jcode

jcode

Website · Docs · OpenAI Ultrafast · SDK · Benchmarks · Screenshots · Install · Contributing

Screenshots

Browse the interactive gallery on jcode.sh.

Installation

# macOS & Linux
curl -fsSL https://jcode.sh/install | bash

Windows installation instructions.

Performance & Resource Efficiency

jcode is built to be as performant and resource efficient as possible. Every metric is optimized to the bone, which is important for scaling multi-session workflows. Here we sample a few metrics to show the difference: RAM usage and boot up.

RAM comparison

Headless sessions (swarm workers)

Swarm workers run headless, so this is the number that matters when you fan out agents. Each session completed 5 real model turns (file listing, file read, repo search, summary, reply) with tool calls, then total PSS of every process was measured. jcode sessions share one daemon; Claude Code runs one claude -p --input-format stream-json process per session. Both used claude-sonnet-4-6.

Concurrent headless sessions	jcode	Claude Code	Comparison
1	32.6 MB	261.0 MB	8.0× less RAM
5	51.0 MB	908.6 MB	17.8× less RAM
10	66.7 MB	1749.7 MB	26.2× less RAM
20	90.6 MB	3376.8 MB	37.3× less RAM
Each additional session	~3.1 MB	~164 MB	~54× less RAM

Measured 2026-09-29 on Linux with jcode v0.89.19-dev (default build, local embeddings not compiled in) and Claude Code 2.1.267. Reproduce with python3 scripts/bench_headless_memory.py.

Time to first frame

Tool	Time to first frame	Range	Comparison
jcode	14.0 ms	10.1–19.3 ms	baseline
Antigravity CLI	383.5 ms	363.1–415.4 ms	27.4× slower
pi	590.7 ms	369.6–934.8 ms	42.2× slower
Codex CLI	882.8 ms	742.3–1640.9 ms	63.1× slower
OpenCode	1035.9 ms	922.5–1104.4 ms	74.0× slower
GitHub Copilot CLI	1518.6 ms	1357.4–1826.8 ms	108.5× slower
Cursor Agent	1949.7 ms	1711.0–2104.8 ms	139.3× slower
Claude Code	3436.9 ms	2032.7–8927.2 ms	245.5× slower

Measured on this Linux machine across 10 interactive PTY launches.

Time to first input

(time until typed probe text appears on the rendered screen; Antigravity uses its internal input-ready log marker because the sign-in screen suppresses probe echo.)

Tool	Time to first input	Range	Comparison
jcode	48.7 ms	30.3–62.7 ms	baseline
Antigravity CLI	383.7 ms	363.4–415.7 ms	7.9× slower
pi	596.4 ms	373.9–955.2 ms	12.2× slower
Codex CLI	905.8 ms	760.1–1675.7 ms	18.6× slower
OpenCode	1047.9 ms	931.1–1116.9 ms	21.5× slower
GitHub Copilot CLI	1583.4 ms	1422.8–1880.0 ms	32.5× slower
Cursor Agent	1978.7 ms	1727.3–2130.0 ms	40.6× slower
Claude Code	3512.8 ms	2137.4–9002.0 ms	72.2× slower

Measured on this Linux machine across 10 interactive PTY launches. Antigravity CLI was unauthenticated for this run; its sign-in screen rendered normally and emitted an internal CLI ready for user input marker, but did not echo the typed probe.

Additional clients / memory scaling

Tool	Extra PSS per added session	Comparison
jcode (local embedding off)	~9.9 MB	baseline
jcode	~10.4 MB	1.1× more RAM
pi	~76.5 MB	7.7× more RAM
Codex CLI	~21.6 MB	2.2× more RAM
OpenCode	~318.4 MB	32.2× more RAM
GitHub Copilot CLI	~158.1 MB	16.0× more RAM
Cursor Agent	~157.5 MB	15.9× more RAM
Claude Code	~212.7 MB	21.5× more RAM
Antigravity CLI	~86.4 MB	8.7× more RAM

• jcode v0.9.1888-dev (be386f2)
• pi 0.62.0
• codex-cli 0.120.0
• opencode 1.0.203
• GitHub Copilot CLI 1.0.24 for the 1-session rerun, GitHub Copilot CLI 1.0.27 for the 10-session rerun
• Cursor Agent 2026.04.08-a41fba1
• Claude Code 2.1.86 (Claude Code)
• Antigravity CLI 1.0.0

Swarm

Spawn two or more agents in the same repo, and they will automatically be managed by the server to allow native collaboration. When agent A edits a file that agent B has read (code shifting under its feet), the server notifies agent B. Agent B can ignore it if it is not relevant, or it can check the diff to make sure that it doesn't conflict. Each agent has messaging abilities, capable of DMing just one agent, broadcasting to all other agents hosted by the server, or just agents working in that repo. This allows you to spawn multiple sessions in the same repo, and have all conflicts automatically resolved.

Agents are also able to spawn their own swarms autonomously. They have a swarm tool which allows them to spawn in their own teamates to accomplish tasks in parallel. Doing so turns the main agent into a coordinator and the spawned agents into workers. Groups of agents, their messaging channels, their completion statuses, etc are all automatically managed. This can be done headlessly or headed.

Swarm modes keep root reasoning separate from worker effort. Configure each mode in ~/.jcode/config.toml:

[agents]
swarm_root_effort = "low"        # /effort swarm
swarm_deep_root_effort = "high"  # /effort swarm-deep

Both default to max. Accepted levels are none, minimal, low, medium, high, xhigh, and max, mapped to the provider's supported range. The effort switcher shows the configured root level. These settings do not change worker swarm_effort. Environment overrides are JCODE_SWARM_ROOT_EFFORT and JCODE_SWARM_DEEP_ROOT_EFFORT.

OAuth and Providers

Supports a wide range of providers, including Claude OAuth and OpenAI Codex OAuth.

Customizability / Self-Dev

Jcode is inventing a new form of customizability. One that doesn't limit you to what a plugin or extension can do. Tell your jcode agent to enter self dev mode, and it will start modifying its own source code. Jcode is optimized to iterate on itself. There is significant infrastructure around self development, which allows it to edit, build, and test its own source code, then reload its own binary and continue work in your (potentially many) sessions, fully automatically.

It is recommended that you use a frontier model for this. The jcode codebase is not a simple one, and weaker models can make subtle, breaking changes. GPT 5.5 or the latest available frontier model works well.

Misc.

The devil is in the details. There are many undocumented optimizations and niceties that jcode implements. Some examples:

Anthropic's Claude cache goes cold after 5 minutes. If you initiate Claude after these 5 minutes, you have a cache miss, potentially costing you lots of tokens. The ui warns you when the cache went cold, and notfies you if there was an unexpected cache miss.

jcode comes with instructions on how to set up Firefox Agent Bridge. Ask you agent to set it up, and then you will have browser automation in jcode as well.

Agent grep is a grep tool I made for the jcode agent. It adds file structure information (ie the list of functions, their displacement, etc) to the grep return, so that the agent can infer more of what the file doesn without actually reading the file. It also implements a harness-level integration that adaptively truncates returns based on what the agent has already seen. This saves on context a lot.

Inputs are by default interleaved with the working agent. It sends the input as soon as it safely can without breaking the KV cache. Submit with shift enter instead, and it will send a queue send, and wait for the agent to fully finish its turn before sending.

Resume sessions from different harnesses. Claude code broke on you? Resume the session from jcode and continue where you left off. Session resume is supported for codex, claude code, opencode, and pi.

Skills are not all loaded on startup. The conversation is embedded as a semantic vector, and will automatically inject a skill if there is an embedding hit similar to memories. The agent has a skill tool for you to manually activate a skill at anytime. You may also activate via slash commands.

Browser Automation

The built-in browser tool lets agents navigate pages, fill forms, upload files, and capture screenshots through Firefox Agent Bridge. Ask your agent to set it up and handle browser tasks directly.

See Browser Provider Protocol for details.