1UP · KRONN KRONN FR · EN · ES
▶ INSERT 0 TOKENS TO START ◀

MANY AGENTS.
ZERO CHAOS.

many agents · zero chaos

You juggle multiple AI CLIs (Claude Code, Codex, Gemini, Ollama) across multiple projects. Kronn brings shared memory, orchestration, and continuity. Local-first. 0 tokens on mechanical steps.

v0.8.6 · early access · API and step types may still change before v1.0
LEVEL 1 · THE BRIDGE

One shared brain.
All your CLIs plugged into it.

Four AI CLIs that ignore each other. Three diverging MCP configs. Fourteen copy-pasted prompts. Accepted state of the art. Kronn proposes a shared brain, local, plugged into everything.

· SYSTEM MAP ·
Claude Code
Codex
Gemini
Ollama
KRONN
shared brain · local-first
Living memory

Locally persisted discussions. Same context picked up by any CLI, OR N CLIs live in the same conversation.

Prompts

Versioned Quick Prompts, measured, ✨ AI Improver.

Routing

Mechanical steps without an agent. AI cost = 0 tokens.

LEVEL 2 · REALITY CHECK

Before.
After.

Here is what you do today, and what Kronn does for you.

Without Kronn
  • Your prompts copy-pasted across three Notion files and a dusty Markdown.
  • Switching Claude Code → Codex causes a full context reset.
  • You play messenger between agents, manually copy-pasting every message across 3 windows.
  • Three MCP configs maintained in parallel, divergence guaranteed.
  • A Slack notification goes through an agent, so consumes tokens on every send.
  • Sending an email or pushing a log also goes through an agent (MCP mail, plugin), tokens on every call.
  • You re-explain your architecture on every new discussion.
With Kronn
  • Quick Prompts versioned with {{vars}}, metrics per iteration.
  • Same discussion, handed to the next CLI via MCP, zero copy-paste.
  • 3 CLIs (Claude, Codex, Gemini…) talk inside a single Kronn discussion via disc_join. Zero human messenger.
  • One interface, auto-propagation to all installed CLIs.
  • Step Notify runs agentless: zero tokens.
  • Step ApiCall (Resend, Mailjet, log webhook): zero tokens, the API directly, no LLM.
  • docs/AGENTS.md generated by audit, auto-read by every agent.
▸ Fig. I · decomposition
Mega-prompt vs Kronn workflow · same outcome, 5× cheaper Two-row diagram. Top row: the usual anti-pattern, a mega-prompt asking an agent to do twelve things at once, around twelve thousand tokens, fragile, debug impossible. Bottom row: the Kronn way, five sequential steps (ApiCall, Agent, Exec, Agent, Notify), three deterministic at zero tokens and two bounded Agent steps. Same outcome, much cheaper. The usual approach · one mega-prompt "Do the 12 things" fetch + analyse + write + test + review + commit + notify… 💸 way (too) many tokens · unpredictable AI black box 12 concerns in 1 call ❌ debug = impossible Output fingers crossed? The Kronn way · decompose, code where you can, AI where you must ApiCall fetch ticket 0 token Agent draft impl N tokens Exec cargo test 0 token Agent review diff N tokens Notify webhook 0 token Better outcome fewer tokens 5 steps · 3 deterministic (0 tokens) · AI only fires on the 2 reasoning steps → fewer hallucinations · faster · predictable cost · debuggable
KRONN STATEMENT

You still write prompts by feel.
You copy-paste your context from one agent to another.
You pay the AI to send a Slack webhook.
You maintain three MCP configs that diverge the moment you blink.

Stop.

Kronn ends this era.

LEVEL 3 · LOADOUT

Everything
that ships in the box.

No roadmap, no soon™. Everything is shipped, in the code, verifiable. Exhaustive list of what Kronn gives you today.

01Memory & context
  • Discussions persisted locally (SQLite)
  • Cross-CLI resume: Claude exports, Codex picks up (0.8.4)
  • Live multi-CLI: N agents inside one discussion (0.8.6)
  • docs/AGENTS.md generated by audit
  • User-context injection ~/.kronn/user-context/
  • Linked repos : contexte multi-repo
  • Canonical state .kronn.json per project
  • Anti-secret filter on the audit
02Quick Prompts
  • Variables {{vars}} + conditional sections
  • Versioning + metrics for tokens · duration · cost
  • Visual diff between versions
  • ✨ AI Improver: suggests and deploys in 1 click
  • Compare across N agents in parallel
  • Chains: DnD reorder + {{previous_qp.output}}
  • Binding profile · skill · directive
03Workflows
  • 7 step types: Agent · ApiCall · Batch · Notify · Exec · Gate · BatchQP
  • 0 tokens on mechanical steps (by definition)
  • AI Workflow Architect : dry-run + 1 clic
  • Batch workflows (fan-out · chaining · git worktree · integrated diff)
  • Loops + state scratchpad + guards
  • Rollback on failure
  • CRON-schedulable : workflows full-auto
  • AutoPilot loop (audit → tickets → workflow)
  • Feasibility-Gated implementation
  • Export / Import per-item
04Agents · plugins · local
  • 7+ supported CLIs (Claude Code · Codex · Gemini · Ollama · Vibe…)
  • Eco mode via RTK: one-click token-killer proxy on supported agents
  • Ollama 100% local + TTS Piper / STT Whisper (FR · EN · ES)
  • MCP plugins: Atlassian · Linear · Notion · …
  • API plugins: Resend · Mailjet · webhooks
  • Custom API plugin · declare your own endpoints
  • Multi-user P2P (chat + shared discussions)
  • Tauri desktop app (Mac · Windows · Linux)
  • AGPL-3.0 · no telemetry · no cookies
▸ KRONN PATTERN · PLUGINS
MCP Plugin
the agent explores
Atlassian · Linear · Notion · Confluence · Stripe …
AI Helper
switch MCP → API
dedicated agent, guided dialogue to generate the API plugin
API Plugin
deterministic · 0 tokens
Resend · Mailjet · webhooks · Custom API …

You start with an MCP plugin, the agent learns how the vendor responds, tests, iterates. Once the sequence stabilizes, a dedicated AI Helper walks you through switching to an API plugin: the AI leaves the loop, execution becomes deterministic and token-free.

LEVEL 4 · COMPOUND EFFECTS

Five structural
mechanisms.

No promised numbers, just the building blocks that let regular usage lower cost and raise quality over time. Each one is in the code, reproducible, verifiable.

01
Audit

Full project scan: stack, conventions, tech debt. Generates tier-loaded docs/AGENTS.md. AutoPilot (opt-in) turns debts into Jira tickets via MCP and pre-configures the processing workflow in 1 click.

02
Memory

Locally persisted discussions + 10 MCP tools disc_* for bidirectional cross-CLI continuation. The next CLI reads AND enriches the thread, not just consults.

03
User-context

Your global preferences (commit style, conventions, anti-patterns) live in ~/.kronn/user-context/. All your projects benefit automatically. Each preference added → effect everywhere.

04
Quick Prompts

Reusable prompts with {{vars}}. Versioned and measured (tokens · duration · cost). You iterate, compare, pick the winner.

05
Workflows

Step-based orchestration: Agent step + 5 mechanical steps (Exec, Gate, Notify, ApiCall) at 0 tokens by definition. CRON-schedulable. Feasibility-Gated pattern for big tickets: YAML triage → human Gate → constrained code.

LEVEL 5 · NOT INCLUDED

Ce que Kronn n'est pas.

Stating exclusions wins over skeptics. Here is what Kronn was not built to be.

×Not an IDE

Keep Cursor, VSCode, Vim, or Aider for CLI pair-prog. Kronn orchestrates above, doesn't replace your editor, though you can do quite a bit from Kronn: git worktree, diff access, direct exec, human gates ;)

×Not an LLM

You bring Anthropic, OpenAI, Google, or Ollama 100% local. Kronn locks you into none.

×Not a SaaS

Local-first desktop app. Your data, prompts and discussions never leave your machine.

×Not a framework

You drive via UI, you don't code. Multi-CLI orchestrated, not LangGraph in Python.

×Not a general workflow tool

For non-AI integrations (Slack, ETL, CRM, lifecycle email), keep n8n or Zapier. Kronn targets AI agent orchestration specifically.

×Not a CI/CD

For build, deploy, lint on git, keep GitHub Actions. Kronn orchestrates local-first AI agent CLIs, not repo-bound cloud pipelines.

LEVEL 6 · QUICKSTART

5 minutes.
4 stages.

OK, convinced? Here is the local install. At the end you have an audited project, a measured Quick Prompt, and discussions that survive switching between 2 CLIs.

STAGE 01
Install

Two install paths:
Desktop app · download the binary (Mac, Windows, Linux) from releases.
Self-hosted · git clone -b <tag> then kronn start (CLI / headless, ideal for teams).
Local-first in both cases. Agent metrics stay local (QPs, runs, tokens, cost) · never cloud telemetry.

↓ Releases
STAGE 02
Audit

Open Kronn, add your project, launch audit with Briefing (≈ 2 min of guided Q&A). Kronn then scans your stack, detects conventions, identifies tech debt and generates tier-loaded docs/AGENTS.md.

Launch: 2 min. Background scan: 20-30 min on an average project, more on a monorepo. AutoPilot ready afterward to turn debts into tickets.

STAGE 03
Quick Prompt

Automation tab, create a Quick Prompt. Drop a {{ticket}} variable in the template. Run it on an agent: Kronn records tokens, duration, cost.

You can run the same Quick Prompt on N agents in parallel to compare outputs.

STAGE 04
Workflow from a template

Automation → Workflows tab, click New and pick a preset (Ticket-to-PR, Feasibility-Autopilot, Feature Planner…). Kronn pre-fills the steps Brief → Architecture → Code → Tests → Review with their human Gates. All you do is approve at each Gate.

The Quick Prompts from STAGE 03 work as building blocks inside your custom workflows.

LEVEL 7 · CONCEPTS

The technical vocabulary.

Eight terms that show up everywhere in Kronn. Quick glossary to decode the rest: what it is, what it does.

Quick Prompt

Reusable prompt template with variables {{ticket}}, {{context}}. Versioned. Metrics recorded per version (tokens, duration, cost). ✨ Improver button to iterate with AI.

QP Chain

A sequence of Quick Prompts. Each QP can reference {{previous_qp.output}} to chain the previous output. Drag & drop reorderable.

MCP

Model Context Protocol · open standard (by Anthropic) letting AI agents access external tools. Kronn relies on MCP to bridge the CLIs and expose its own tools via kronn-internal.

StepType

The 7 step types in a workflow: 2 AI (Agent, BatchQuickPrompt) and 5 mechanical (ApiCall, BatchApiCall, Notify, Exec, Gate) at 0 tokens.

AutoPilot

Opt-in loop after audit: Kronn detects tech debt, creates Jira tickets via MCP, pre-configures a workflow to handle them. One click, done.

Feasibility-Gated

Implementation pattern for big tickets: Kronn generates a feasibility YAML manifest, then a human Gate (you approve or reject), then the agent codes following the manifest. Zero surprise.

docs/AGENTS.md

The canonical doc source for a Kronn project. Generated by the audit. Tier-loaded (light always-on, sections on-demand). Auto-read by every CLI that opens the project.

Plugin

4 flavors: MCP (vendor exposes a server, e.g. Atlassian), API (API-only vendor packaged, e.g. Resend), Hybrid (both), Custom API (declare your own endpoints).

LEVEL 8 · SECURITY & PRIVACY

Your machine.
Your data.

Local-first by default. No telemetry, no cookies, no Kronn server. Here is what is protected, and what we don't have yet: honesty before the certification seal.

Local-first
  • Local SQLite DB in ~/.kronn/
  • Discussions, QPs, audits, configs: all on your machine
  • No cloud sync, no telemetry, no cookies
  • Your data only leaves if YOU launch a workflow that calls a cloud LLM
Secrets & keys
  • API keys stored in the OS keychain (Mac/Windows/Linux)
  • Anti-secret filter on the audit: .env and vendor prefixes filtered before docs/AGENTS.md
  • No Kronn env var pushed to third-party CLIs
Exec & Gate
  • Step Exec gated by workflow exec_allowlist (literal argv, no sh -c)
  • Step Gate human-in-the-loop: the agent waits for your explicit approval
  • Audit log: each run persisted with timestamp, tokens, steps executed
What we don't have
  • Not yet v1.0 stable. Current version v0.8.6, early access. The API, step types and schemas may evolve between versions.
  • No SOC2 or GDPR certification at this time
  • Single-user · no RBAC or team management
  • If you use Anthropic / OpenAI / Google, your prompts go to them: their responsibility, not ours
  • AGPL-3.0 → auditable code, but no external security audit at this stage