jcode: A Rust-Powered Coding Agent Harness with MCP & Multi-Provider Support
文章目录
- Multi-Provider Support: Connect to Claude, OpenAI, GitHub Copilot, Gemini, and any OpenAI-compatible API endpoint. Switch between providers on the fly with a single /model command. Native MCP Integration: jcode speaks MCP natively. Configure your MCP servers via ~/.jcode/mcp.json or project-local .jcode/mcp.json, and your AI agents can use them as tools instantly. Autonomous Swarm Mode: Spin up multiple AI agents that coordinate autonomously on complex tasks. Useful for exploratory research, parallel debugging, or handling multi-step refactors where a single agent would lose context.
- Here's what developers are discussing on GitHub Issues: 🔧 Issue #44 — GitHub Copilot Auth on WSL2 (5 comments) Reported by @wahyuabrory: "I've logged into Copilot, but the available models aren't the same as in OpenCode or the Copilot CLI. For example, 5.4 mini isn't listed. There's also a serious performance issue in the WSL environment — it lags heavily, even when just selecting a model, and sometimes freezes entirely." Maintainer response: "Try v9.8, I made a ton of performance improvement changes, as well as a bunch of WSL-specific ones." This thread highlights an important consideration for Windows/WSL users: not all GitHub Copilot models are available via jcode's OAuth flow, and performance in WSL environments has historically been an issue. Upgrading to the latest version is strongly recommended if you're on WSL2. 🛠️ Issue #58 — Windows Port: Named Pipes, Console Suppression & CI (4 comments) Community contributor @h5kk: "Pushed a follow-up commit: windows: suppress new console window when bash tool spawns cmd.exe. When the jcode server runs detached, it has no inherited console. Every Bash tool call invokes cmd.exe /C <command>, and Windows pops up a new console window for each one — which is disruptive." Another follow-up extracted this into a reusable platform::suppress_child_console helper, applied across all server-side spawns. The result: Windows users get a clean, popup-free experience matching Linux/macOS behavior. 📄 Issue #48 — Config Files for Offline Use Not Documented (3 comments) Reported by @aixiansheng: "I tried using the OpenAI compact flags, then typed /login and selected that provider, but it printed: Warning: ignoring invalid JCODE_OPENAI_COMPAT_API_BASE 'http://my.models.net:8009/v1'. Use https://… (or http://localhost). Also, my MCP tools are HTTP servers — there's no local command for them to run." Maintainer: "jcode does support config-file based setup today. The confusing part was that it was under-documented. Use ~/.jcode/mcp.json for global MCP config, or .jcode/mcp.json for project-local." This exchange reveals a key usability gap: self-hosted vLLM users need a clear path to configure HTTP(S) MCP servers and custom API bases. The documentation has since improved, but it's worth noting that jcode does support this — it's just not immediately obvious.
- jcode is a compelling choice for developers who want a lightweight, terminal-first coding agent harness with first-class Rust performance. Its multi-provider flexibility and MCP support make it versatile, while the active community (evidenced by high-quality cross-platform PRs like the Windows port) suggests it's maturing rapidly. The autonomous swarm mode is experimental but promising for advanced use cases. If you're tired of heavy IDE extensions and want something that just works in your terminal, jcode is worth a try. Category: Developer Tools Language: Rust 🦀 License: Open Source ⭐ Stars: 2,560+ @1jehuang · View on GitHub →
jcode is a Rust-powered, terminal-native coding agent harness that lets you connect any LLM provider — Claude, OpenAI, Copilot, or your own vLLM servers — through a unified TUI. It was created as a developer-friendly alternative to heavy IDE plugins, offering deep MCP (Model Context Protocol) integration, autonomous swarm mode, and a scriptable CLI workflow.
Whether you're building a personal coding assistant or wiring up a fleet of AI agents to work in parallel, jcode gives you the infrastructure without the bloat. It's actively maintained, supports both cloud and self-hosted models, and runs entirely in the terminal — no browser required.
- Multi-Provider Support: Connect to Claude, OpenAI, GitHub Copilot, Gemini, and any OpenAI-compatible API endpoint. Switch between providers on the fly with a single
/model command.
- Native MCP Integration: jcode speaks MCP natively. Configure your MCP servers via
~/.jcode/mcp.json or project-local .jcode/mcp.json, and your AI agents can use them as tools instantly.
- Autonomous Swarm Mode: Spin up multiple AI agents that coordinate autonomously on complex tasks. Useful for exploratory research, parallel debugging, or handling multi-step refactors where a single agent would lose context.
/model command.~/.jcode/mcp.json or project-local .jcode/mcp.json, and your AI agents can use them as tools instantly.Here's what developers are discussing on GitHub Issues:
🔧 Issue #44 — GitHub Copilot Auth on WSL2 (5 comments)
Reported by @wahyuabrory:
"I've logged into Copilot, but the available models aren't the same as in OpenCode or the Copilot CLI. For example, 5.4 mini isn't listed. There's also a serious performance issue in the WSL environment — it lags heavily, even when just selecting a model, and sometimes freezes entirely."
Maintainer response: "Try v9.8, I made a ton of performance improvement changes, as well as a bunch of WSL-specific ones."
This thread highlights an important consideration for Windows/WSL users: not all GitHub Copilot models are available via jcode's OAuth flow, and performance in WSL environments has historically been an issue. Upgrading to the latest version is strongly recommended if you're on WSL2.
🛠️ Issue #58 — Windows Port: Named Pipes, Console Suppression & CI (4 comments)
Community contributor @h5kk:
"Pushed a follow-up commit: windows: suppress new console window when bash tool spawns cmd.exe. When the jcode server runs detached, it has no inherited console. Every Bash tool call invokes
cmd.exe /C <command>, and Windows pops up a new console window for each one — which is disruptive."
Another follow-up extracted this into a reusable platform::suppress_child_console helper, applied across all server-side spawns. The result: Windows users get a clean, popup-free experience matching Linux/macOS behavior.
📄 Issue #48 — Config Files for Offline Use Not Documented (3 comments)
Reported by @aixiansheng:
"I tried using the OpenAI compact flags, then typed
/loginand selected that provider, but it printed: Warning: ignoring invalid JCODE_OPENAI_COMPAT_API_BASE 'http://my.models.net:8009/v1'. Use https://… (or http://localhost). Also, my MCP tools are HTTP servers — there's no local command for them to run."
Maintainer: "jcode does support config-file based setup today. The confusing part was that it was under-documented. Use ~/.jcode/mcp.json for global MCP config, or .jcode/mcp.json for project-local."
This exchange reveals a key usability gap: self-hosted vLLM users need a clear path to configure HTTP(S) MCP servers and custom API bases. The documentation has since improved, but it's worth noting that jcode does support this — it's just not immediately obvious.
jcode is a compelling choice for developers who want a lightweight, terminal-first coding agent harness with first-class Rust performance. Its multi-provider flexibility and MCP support make it versatile, while the active community (evidenced by high-quality cross-platform PRs like the Windows port) suggests it's maturing rapidly. The autonomous swarm mode is experimental but promising for advanced use cases. If you're tired of heavy IDE extensions and want something that just works in your terminal, jcode is worth a try.
Category: Developer Tools
Language: Rust 🦀
License: Open Source
⭐ Stars: 2,560+
@1jehuang · View on GitHub →
微信扫一扫,分享到朋友圈
