Narwhal is a native desktop app that connects an AI agent to your Jira, Confluence, Slack, pipelines, and oncall — and runs entirely on your machine.
Real workflows, not demos. Every feature below works today through connected MCP tools and built-in capabilities.
Read, search, update, and triage tickets without opening a browser. The agent compresses Jira's verbose API into clean markdown — 90% fewer tokens, instant context.
"Read ORCHA-789" → summary, status, assignee, description (converted from Jira markup), and recent comments. No raw JSON.
Jira MCP"Find my open tickets" or "Search for blocked stories in sprint 42" → compact table with status icons.
Jira MCPWrite in markdown, the agent auto-converts to Jira markup. No more fighting with Jira's editor.
Jira MCP"Show me the ORCHA epic" → tree of all child stories with status, assignee, and progress.
Jira MCPCreate, edit, and search Confluence pages through conversation. The agent handles the storage format conversion — you just write in plain language.
"Create a design doc for the new auth flow" → full Confluence page with headings, tables, and code blocks.
Confluence MCP"Update the architecture section of page 12345" → edits just that section without touching the rest. No accidental overwrites.
Confluence MCPAutomatically escapes CDATA markers, ampersands, and other XML pitfalls. Learns from errors and fixes them on retry.
Self-Learning"Find the onboarding docs" → searches by title, label, or content across all your Confluence spaces.
Confluence MCPSearch messages, read threads, post updates — all from the agent. No more context-switching between Slack and your work.
"Find the discussion about the API migration" → searches messages with Slack's full query syntax.
Slack MCP"What's the latest in #platform-oncall?" → reads recent messages and thread replies.
Slack MCP"Post a summary of today's work to #team-updates" → composes and sends a message.
Slack MCPThe agent runs on your machine with access to your filesystem, shell, and dev tools. It can read, write, and execute — with your permission.
Browse your project, read source code, create new files, edit existing ones. Full filesystem access within your workspace.
Built-ingit, npm, python, curl — anything you'd run in a terminal. Tiered permissions ensure dangerous commands require approval.
Built-inData analysis, CSV processing, calculations — the agent writes and runs Python in a sandboxed subprocess.
Built-inVisual file tree in the sidebar. Click to preview. Track which files the agent created or modified.
Built-inA separate web service that replaces Confluence and Jira for structured product requirements. One database, one source of truth. AI agents are first-class users.
Every requirement is a typed, filterable row — not a cell in a 300KB Confluence XML page. Filter by domain, priority, milestone, product variant.
NarwhalCloudBrowse requirements, view PRDs, track projects, search across everything — all in a browser. Shareable URLs for any requirement or project.
Web App"How many P0 requirements does Hawkfish have?" — the agent queries the database through typed MCP tools. No markup generation, no XML parsing.
MCP ToolsEvery change to every requirement is logged: who changed it, when, what the old value was. Not page-level versioning — field-level.
Built-inFour agent patterns — supervisor, pipeline, swarm, and context brief — all with hard safety boundaries. Subagents get scoped tools and can't recursively spawn.
"/brief" spawns parallel agents for Slack, Jira, Email, Confluence, and Calendar. Each searches independently, then a synthesizer merges findings into one coherent summary.
SwarmA Jira agent can only use Jira tools. A Slack agent can only search Slack. Spawning tools are stripped at runtime — no infinite recursion.
SafetyRead-only, concurrency-safe tools run simultaneously. If one fails, siblings are aborted. Destructive tools always run sequentially with approval.
Built-inBuild your own agent. Narwhal remembers your preferences, learns your patterns, and lets you create specialized personas for different types of work.
The agent remembers your projects, coding style, team context, and preferences across sessions. All stored as editable Markdown files.
Built-inSwitch between Architect, Code Reviewer, Shell Expert, Writer — or create your own with custom system prompts and behaviors.
Built-inReusable instruction sets that activate on keyword triggers. "Review this PR" loads your code review skill. Create skills as Markdown files.
SkillsAdd new capabilities by connecting MCP servers — Figma, GitHub, databases, internal APIs. The agent discovers and uses new tools automatically.
MCP ProtocolObserves tool usage, detects friction, proposes improvements. Learns which commands to auto-approve. All changes require your explicit approval.
EvolutionOpt-in morning briefing: oncall shifts, open tickets, Jira activity, pipeline health. Configure sources and schedule in Settings.
Built-inNo cloud servers. No data leaving your machine. Just a native app that talks to your tools.
Download the DMG, drag to Applications. Narwhal authenticates through Midway — same credentials you already use.
MCP servers for Jira, Confluence, Slack, and more are pre-configured. Add custom servers in Settings → MCP Servers.
Ask it to read a ticket, write a doc, search Slack, run a command. The agent picks the right tools and asks permission when needed.
Everything you need to get Narwhal running. Takes about 5 minutes.
Before you install Narwhal, make sure you have these. If you're an RBKS employee on a Mac, you almost certainly already do.
M1, M2, M3, or M4 Mac. Intel Macs are not currently supported. Check: Apple menu → About This Mac → Chip.
You need an active Midway session. If you haven't run it today, open Terminal and run:mwinit -o
This is the same auth you use for Confluence, Jira, and every other internal tool.
Required for the Midway authentication step. You already have one if you access any internal Amazon tools.
That's it. No Python install, no Homebrew, no AWS CLI, no config files. Narwhal bundles its own Python runtime and all dependencies.
Open Terminal and run: mwinit -o — tap your security key when prompted. This refreshes your Midway session (needed for login and for Confluence/Jira access).
Download Narwhal-latest.dmg (~190 MB). Open the DMG and drag Narwhal to your Applications folder.
macOS blocks unsigned apps. Run this once in Terminal:
sudo xattr -cr /Applications/Narwhal.app
Enter your Mac password when prompted. You'll see some "Operation not permitted" warnings — those are normal and harmless.
Open Narwhal from Applications. It will open your browser for Midway login — tap your security key. After authentication, the app is ready to use.
Midway sessions expire (typically every 24 hours). If Narwhal shows an auth error, run mwinit -o again and click "Re-authenticate" in the app.
Narwhal checks for updates on launch and every 4 hours. When a new version is available, you'll see a banner: "Update Available — Restart to install." One click and you're on the latest version.
Common questions about Narwhal, how it compares to other tools, and what makes it different.
Kiro CLI is a general-purpose AI coding assistant that runs in your terminal. Narwhal is a native desktop app purpose-built for RBKS product and engineering workflows. Key differences:
| Narwhal | Kiro CLI | |
|---|---|---|
| Interface | Native macOS app with GUI — sidebar, preview panel, workspace browser, multi-tab conversations | Terminal-based CLI |
| Tool integrations | Pre-configured MCP servers for Ring Jira, Ring Confluence, RBKS Slack, NarwhalCloud, email, calendar | General-purpose tools, user configures MCP servers |
| Multi-agent | 4 orchestration patterns (supervisor, pipeline, swarm, context brief) with scoped subagents | Single agent with subagent delegation |
| Memory | Persistent memory across sessions — remembers your projects, preferences, team context | Per-session context only |
| Self-evolution | Observes its own behavior, learns from friction, proposes improvements | No self-modification |
| NarwhalCloud | Built-in access to 1,399 URF requirements, projects, PRDs via typed MCP tools | No requirements platform |
| Context brief | "/brief" queries Slack + Jira + Email + Confluence in parallel, synthesizes a status update | No cross-source synthesis |
| Target user | RBKS PMs, engineers, TPMs — people who live in Jira/Confluence/Slack | Software engineers writing code |
They're complementary. Use Kiro CLI for coding in your terminal. Use Narwhal for everything else — tickets, docs, status updates, requirements, cross-tool queries.
Amazon Q Developer is focused on code generation and IDE integration. Narwhal is focused on operational workflows — the work you do outside your IDE. Narwhal reads your Jira tickets, writes your Confluence docs, searches your Slack history, manages your requirements, and synthesizes information across all of them. Q Developer doesn't connect to any of these.
Your conversations, memory files, skills, and session history are stored locally at ~/.orcha/. The only external calls are: (1) AWS Bedrock for LLM inference (Claude), (2) MCP tool calls to Jira/Confluence/Slack APIs, and (3) NarwhalCloud API for requirements data. No data is sent to any third-party service.
Claude Sonnet 4 via AWS Bedrock. A lightweight side-model (Haiku) handles tool routing and memory selection to keep costs low. All inference runs through the RBKS Bedrock account — no API keys needed.
Yes, with a tiered permission system. Read-only tools (file reads, searches) auto-approve. Write tools (file edits, git commits) show a confirmation. Destructive tools (rm, deploy scripts) require explicit approval with a preview of what will run. You can customize these tiers in Settings.
This is macOS Gatekeeper blocking the unsigned app. Run this in Terminal:
sudo xattr -cr /Applications/Narwhal.app
You'll see some "Operation not permitted" warnings — those are normal. After this, Narwhal opens normally.
Your Midway session has likely expired. Run mwinit -o in Terminal, tap your security key, then relaunch Narwhal. If the app is already open, click Settings → Re-authenticate.
NarwhalCloud is a separate web service at cloud.narwhal.rbks.amazon.dev that stores structured product requirements, projects, and documents. It replaces Confluence for requirements tracking — every requirement is a queryable database row instead of a cell in an XML page. PMs can browse it in a browser; the desktop app's AI agent queries it through MCP tools. Read more in the Architecture guide.
Yes. Go to Settings → MCP Servers → Add Server. Point it at any MCP-compatible server (Figma, GitHub, databases, custom APIs). The agent automatically discovers and uses the new tools. Config is stored at ~/.orcha/mcp/config.json.
RBKS Core AI team. The NarwhalLLM SDK, desktop app, and NarwhalCloud platform are all built from scratch — no LangChain, no Strands, no third-party agent frameworks. See the architecture.
Download the app, connect your tools, and see what an AI agent can do when it actually knows your work.