████░░░░░░] 2% │ MCP: 2/5

💰 Token Usage & Budget Control

Real-time context window tracking with a color-coded progress bar in the status bar:

Set cost guardrails in config.json:

{
  "budget": {
    "max_cost_per_session": 5.00,
    "warning_threshold": 0.8
  }
}

The agent receives in-context warnings when nearing limits and stops if the budget is exceeded. Model pricing is auto-fetched from models.dev.

🧠 Context Management

• Auto-compaction — when the context window fills up, older conversation is summarized while preserving the most recent messages
• Manual compaction — type /compact anytime to free up context
• Smart prompt caching — reduces redundant prompt computation across turns
• AGENTS.md support — global (~/.jcode/AGENTS.md), project-level, and local (.local.md, git-ignored) agent instructions with @include directives

🛠 Skills

Domain-specific skills loaded on demand. Built-in skills include PR review and security review. Add your own skill packs to ~/.jcode/skills/ or <project>/.jcode/skills/.

Skills register as slash commands — type /review-pr or /security-review to activate.

⚡ Subagents & Background Tasks

• Subagents — delegate subtasks to independent child agents (explore, general, or coordinator type) with up to 3 levels of nesting
• Background commands — long-running builds/tests run async; check with /bg or the check_background tool
• Status tracking — Bg: 3 running shown in status bar; task IDs for programmatic access

📼 Session Resume

Every conversation is recorded as JSONL. Resume any past session:

  ┌──────────────── Resume Session ─────────────────┐
  │  > 2026-03-12  gpt-4o      fix nginx crash       │
  │    2026-03-11  gpt-4o      refactor auth module  │
  │    2026-03-10  claude-3.5  add pagination logic  │
  └─────────────────────────────────────────────────┘

jcode sessions          # list sessions
jcode --resume <UUID>   # pick up where you left off

🌐 Web Interface

Start a browser-based UI with jcode web. Chat interface, file browser, built-in terminal, and full agent control — all accessible from http://localhost:8080.

🧭 Context Awareness

At startup the agent automatically detects:

• Git branch, dirty status, last commit
• Project type (Go, Python, JS, Rust, Java, …)
• Directory structure
• SSH environment labels
• Available skills

No manual configuration needed — the agent adapts to your project.

Keyboard Shortcuts

Slash Commands

Configuration

Config lives at ~/.jcode/config.json. Key sections:

jcode doctor    # verify model + MCP connectivity
jcode version   # show version, commit, build time
jcode update    # update to latest version

Documentation

📖 Full documentation is available at cnjack.github.io/jcode

Changelog

See CHANGELOG.md for release notes and version history.

License

MIT