🦀 v0.12.21 · beta · MIT licensed

Gajae Code

Encode intention. Decode software.

An external coding-agent harness for interviews, approved plans, and durable verification. Run gjc from the repository or worktree you choose, then integrate direct clients through the SDK WebSocket, Coordinator MCP, or ACP.

shape · interview act · reviewed plans prove · evidence
Download latest Read the Docs
Gajae Code autonomous coding-agent hero illustration
What is Gajae Code?

A focused harness, not a hidden plugin

Gajae Code (gjc) runs from the repository or worktree you choose, then gives the agent a deliberately small workflow: clarify before guessing, plan before mutation, and track every goal to completion with evidence.

It runs beside your existing tools — start it alongside Codex CLI, Claude Code, OpenCode, or Claw Code when you want structured planning, persistent evidence, an SDK integration, or an isolated worktree.

recommended workflow 
# run gjc in the repository or worktree you want to change
$ gjc


# inside the session: shape -> act -> prove
> /skill:deep-interview # clarify intent
> /skill:ralplan        # plan & critique
> gjc ultragoal create-goals --brief-file plan.md
> gjc ultragoal complete-goals
Latest stable · v0.12.21

Gajae Code v0.12.21

Published 2026-08-09. Release binaries and npm packages are available.

Read what’s new GitHub Release
Integration surfaces

Choose the client boundary that fits

Use the SDK WebSocket for direct control of one session, Coordinator MCP for multi-session automation, and ACP when an editor client is the right front end. Legacy Bridge and RPC modes are not current interfaces.

SDK & Migration

Build direct session clients on the loopback SDK WebSocket. The preserved route also explains migration away from historical Bridge and RPC ingress modes.

 computer-use desktop control — See. Click. Type. Control.

computer — desktop control (experimental)

The opt-in computer tool is available on macOS. Its nine actions are screenshot, click, double_click, move, drag, scroll, type, keypress, and wait; Screen Recording and Accessibility permissions are required.
The method

Shape · Act · Prove

One small loop, done well. No sprawling skill zoo — Gajae Code improves by making this method better.

01 — Shape

Interview before guessing

deep-interview turns vague requests into concrete requirements, boundaries, and explicit non-goals before any code is touched.

02 — Act

Plan before mutation

ralplan builds and critiques the implementation plan, then ultragoal carries the approved plan through execution, revision, verification, and evidence.

03 — Prove

Finish with evidence

ultragoal tracks goals, revisions, checks, and completion evidence so a finished task ships with a verifiable trail, not a hopeful claim.

Workflow surface

A path you can read

The default workflow is intentionally explicit. You always know what stage you are in.

deep-interview

Clarify ambiguous requirements before planning or code changes.

ralplan

Build and critique an implementation plan before mutation.

ultragoal

Track goals through execution, revision, verification, and evidence.

team (optional)

Coordinate focused workers when parallel execution is worth it.

Capabilities

Built for reviewable autonomy

Everything you need to run an agent you can actually trust with a repository.

Interview-first intent

Requirements are clarified up front with deep-interview, so the agent works from a shared, concrete understanding instead of a guess.

Reviewed plans

ralplan drafts and critiques the approach through planner, architect, and critic perspectives before a single file is mutated.

Durable evidence

ultragoal records goals, revisions, checks, and completion evidence under .gjc/ — a verifiable trail for every task.

Client integrations

Use the SDK WebSocket to control a direct session, Coordinator MCP to automate multiple sessions, or ACP to connect an editor client.

Worktree isolation

Create an isolated git worktree from a new branch for risky or reviewable work, keeping the primary checkout clean.

Coordinator MCP

Coordinate multiple sessions with bounded turns, structured questions, durable status, and reports through the Coordinator MCP surface.

Research mode (rlm)

Run gjc rlm for a notebook-style research session over a persistent Python kernel, with a hard-gated read-only toolset, a live .ipynb, and a synthesized report on exit.

Desktop control (experimental)

The opt-in computer tool is available on macOS with nine actions: screenshot, click, double-click, move, drag, scroll, type, keypress, and wait.
Architecture

A layered runtime

From CLI launch to model streams, every layer has a clear responsibility.

CLI & client surfacescoding-agent
The gjc binary registers commands (setup, deep-interview, ralplan, ultragoal, team, rlm) and supports interactive sessions, print output, SDK WebSocket clients, Coordinator MCP automation, and ACP editor clients.
Agent runtimeagent
A stateful turn loop transforms context, calls the model stream, executes tool calls, appends results, and emits lifecycle events.
Provider boundaryai
A model registry, provider implementations, auth, streaming, usage accounting, and retry/overflow handling sit behind one clean interface.
Toolsread · edit · bash · lsp
A built-in tool registry for file, code, and runtime operations — reading, editing, AST tools, search, LSP, the browser, subagents, and more.
Native & TUInatives · tui
Rust N-API helpers accelerate search, diff, syntax, and shell work; the terminal UI framework renders the interactive experience.

Explore the architecture →

Get started

Install the latest GJC binary

No Bun required. We detect your platform and link directly to the newest standalone release.

# 1 · choose your platform
View latest downloads

Choose the platform and architecture for this machine.

Need another build? View every latest release asset. Read the full installation guide. Standalone binary source availability remains on GitHub.

Workflow surface

Four skills, four role agents

A small, fixed surface — no default skill sprawl.

Workflow skills

SkillWhat it does
deep-interviewClarifies ambiguous requirements before planning or code changes.
ralplanBuilds and critiques an implementation plan before mutation.
ultragoalTracks goals through execution, revision, verification, and evidence.
teamCoordinates focused workers when parallel execution is worth it.

Role agents

AgentWhat it does
executorBounded implementation, fixes, and refactors.
architectRead-only architecture and code-review assessment.
plannerRead-only sequencing and acceptance criteria.
criticRead-only plan critique and actionability review.

Skills reference →

Safety & proof

Trust comes from evidence

Autonomy is only useful when the result is reviewable. Gajae Code is built to leave a trail.

Plan-before-mutation

Approaches are reviewed and critiqued before any file changes, reducing wasted runs and surprise edits.

Completion evidence

Goals, revisions, checks, and verification results are recorded as durable artifacts you can read after the fact.

Isolated worktrees

Risky work runs in a dedicated git worktree that gjc creates from a new branch, so your primary checkout stays clean and the diff stays scoped.

Scoped client control

The SDK WebSocket, Coordinator MCP, and ACP expose explicit client boundaries rather than historical Bridge or RPC ingress modes.

How receipts & proof work →

Documentation

Dive deeper

Guides and references for every layer of Gajae Code.

Getting Started

Install, launch, and run your first shape → act → prove loop.

Architecture

How the CLI, agent loop, provider boundary, and tools fit together.

The Harness

Running gjc as an external runner beside your other agents.

SDK & Migration

Direct session control uses the SDK WebSocket; Coordinator MCP automates multiple sessions; ACP serves editor clients.

Skills & Agents

The four workflow skills and four role agents in detail.

Receipts & Proof

The evidence model behind durable, verifiable completion.

Give your agent a workflow it can prove

Install Gajae Code, run it beside the tools you already use, and ship work with a trail you can read.

Get Started Star on GitHub Join Discord