Jcode is a terminal coding agent written in Rust, designed to keep memory use low so that many agent sessions can run on one machine. Sessions share a single daemon, and headless sessions act as swarm workers. It runs on macOS, Linux, and Windows, with a desktop build in beta.
Measured memory (PSS) with Claude Sonnet 4.6 as the model:
| Tool | 1 active session | 10 active sessions |
|---|---|---|
| Jcode (local embedding off) | 27.8 MB | 117.0 MB |
| Jcode | 167.1 MB | 260.8 MB |
| Codex CLI | 140.0 MB | 334.8 MB |
| Pi | 144.4 MB | 833.0 MB |
| Cursor Agent | 214.9 MB | 1632.4 MB |
| Antigravity CLI | 243.7 MB | 1021.2 MB |
| GitHub Copilot CLI | 333.3 MB | 1756.5 MB |
| OpenCode | 371.5 MB | 3237.2 MB |
| Claude Code | 386.6 MB | 2300.6 MB |
In a swarm, two or more agents work in the same repository under the server's management. When one agent edits a file another has read, the server notifies the second agent, and agents can message each other directly. An agent can also spawn its own swarm with a swarm tool, becoming a coordinator over worker agents, headless or headed.
Jcode embeds each turn as a semantic vector and queries a memory graph for related entries on every turn, optionally through a memory sideagent that checks relevance before injecting results. Memories are extracted periodically, consolidated in an ambient mode, and searchable through explicit memory tools. Skills load the same way, injected when the conversation matches them rather than all at startup.
Jcode signs in through subscription OAuth flows for Claude, OpenAI (ChatGPT and Codex), Google Gemini, and GitHub Copilot, and connects to direct API providers, aggregators, and local servers such as Ollama and LM Studio. It can resume sessions started in Codex, Claude Code, OpenCode, and Pi. A self-dev mode lets the agent edit, build, test, and reload its own source code.
Features
- Low memory footprint: sessions share one daemon, with 117 MB total measured for 10 active sessions with local embedding off
- Swarms: multiple agents in one repository get file-conflict notifications and direct messaging, and an agent can spawn its own workers
- Semantic memory: turns are embedded and matched against a memory graph, with periodic extraction, consolidation, and explicit memory tools
- Session search: retrieval over previous sessions for context from past work
- Cross-harness resume: continues sessions started in Codex, Claude Code, OpenCode, or Pi
- Subscription logins: OAuth flows for Claude, ChatGPT and Codex, Gemini, and GitHub Copilot, plus multi-account switching
- Side panel: shows files, agent-written notes, or diffs that update in real time, with inline Mermaid diagram rendering
- Agent grep: a grep tool that adds file structure such as function lists to results and truncates output based on what the agent has already seen
- Browser tool: built-in browser control through Firefox Agent Bridge, with actions such as open, click, type, and form filling
- Embedded skills: skills are injected when the conversation matches them, or activated with a skill tool or slash command
- Self-dev mode: the agent modifies, builds, tests, and reloads its own source code
- Cache warnings: the UI warns when the Claude prompt cache has gone cold and flags unexpected cache misses
