Planner
plannersplits the work into a graph of slices with dependencies
graph · slice · deps
v2.65.0 · Apache-2.0
Zelari-Code · Open source
Zelari-Code, the agent that doesn't trust itself
An open-source coding orchestrator: you choose the model, the proof is mandatory. Kraken splits the work and delegates it to specialised tentacles, and a turn cannot close as successful until verification passes.
Real screenshot · v2.65.0
zelari-code@studio: ~
$ zelari-code
[ok] Kraken online · 4 tentacles ready
[ok] strict-done ON · worktree ON
[kraken] lead · splits the work into slices
01The problem
The problem with chat coding
A single chat thread is not a development workflow, and without verification you do not know whether the work is really done.
Decaying context
In long sessions context fades: the model forgets decisions made two hours earlier.
"Done" on trust
The agent says it works and nobody runs the tests or the build before closing.
Colliding writes
Several steps edit the same files in the same tree, with no isolation and no rollback.
Gut-feeling permissions
Either you approve every command by hand or you let everything through: there is no policy in between.
Little audit
Hard to trace what happened, why a command was denied or who wrote what.
The Zelari-Code promise
Kraken splits and delegates, every writer works in an isolated worktree, a blind reviewer judges from real output and the strict-done gate refuses to close with open verification. Everything lands on the session spine, replayable at any time.
02Features
What makes it different
Not an autocomplete, not a chat agent: an orchestrator that proves the work instead of claiming it.
01
Kraken and tentacles
A coordinating lead and four roles: Planner, Explorer, Builder and Checker. Each role can use a different model, even from another provider. The six-role council is still there when the work earns a second opinion.
02
Isolated worktrees
Every writing tentacle works in a private git worktree seeded from your tree, uncommitted changes included. When it is done it squash-merges back; on a conflict it rolls back without touching your work.
03
Strict-done
The Checker is blind: it only sees the real output of tests, typecheck and build. A written PASS with no executed commands does not count; with open verification the turn does not close as successful (exit 4).
04
Policy-based permissions
Allow/ask/deny rules evaluated before every tool, strict, standard and yolo presets, and a deterministic command classifier (safe, review, destructive, blocked). Every denial is recorded.
05
Memory and sessions
A replayable session spine, the .zelari/ workspace, deterministic compaction and resumable sessions. Outcome memory only records work that passed verification.
06
Provider-agnostic, cost-aware
Grok, ChatGPT, Claude, Muse, GLM, MiniMax, DeepSeek, Xiaomi MiMo or any OpenAI-compatible endpoint. A prompt-cache-friendly layout: since 2.65, about 12,500 fewer tokens per lead request.
- Grok
- ChatGPT
- Claude
- Muse
- GLM
- MiniMax
- DeepSeek
- Xiaomi MiMo
- OpenAI-compatible
03How it works
From request to merge, with proof
The default flow in kraken mode: from request to merge, with verification as the condition for closing.
Lead
Kraken reads the request, splits it into slices and decides what to do itself and what to delegate.
Explore
Read-only tentacles map the relevant code and report what they found.
Build
Builders write in isolated git worktrees, starting from your current working tree.
Verify
A blind Checker runs tests, typecheck and build: the verdict comes only from real output.
Merge & gate
Squash-merge into your tree. With open verification debt the turn does not close as successful.
04Architecture
Kraken and its tentacles
A coordinating lead and four specialised roles, each with its own model and thinking effort.
KRAKEN · lead
plan · delegate · gate
Planner
planner
Explorer
explore
Builder
general
Checker
verify
verifysquash-mergestrict-done gate
Explorer
explorereads and maps the code, never writes
read-only · quick/medium/deep
Builder
generalwrites in a private worktree, then squash-merges
.zelari/worktrees/ · merge
Checker
verifyjudges only from real command output
npm test · tsc → VERDICT
On demand
Council on demand
When the work earns a second opinion: six roles in sequence for planning, ideas, knowledge mapping, critique and synthesis. From the CLI with --mode council; in Desktop 2.65 the modes are Kraken and Zelari.
Caronte01
Orchestrator · Director
frames the problem and delegates
mandate → specialists
Nettuno02
Planner
buildable plan, phases and tasks
createPlan · phases × tasks
Gerione03
Ideator
creative divergence, then top pick
journey · IA · tokens
Plutone04
Knowledge Architect
navigable concept map
knowledge-map
Minosse05
Quality Critic
exposes gaps and risks, no product code
risks · scores
Lucifero06
Synthesizer · Implementer
synthesis; implements only when enabled
synthesis · write/edit
05Session
One run, start to finish
Illustrative flow in kraken mode, headless with the strict-done gate. Simplified output; the tentacles, worktree and gate are the real ones.
zelari@studio: zelari-code --headless --task "Add pagination to GET /tasks" --strict-doneIllustrative demo
$ zelari-code --headless --task "Add pagination to GET /tasks" --strict-done
[kraken] lead · plan in 3 slices ✓
[explore] /tasks route + existing tests · read-only ✓
[general] worktree .zelari/worktrees/t2 · 4 files ✓
[verify] npm test · tsc --noEmit → PASS ✓
[merge] squash-merge into the lead tree ✓
[gate] strict-done · evidence on the spine ✓
> exit 0 · verified, not claimed
06Interfaces
One engine, six ways to use it
CLI
zelari-code: scriptable, with /update, --doctor --fix and slash commands.
TUI
Keyboard-first interactive console, configurable status line, Ctrl+O for full tool output.
Desktop
Tauri workbench with tabs, Kraken activity, MCP and SSH. Runs the CLI as a sidecar.
Headless / CI
--headless --task with JSON output, exit codes 0–4 and trajectory export in ATIF format.
Editors via ACP
zelari-code acp speaks the Agent Client Protocol: Zed, Neovim and an experimental VS Code extension.
Android companion
Your phone as a thin client over Tailscale or LAN: approvals, questions and run steering.
07Desktop · v2.65.0
A workbench, not a chat window
A workbench, not a chat window: per-project tabs, live tentacle activity, models and permissions one click away. Same engine as the CLI.
Tabs and projects
Conversations open as tabs and the sidebar groups them by project. Parallel chats on different projects stay isolated.
Tentacles pill
Pick the delegation level (Automatic, Prefer, Maximum, Lead only) and a model and thinking effort per role, right from the composer.
Kraken activity
Every tentacle with its status, duration, model and actual provider. One whose tool calls mostly failed is flagged as "degraded".
Runs dashboard
Every run from every conversation in one drawer, with project, starting prompt and live status.
MCP and SSH extensions
Catalog or custom MCP servers, SSH connections with a command allowlist, and readable system notices that say what happened and what to do.
Android companion
Pair your phone with a QR code: approve or deny tools, answer the agent's questions and steer the run while you are away from your desk.
08Download
Download Zelari Desktop
Optional Tauri shell for long sessions, tentacle activity and parallel chats. Same engine as the CLI.
Installer v2.65.0. The installer does not install or update the CLI: use npm i -g zelari-code (or Settings → Update CLI). On first launch a setup guide checks Node and the CLI.
Requirements and setup
- Node.js ≥ 20.17 and npm ≥ 10
- CLI: npm i -g zelari-code, or npx zelari-code@latest with no install
- An LLM provider: API key or OAuth login (/login grok, chatgpt, anthropic, muse)
- Git recommended; on Windows the bash tool needs Git Bash
- Windows x64 · macOS arm64 · Linux x64
v2.65.0 · Desktop shell · Apache-2.0
09Use cases
Where it makes the difference
Where Zelari-Code makes a difference compared with plain chat coding.
End-to-end features
From brief to merge: Kraken plans, Builders implement in parallel and the Checker verifies before closing.
Review and refactoring
Multi-file refactors with no risk to your tree: isolated worktrees and rollback when a merge fails.
Architecture decisions
The council weighs plan, alternatives and risks before any code is written.
Unattended runs
Headless with --permissions yolo and strict-done: the exit code tells you whether the work is really verified.
Team & handoff
Replayable sessions with reports: intent, decisions, denials and outcomes in a single thread.
10Comparison
Chat coding versus orchestration
What changes compared with a generic chat assistant.
| Capability | Chat coding | Zelari-Code |
|---|---|---|
| Architecture | Chat codingsingle-agent chat | Zelari-CodeKraken + tentacles, council on demand |
| Definition of done | Chat codingon trust | Zelari-Codeblind Checker + strict-done |
| Write isolation | Chat codingsame tree | Zelari-Codegit worktree per tentacle |
| Permissions | Chat codingall or nothing | Zelari-Codeallow/ask/deny policy |
| Project memory | Chat codingephemeral | Zelari-Codepersistent (.zelari/) |
| Audit | Chat codingscattered logs | Zelari-Codesession spine + replay |
| Models | Chat codingoften lock-in | Zelari-Codeone per role, any provider |
| Interfaces | Chat codingvaries | Zelari-CodeCLI · TUI · Desktop · ACP · mobile |
11Security
Control and safety
Kraken works for you, not instead of you without supervision.
Verification before done
No turn closes as successful with open verification: the strict-done gate is on by default in kraken build mode.
Permission policy
Allow/ask/deny rules before every tool, with deny always winning. Every denial lands on the spine with the rule that caused it.
Isolated writes
A git worktree for every writer, a snapshot of your uncommitted work and automatic rollback when a merge fails.
Classified commands
A deterministic classifier, no LLM involved, tags every command as safe, review, destructive or blocked.
Resumable, replayable sessions
Resume, replay and end-of-session reports rebuilt from the spine: you can always trace the why.
Open source Apache-2.0
Source on GitHub, packages on npm (zelari-code and @zelari/core). Built by Anathema Studio.
12FAQ
Frequently asked questions
Quick answers before the full documentation.
Commercial support or custom integrations?
ContactYes, Apache-2.0. Source on GitHub (N-THEM-Studio/zelari-code), the CLI on npm as zelari-code and the reusable runtime as @zelari/core. Current version: v2.65.0.
You need Node.js ≥ 20.17. To try it with no install: npx zelari-code@latest. For daily use: npm install -g zelari-code, which also enables /update. On Windows, run zelari-code --doctor once.
The work is split and delegated to tentacles with different roles and permissions; writers work in an isolated worktree and a blind reviewer verifies with real commands. "Done" means verified: with open verification the turn does not close as successful.
Mode chooses how work is dispatched: kraken (default), council or zelari (missions that keep going until the result is proven). Phase (plan or build) decides whether the agent may change the project. They are independent axes.
Yes. It is a six-role pipeline (Caronte, Nettuno, Gerione, Plutone, Minosse, Lucifero) for work that earns a second opinion. From the CLI use --mode council; in Desktop 2.65 the available modes are Kraken and Zelari.
xAI Grok, ChatGPT, Anthropic Claude, Muse, GLM/Z.AI, MiniMax, DeepSeek, Xiaomi MiMo and any OpenAI-compatible endpoint. OAuth login via /login for Grok, ChatGPT, Anthropic and Muse. You can give each tentacle a different model.
No. CLI, TUI and headless are enough. Desktop is optional: it runs zelari-code --headless as a sidecar and does not replace npm i -g zelari-code.
Yes, through the Agent Client Protocol: zelari-code acp works with Zed, Neovim and other ACP clients. There is also a VS Code extension, still experimental.
13Documentation
All the detail, in the docs
Kraken and tentacles, mode and phase, permissions, interfaces and install. The CLI reference lives in the repo (docs/GUIDE.md).
- TypeScript
- Node.js ≥ 20.17
- Ink + React (TUI)
- Tauri 2 + React (Desktop)
- Kraken super-agent
- Git worktrees
- Strict-done gate
- Session spine
- MCP
- ACP
- Provider-agnostic LLM
- Apache-2.0
Next step
Try Zelari-Code from your terminal
Open source Apache-2.0: run npx zelari-code@latest, install from npm or download Desktop 2.65.0. Contact us for commercial support and custom integrations.
npx zelari-code@latest · Apache-2.0
