# jcode

> An open source terminal coding agent written in Rust, featuring agent swarms, semantic memory, and extreme resource efficiency for multi-session AI workflows.

jcode is an open source terminal coding agent written in Rust by solo founder Jeremy Huang, a YC Summer 2026 batch company under Solo Systems. It is MIT licensed and available on macOS, Linux, and Windows, installable via a single curl command or Homebrew. The project positions itself as a harness-level research platform, publishing all benchmarks and agent transcripts publicly.

## What It Is

jcode is a CLI-based AI coding agent harness — the layer between the user and the underlying language model that orchestrates tool calls, memory, context management, and multi-agent coordination. Unlike IDE plugins or web-based coding assistants, jcode runs entirely in the terminal as a TUI (text user interface) and is designed to be spawned in many parallel sessions simultaneously. It supports bring-your-own API keys, OAuth with existing Claude and OpenAI subscriptions, and local models via Ollama or LM Studio, as well as hosted token plans for zero-setup access.

## Harness Architecture and Intelligence Features

jcode's core thesis, stated on its homepage, is that "the next leaps in agent capabilities will come from the harness level." Several concrete mechanisms implement this:

- **Confidence stepping**: The todo tool tracks per-task confidence scores at assignment and completion. When the harness detects a large confidence spike (low at assignment, suddenly 100% at done), it forces the agent to re-verify rather than accept the claim. According to the project's Terminal-Bench 2.1 results, this raised the pass rate for trials that finished in time from 88% to 92%.
- **Hill-climbable goals**: Every agent goal receives a hill-climbability score (0–100) measuring how quantifiable and iterable its progress is. Low-scoring goals trigger a harness-level pushback asking the agent to reframe the task into a verifiable objective.
- **Auto-poke**: When a turn ends with incomplete todos, the harness automatically re-prompts the model to continue rather than letting it declare early victory.
- **Agent swarms**: Multiple jcode agents can work in the same repository simultaneously. The server notifies agents when a file they have read is modified by another agent, and agents can message each other directly or broadcast to the group.

## Performance and Resource Efficiency

jcode is written in Rust and optimized for low per-session overhead, which the project frames as essential for scaling swarms. According to vendor-published benchmarks on the project's homepage and README:

- Extra proportional memory (PSS) per additional session: ~10.4 MB for jcode vs ~212.7 MB for Claude Code
- Time to first frame: 14.0 ms for jcode vs 3,436.9 ms for Claude Code
- Time to first input: 48.7 ms for jcode vs 3,512.8 ms for Claude Code

These figures are presented by the project as self-measured on a single Linux machine and should be treated as vendor-published claims.

## Context Engineering and MCP Integration

jcode uses an append-only conversation structure to keep the model's KV cache warm across turns. Tool schemas are served from an on-disk cache so they never invalidate the prompt prefix. MCP tools are advertised from a schema cache at session start, with actual server connections happening lazily in the background, so neither startup nor the prompt cache is blocked by slow MCP handshakes. Background tasks can be started with `run_in_background`, allowing the agent to monitor builds or test runs without polling loops.

## Update: v0.62.1 and Active Development

The latest tagged release is v0.62.1, published July 30, 2026. Recent releases show rapid iteration:

- **v0.61.0** (July 27): `/remote` command for reaching a running session from another machine over WebSocket; destructive shell command gate; desktop preview improvements.
- **v0.58.0** (July 25): OpenAI max-effort reasoning timeout fix; `/model` picker with fuzzy matching.
- **v0.57.0** (July 25): Claude Opus 5 added to the built-in Anthropic catalog.
- **v0.56.0** (July 24): Redesigned TUI header; versioned harness API socket for external apps to drive jcode sessions.

The project's base system prompt has shrunk 73% since v0.1 (from 2,476 tokens to 659 base tokens), which the project tracks as a signal that frontier models require less scaffolding over time. The GitHub repository shows 13,523 stars and 1,485 forks as of the data snapshot.

## Features
- Open source MIT licensed terminal coding agent
- Agent swarms with multi-agent coordination in the same repository
- Semantic memory with vector embeddings and cosine similarity recall
- Confidence stepping to prevent false task completion claims
- Hill-climbable goal scoring to push agents toward verifiable objectives
- Auto-poke persistence to keep agents working until todos are complete
- Append-only context engineering for high KV cache hit rates
- MCP tool support with on-disk schema cache for zero-startup-delay
- Background task management with live progress rendering in TUI
- Self dev mode: agent modifies and reloads its own source code
- Bring your own API keys or OAuth with Claude/OpenAI subscriptions
- Local model support via Ollama and LM Studio
- Browser automation via Firefox Agent Bridge
- Inline LaTeX, Mermaid diagram, and image rendering in TUI
- Session resume from Claude Code, Codex, OpenCode, and pi
- Headless jcode run for non-interactive single-prompt execution
- Skills system with semantic injection based on conversation context
- Multi-account switching per provider
- Ambient mode for background memory consolidation
- jcode bench: uncontaminatable optimization benchmarks

## Integrations
Anthropic Claude, OpenAI / ChatGPT, Google Gemini, GitHub Copilot, Azure OpenAI, Ollama, LM Studio, OpenRouter, DeepSeek, AWS Bedrock, Fireworks, MiniMax, Groq, Mistral, Perplexity, Together AI, DeepInfra, xAI, Cerebras, Hugging Face, Nebius, MCP (Model Context Protocol), Firefox Agent Bridge, Tailscale (iOS remote access), Stripe (subscription billing)

## Platforms
WINDOWS, MACOS, LINUX, IOS, API, CLI

## Pricing
Open Source, Free tier available

## Version
v0.62.1

## Links
- Website: https://jcode.sh
- Documentation: https://jcode.sh/docs
- Repository: https://github.com/1jehuang/jcode
- EveryDev.ai: https://www.everydev.ai/tools/jcode