jcode

jcode is an open-source terminal AI coding harness built to be resource-efficient and fast. The project positions itself as the most RAM-efficient and most intelligent harness, and the README backs that up with benchmark tables comparing jcode against tools such as pi, Codex CLI, OpenCode, GitHub Copilot CLI, Cursor Agent, Claude Code, and Antigravity CLI. Reported figures include roughly 27.8 MB PSS for a single session with local embedding disabled, about 14 ms time to first frame, and about 48.7 ms time to first input, with per-session memory scaling around 10 MB. Install is a one-line curl script on macOS/Linux and a PowerShell one-liner on Windows 11, and the tool ships for Linux, macOS, and Windows under an MIT license. The core experience is a TUI with a chat area, side panels, inline Mermaid diagram rendering, info widgets that only occupy negative screen space, a custom scrollback implementation, and a companion terminal project called Handterm that adds a native scroll API. Rendering is claimed to run at over a thousand fps, and the project includes a custom Mermaid rendering library (mermaid-rs-renderer) with no browser or TypeScript dependency. Alignment can be switched between left and centered, and emoji can be disabled globally via config or an environment variable. A distinctive feature is the agent memory system. Each turn is embedded as a semantic vector and queried against a graph of memories using cosine similarity. Hits are injected into the conversation, optionally verified by a memory sideagent that can do additional retrieval work. Memories are extracted periodically based on semantic drift, turn counts, or session end, and consolidated in an ambient mode that reorganizes entries and checks for staleness and conflicts. Explicit memory tools let the agent actively search or store memories, and session search provides traditional RAG over prior sessions. The Swarm feature lets multiple agents run in the same repository under a shared server. When one agent edits a file another has read, the server notifies the other agent, which can ignore the change or inspect the diff to avoid conflicts. Agents can message each other via DM, broadcast, or repo-scoped channels, and they can autonomously spawn their own swarms, turning the main agent into a coordinator and spawned agents into workers. Group messaging channels, completion statuses, and coordination are managed automatically, and this can run headlessly or headed. Provider support is broad. Built-in OAuth login flows cover Claude, OpenAI/ChatGPT/Codex, Google Gemini, GitHub Copilot, Azure OpenAI, Alibaba Cloud Coding Plan, Fireworks, Novita AI, MiniMax, Meta Model API/Muse, LM Studio, and Ollama, plus a custom OpenAI-compatible endpoint. Native OpenAI providers use Responses WebSocket v2 with opportunistic background prewarming and HTTPS fallback. OpenAI-compatible providers include built-in named profiles for services such as OpenRouter, DeepSeek, Moonshot AI, Baseten, Hugging Face, Nebius, Scaleway, and others, and arbitrary endpoints can be added via a one-shot provider profile command that writes to ~/.jcode/config.toml and stores secrets in a private app config directory. Anthropic Messages-compatible gateways are supported with bearer, custom-header, or no authentication and gateway-specific headers. Local runtimes like Ollama and LM Studio are supported with streaming chat completions, function/tool calling, and OpenAI-style image content for vision-capable local models. Configuration options include per-model context windows, extra_body fields for non-standard request bodies, stream idle timeouts that scale with reasoning effort, and reasoning effort controls. The project also provides an SDK, documentation, benchmarks, and a Discord community. Updates can be performed in the TUI with /update or from the terminal with jcode update, with a stable channel by default and a main channel for source-branch updates. Development builds compare the running binary's Git commit against the release tag and stop rather than risk a downgrade if ancestry cannot be verified. Overall, jcode targets developers who want a fast, memory-light terminal AI coding environment with persistent memory, multi-agent collaboration, and flexible model-provider configuration.