English | 한국어 | 中文 | 日本語

G A J A E - C O D E

Encode intention. Decode software.
_{The coding agent that runs on the plan you already pay for — and answers to your phone.}

🖥️ Explore Gajae-Code App → Desktop & self-hosted web GUI
_{Experimental, community-built third-party project — not an official first-party Gajae-Code app.}

Quick Start · Why · Coding Plans · Phone · Workflow · Token Diet · Controllers · Agent Shells · Docs

Connect a supported subscription or API key, choose a planning workflow when needed, execute with evidence — and answer the agent's questions from your terminal, your phone, or your own bot.

Gajae-Code (gjc) is an external coding-agent harness: drop it into any repository or worktree. Supported subscription routes use your provider's plan; API-key routes follow that provider's billing and usage limits.

Gajae-Code is an experimental, beta-stage project. Expect rough edges and verify outputs before relying on it for important work.

Most coding agents fail on three fronts: they bill you twice, they mutate before they understand, and they go silent the moment you step away from the keyboard.

Quick Start

Install — prebuilt binaries for Linux (x64/arm64, glibc only), macOS (arm64/x64), and Windows (x64). Bun is not required:

curl -fsSL https://raw.githubusercontent.com/Yeachan-Heo/gajae-code/v0.15.3/scripts/install.sh -o gjc-install.sh
sh gjc-install.sh
gjc

Piping main executes mutable content; use it only if you explicitly want the latest installer:

curl -fsSL https://raw.githubusercontent.com/Yeachan-Heo/gajae-code/main/scripts/install.sh | sh

Windows (PowerShell), tagged:

Invoke-WebRequest -UseBasicParsing https://raw.githubusercontent.com/Yeachan-Heo/gajae-code/v0.15.3/scripts/install.ps1 -OutFile gjc-install.ps1
powershell -File gjc-install.ps1

First use — inside an interactive gjc session:

/login                       pick a provider / coding plan
/skill:deep-interview        clarify ambiguous requirements
/skill:ralplan               build and critique the plan

After approving the plan, ask the active GJC session to run this through its Bash tool, which supplies GJC_SESSION_ID, to create durable goals. With goal mode enabled, the session can activate inline goal mode and begin work in the same turn, before another user action; run this only when ready for execution. The command does not start the durable scheduler. A standalone shell without that session context fails:

gjc ultragoal create-goals --brief-file <approved-plan>

Run modes:

gjc                                # run in the current checkout
gjc --tmux                         # tmux-backed leader session
gjc --tmux --worktree my-task      # isolated worktree for risky work
gjc @screenshot.png "What should I change?"   # image input

Nightly channel: sh gjc-install.sh --channel nightly (use the tagged installer downloaded above). Full install matrix, Windows setup, update channels, and shell completion: docs/install.md. Bun is only needed to build from source.

Bring your coding plan

Log in once and run GJC on the subscription you already pay for. Run /login inside a session and pick your plan:

/login supports both OAuth and key-based onboarding; for example, Z.AI and MiniMax Coding Plan require a plan API key, not an OAuth browser login. More login options — Google Gemini CLI, GitLab Duo, Perplexity Pro/Max, Fire Pass, Xiaomi Token Plan — are covered in docs/models.md. Availability, entitlement, and usage limits depend on the provider.

New: coding-plan presets

Key-based coding plans onboard with one command — the preset writes the API type, base URL, env-var reference, compatibility flags, and model configuration. The two discovery-backed presets below resolve a live model catalog at runtime; setup itself does not validate credentials:

gjc setup provider --preset commandcode-goat   # Command Code GOAT plan (CMD_API_KEY)
gjc setup provider --preset cline-pass         # ClinePass (CLINE_API_KEY)

• Command Code GOAT — discovers the provider's live /provider/v1/models catalog. Claude model IDs use the Anthropic Messages transport; other models use OpenAI-compatible Chat Completions. /login verifies the key with a small inference probe before persisting it; this can consume provider quota and is separate from preset setup. Aliases: commandcode, goat.
• ClinePass — no hardcoded models; GJC fetches Cline's live catalog the same way Cline generates its own. Aliases: clinepass, cline.
• Also available as presets: minimax, minimax-cn, glm, alibaba-token-plan — or /provider add --preset <name> inside the TUI.

Answer from your phone

After you configure notifications (off by default), the agent can ping you on Telegram when it needs a decision — and you answer from anywhere:

• Coordinator/lifecycle session forum topics with live/finalized output, context updates, image attachments, inline buttons, free-text replies, and typing indicators.
• Configure once from /settings → Notifications in a running session, or headless via gjc notify setup|status|health|test|recovery. Tokens are masked on entry and never displayed again.
• gjc daemon lets sessions sharing a local agent directory and bot/chat configuration reuse a long-poll owner. Other GJC profiles, hosts, or applications polling the same token can still cause Telegram 409 conflicts.
• Discord and Slack delivery ship alongside; the generic action_needed/reply protocol lets any bot or mobile app route answers back without terminal scraping.

Telegram onboarding · Discord · Slack

Plan before mutation

A deliberately small workflow surface — four skills, four role agents, nothing else:

deep-interview -> ralplan -> ultragoal
               └─ optional autoresearch mission when research must ground the plan

These workflows are explicitly invoked, not an automatic prerequisite for every request. Clear, low-risk changes can use direct execution. Planning workflows do not authorize implementation by themselves: execution requires explicit approval or an admitted ralplan runtime handoff.

Also available: the experimental computer tool for Apple Silicon macOS desktop control (not a fifth bundled workflow skill). It is callable by default on supported hosts, subject to supervisor and macOS permissions; set computer.enabled=false to disable it. See Python REPL and docs/tools/computer.md.

Custom skills

GJC uses the Claude Code / Codex SKILL.md file convention, but loads runtime skills directly only from canonical GJC locations — no configuration required:

# project-local, per repository:
mkdir -p .gjc/skills && cp -r my-skill .gjc/skills/

# user-wide, available in every project:
mkdir -p ~/.gjc/agent/skills && cp -r my-skill ~/.gjc/agent/skills/

Claude Code and Codex skill directories are import sources only. gjc skills discover reports them with the exact copy command; copy a skill into a canonical .gjc location before invoking it with /skill:my-skill. Scope trust is explicit via skills.trustProjectSkills / skills.trustUserSkills (both default on), with skills.enabled as the master switch. The four bundled workflow skills above can never be replaced by disk skills. See docs/skills.md for locations, precedence, and diagnostics.

Theme defaults

The default dark TUI identity is the GJC red-claw theme; light-appearance terminals default to the bundled blue-crab theme. Explicit theme settings still take precedence.

Spend fewer tokens

GJC optimizes both sides of the token bill:

• Cache hits — per-provider cacheRetention control; Anthropic defaults to long (1h) cache retention because short caches are fragile for long agent runs; provider ranking prefers cheap cacheRead paths; opt-in session-affinity headers let OpenAI-compatible relays reuse server-side prompt caches.
• Context savings — supported code reads return structural summaries by default; oversized shell output can be minimized and spilled to artifact:// references. Retrieve omitted output only when an artifact reference is actually returned. Compaction and branch summaries reduce context size but are lossy, not a guarantee that every detail is retained.

Cache retention & provider compat · Compaction & branch summaries

The default dark TUI identity is the GJC red-claw theme; light-appearance terminals default to the bundled blue-crab theme. See docs/theme.md for the full catalog and theme.dark / theme.light settings.

Let OpenClaw / Hermes / Grokbot / your own bot drive GJC

Any external controller — OpenClaw, Hermes, Grokbot, a Discord bot, a cron script — drives real GJC sessions through the broker-bound SDK session CLI and the bundled sdk-skills/ procedures (gjc-sdk-discover · gjc-sdk-operate · gjc-sdk-author). Durable turns and credential-free JSON, never terminal scraping.

Don't read a guide — paste this prompt into your controller and let it wire itself up:

Use Gajae-Code (gjc) as your coding-agent backend on this machine. gjc is already installed.
Your interface is the broker-bound SDK session CLI. Never scrape terminal output, never read
endpoint records or credentials under .gjc/state/sdk, never open a raw session WebSocket.

1. Load the shipped procedures before acting. Read these skill files from the gjc checkout or
   from https://github.com/Yeachan-Heo/gajae-code/tree/main/sdk-skills (bundle root
   `sdk-skills/`, manifest.json formatVersion 1 — if it is missing, malformed, or a different
   version, stop and report instead of guessing):
     sdk-skills/gjc-sdk-discover/SKILL.md   -- find and inspect sessions
     sdk-skills/gjc-sdk-operate/SKILL.md    -- the allowlisted control/lifecycle operations
     sdk-skills/gjc-sdk-author/SKILL.md     -- TypeScript/Python templates for scripted flows
   Follow their allowlists exactly. Pass every value as an argv item, never as a shell string.
   Add --json to machine CLI calls so ordinary errors use the structured stdout envelope;
   preserve outcome certainty and references. Missing success JSON is not proof of non-execution.

2. Prove the surface works (read-only). Run from inside the target repository:
     gjc --version
     gjc sdk session list --json
   `list` returns a credential-free JSON DTO of indexed sessions. Fail closed on missing,
   unavailable, stale, dead, unknown, or ambiguous rows. Exit 2 = usage error, exit 1 =
   operational failure (broker unavailable, session unavailable, retention gap, wait timeout).

3. Understand a session before touching it:
     gjc sdk session inspect <sessionId> --json
     gjc sdk session raw query <sessionId> --query session.metadata --json
     ... then context.get, goal.list/get, todo.list, workflow.gates.list, session.stats
   These reads are not an atomic snapshot: label every reported field confirmed / inferred /
   stale / unavailable / unknown. Never invent a missing value.

4. Start work in an isolated session:
     gjc sdk session raw global --op session.create \
       --idempotency-key <fresh-uuid> --json-input '{"cwd":"/abs/path/to/repo"}' --json
   Lifecycle ops allowed: session.create, session.fork, session.resume, session.close.
   session.delete is NOT allowed. session.get_endpoint is refused unconditionally.

5. Drive a turn and reconcile it:
     gjc sdk session send <sessionId> --text "<task>" --op-ref <fresh-ulid> --json
     gjc sdk session status <sessionId> <opRef> --json # lossless turn.result lookup
     gjc sdk session tail <sessionId> --until-idle --json # replay + live follow
   Use `send --wait --timeout-ms <ms>` for a bounded wait; a wait window that elapses reports
   wait_timeout and never cancels the running turn. One fresh op-ref per logical prompt --
   `unknown` means uncertainty, never proof of non-execution, so reconcile with `status`
   instead of replaying a prompt.

6. Answer durable workflow-gate questions:
     gjc sdk session raw control <sessionId> --op workflow.gate_answer --json-input '{...}' --json
   For gate answers use the durable workflow gate ID plus expectedSessionId; a transient
   action_needed.id is never durable authority. The current host does not implement the
   protocol's general ask.answer operation. Other allowed per-session controls:
   turn.prompt, turn.steer, turn.follow_up, todo.replace, session.switch, session.rename.

7. Show the human the exact operation and target before any mutating call, and treat the
   approval as single-use: if the operation, input, or target changes, ask again.

Long prompts are safe to leave running: the SDK prompt deadline is a progress-aware inactivity lease (sdk.promptDeadlineMs, 60 min default) bounded by sdk.promptMaxRuntimeMs (6 h default), renewed only by attributable tool execution for the accepted turn — not by heartbeats or streaming text.

Need event-driven fan-out across many worktrees instead of one session at a time? The native Coordinator MCP bridge (gjc mcp-serve coordinator) exposes the delegation tools for that shape. gjc setup hermes renders a configuration preview; installation requires --install with --target or --profile-dir (see the guide).

• External controller / bot integration guide — provider-independent smokes; docs/aside-integration.md covers the opt-in search/context sidecar and the /aside composer command
• SDK session CLI · SDK & wire protocol · SDK app guide · External-control readiness

Run GJC inside Paseo, Orca, or T3 Code

Prefer a desktop/mobile agent shell over a bare terminal? GJC plugs into the three popular ones — at three honestly different levels of support.

Host	Support	What you get	Setup
Paseo repo	★★★★★ first-class	Native ACP provider installed by GJC itself. Model catalog, Default/Plan modes, thinking levels, real permission prompts, cancel that can terminate owned subagents, mobile control.	gjc setup paseo then paseo daemon restart
Orca repo	★★★★☆ works, one field	GJC runs as a custom CLI agent, one worktree per session, with Orca's diff review, terminal splits, SSH worktrees, and mobile companion. No usage tracking or account hot-swap yet.	Settings → Agents add command gjc
T3 Code repo	★★★☆☆ experimental	T3 Code ships harnesses for Codex, Claude, Cursor, Grok and OpenCode only — there is no GJC harness upstream yet. Run GJC beside it; the native provider is proposed upstream.	not one-command yet — see the guide

Paseo, in one paste:

gjc setup paseo            # writes the ACP provider entry, backs up, never restarts your daemon
paseo daemon restart
paseo provider ls          # gjc must read `available`
paseo run --provider gjc --cwd /path/to/repo "your prompt"

gjc setup paseo --check    # pass / stale / drift, with a machine-readable --json
gjc setup paseo --remove   # rolls back only the keys GJC itself created

Orca, in one field: install GJC (see docs/install.md), then add a custom agent with command gjc and no arguments. Orca pre-fills a permission-bypass flag for agents that expose one — GJC has none by design, so leave the arguments empty and keep GJC's own approval gates.

Full integration guide → docs/terminal-app-integrations.md — per-host setup, verification, cancel semantics, troubleshooting tables, and what each host cannot reach yet.

Documentation

Start at gajae-code.com or docs/:

• Install & updates · Environment variables · Keybindings · Themes · UI language
• Models & providers · Custom providers & multi-account routing · Multi-vendor profiles · Auth broker
• Customization authority, import, and trust · Skills · Hooks · Standalone MCP · Plugin bundles
• Terminal app integrations: Paseo · Orca · T3 Code
• Telegram · Bot integration · SDK · SDK session CLI
• Sessions · Compaction · Memory · Secrets
• Codebase overview · Contributing / dev setup
• macOS Option/Alt key setup (iTerm2) · GEO visibility benchmark

The default dark TUI identity is the GJC red-claw theme; light-appearance terminals default to the bundled blue-crab theme. See Themes to swap or build your own.

SDK extensions

Local customization: /extensions

In an interactive session, /extensions is the primary customization setup surface — it configures skills, hooks, and MCPs across the project (<project>/.gjc/) and user-global (~/.gjc/agent/) scopes, with status/provenance diagnostics, enable/disable/remove, and a guided Import-from-Claude-Code/Codex flow (normalized preview, explicit confirmation, skip/rename/overwrite collision policy, atomic writes with rollback). Non-interactive setups use gjc mcp for MCP servers and gjc migrate for Claude Code/Codex imports.

Skill migration and bundled skill inspection

When moving a workflow into GJC, inspect the bundled defaults before installing or overwriting anything:

gjc skills list
gjc skills read ralplan
gjc setup defaults --check

gjc setup defaults installs the four bundled GJC workflow skills into your user .gjc directory and preserves existing local files by default. If --check reports missing or different files, compare the embedded copy with gjc skills read <name> first; use gjc setup defaults --force only when you intentionally want to replace local default workflow skill files.

Works beside your existing agent or bot

For evaluating Aside as an opt-in search/context retrieval sidecar, see docs/aside-integration.md. For generic third-party bot setup and provider-independent smokes, see docs/bot-integration.md. For external-control readiness, see docs/external-control-readiness.md. For the wire protocol and machine interfaces, see docs/sdk.md.

SDK Extensions

• gjc-remote — a real-world SDK extension for controlling allowlisted GJC sessions on remote hosts from Discord.
• oh-my-gajae-code — a community plugin marketplace for installing additional workflow skills and slash commands.
• gjc-agy-skill — a third-party community GJC skill integrating Antigravity CLI for vision/OCR, image generation, and Gemini print-mode one-shot workflows.
• GJC multivendor setup guide — role-based provider profiles and installable model bundles for multivendor GJC setups.

Configuration

Provider retry budgets can be configured in ~/.gjc/agent/config.yml (the default global agent directory). The defaults for these keys are:

retry:
  requestMaxRetries: 5
  streamMaxRetries: 5
  maxRetries: 3
  maxDelayMs: 300000

requestMaxRetries applies before a stream is established. streamMaxRetries applies only to replay-safe transient stream failures. Invalid auth, unsupported models/providers, malformed requests, context overflow, user aborts, and permanent quota failures remain fail-fast.

Launch-time updates

Interactive startup checks GitHub releases for a newer GJC version in the background by default. This check is notify-only and non-mutating: GJC never installs or replaces itself during launch. On a supported platform, gjc update downloads and atomically replaces the matching GitHub release binary (package-manager shims are not overwritten). Source checkouts and dev:link executables must be updated through that checkout; gjc update refuses to self-overwrite them. Unsupported platforms should rerun the documented installer.

Run gjc config set startup.checkUpdate false to disable the launch-time check. Network failures are ignored so they do not block startup.

Good to read together

• GJC multivendor setup guide — a community guide for role-based provider/profile selection across Anthropic, OpenAI/Codex, Google/Gemini, xAI/Grok, and opencode-go. Treat its presets as user-level configuration guidance rather than bundled defaults; verify model availability and provider auth in your own environment before adopting them.

TUI identity

The default dark TUI identity is the GJC red-claw theme, while light-appearance terminals default to the bundled blue-crab theme. Three additional bundled migration themes — claude-code, codex, and opencode — mirror the look of those tools for easy eye-migration and are selectable from Settings or /theme. Explicit user theme settings still win.

Bundled theme grid

Pick from Settings (Appearance -> Dark theme / Light theme) or /theme.

Troubleshooting

When a tool, skill, hook, extension, slash command, MCP server, or plugin bundle does not appear as expected, start here:

gjc customize doctor         # human-readable provenance and remediation
gjc customize doctor --json # stable JSON for CI/setup tooling

gjc customize doctor is the single read-only troubleshooting surface. It reports every discovered customization, its source convention and scope (gjc, Claude project, Codex project, plugin, explicit config), effective precedence and shadowing, loaded/enabled/disabled/quarantined/rejected/stored-only status, bounded reason codes, remediation commands, trust requirements, and whether a restart/new session is required. Credentials, endpoint tokens, auth headers, and unsafe raw config dumps are never printed.

Development

bun install
bun run build:native
bun run dev:link       # global `gjc` runs this checkout's source
bun run dev:doctor     # verify the link

See CONTRIBUTING.md and docs/codebase-overview.md for the package map and gates.

Contributors & lineage

Thanks to Yeachan-Heo, IYENTeam, HaD0Yun, probepark, and snowykr. Repository maintainers and their GitHub access are listed in MAINTAINERS.md. GJC builds on lessons from a small family of agent harnesses; historical attribution lives in NOTICE.md.

License

MIT. See LICENSE.

"Encode intention. Decode software."

The plan comes first. The mutation earns its place.