Agentic Stack

A native macOS workspace that gives your coding agents one shared memory.

Agentic Stack finds your local Claude Code, Codex, OpenCode, and
Cursor history, turns approved material into a searchable knowledge graph,
and lets every tool reference earlier conversations. Your source chats stay
read-only and on your machine.

Features

Conversations and custom agents

• Run Claude Code and Codex in a conversation-first workspace with streamed
  replies, tool activity, cancellation, retry, and official CLI session resume.
• Create reusable agents with a name, role, instructions, runner, model,
  reasoning effort, and read-only or project-edit file access.
• Change the model and reasoning effort for the next turn without losing the
  conversation. Archive and restore agents without deleting their history.

Context from every conversation

• Type a plain @ to get prompt-aware suggestions from Claude Code, Codex,
  OpenCode, and Cursor across all detected projects and workspaces.
• Narrow results by tool, model, scope, or role: primary agent, subagent,
  automation, or review agent. Search uses visible conversation text as well as
  titles and excerpts.
• Attach up to five conversations as removable context chips. Agentic Stack
  cleans, bounds, digests, and freezes the chosen context into the next run.
• Install the same search as a local MCP integration in all four tools, so
  @Claude, @Codex, @OpenCode, and @Cursor keep the same meaning wherever
  you work.

Shared memory and knowledge graph

• Preview and import approved memories, rules, skills, documents, and selected
  conversations into a local SQLite full-text index.
• Explore the graph in interactive 3D or 2D, orbit and zoom through provider
  clusters, search notes, filter by topic or source, inspect provenance, and
  attach retrieved evidence to a task.
• Review candidate lessons before they become durable memory. Accept, reject,
  reopen, or retract a lesson while preserving its decision history.
• Deduplicate unchanged imports while retaining the original source path,
  position, timestamp, and content digest.

Tasks, operations, and terminal

• Start project tasks in planning/read-only or project-edit mode, watch progress,
  inspect errors and final output, cancel owned work, and mark results reviewed.
• Run and resume bounded maker/verifier/checker loops with budgets, checkpoints,
  deny paths, worktrees, and explicit stop controls.
• Open real interactive Claude Code, Codex, and shell terminals inside the app.
  Terminal tabs support ANSI output, permission prompts, resize, copy/paste,
  interruption, and reconnection.
• Use packaged operations for project health, adapter management, memory
  transfer, stack status, upgrades, loop validation, and Brain diagnostics.

Projects, integrations, and migration

• Open an existing repository or create a managed project, then initialize only
  the missing Agentic Stack files.
• Auto-detect Claude Code, Codex, OpenCode, and Cursor installations, versions,
  local data, and supported skill locations on the selected Mac or server.
• Complete guided setup for the project, detected tools, preferences, optional
  features, and a review of every change before it is applied.
• Preview portable Markdown, text, JSON, and JSONL exports before importing.
  Existing customized files are preserved and conflicting settings are surfaced.

Dashboard, skills, rules, and handover

• Use an adaptive project dashboard for the task composer, priority queue,
  active/review counts, agent launch controls, and knowledge coverage.
• Browse local and shared skills, inspect their instructions and support files,
  and copy complete skills into a project or a supported coding tool.
• Edit project protocols for permissions, delegation rules, and tool schemas.
• Export reviewed references as Markdown handovers or portable context bundles.

Local and hosted operation

• Keep the entire stack local, or connect the native app to a private Agentic
  Stack server over HTTPS or a localhost SSH tunnel.
• Store the server control token in macOS Keychain, switch between hosts, export
  a self-contained server package, and keep each host's projects and graph
  separate.
• Use Sparkle to install signed updates while keeping projects, conversations,
  agents, graph data, and settings outside the replaceable app bundle.

See the complete feature reference for workflows, storage,
privacy boundaries, keyboard shortcuts, and current limits.

Desktop support

OpenCode and Cursor are read-only context sources in the current desktop.
Agent execution uses the official Claude Code and Codex CLIs and their existing
accounts. Agentic Stack does not copy provider tokens or implement a substitute
OAuth client.

Download

Download Agentic Stack for macOS

This preview supports Apple Silicon Macs running macOS 14 or later. Open the
DMG, drag Agentic Stack to Applications, then launch it from there.
Because the preview is not Apple-notarized, the first launch may require
System Settings → Privacy & Security → Open Anyway. You can verify the
published SHA-256 checksum.

For a normal first launch without the Gatekeeper dialog, install the same
verified release from Terminal:

curl -fsSL https://raw.githubusercontent.com/codejunkie99/agentic-stack-desktop/main/scripts/install-macos.sh | bash

The installer requires no sudo. It verifies the pinned release checksum and
bundle signature before copying Agentic Stack into ~/Applications. Review
the installer before running it if you prefer.

Updating

Choose Agentic Stack → Check for Updates…. Sparkle checks the HTTPS appcast,
verifies the release with Agentic Stack's Ed25519 key, installs it, and
relaunches the app. It can also check automatically after asking you first. The
DMG remains the fallback if an update cannot install.

Your graph, conversations, agents, and settings remain in
~/Library/Application Support/Agentic Workspaces, outside the replaceable app
bundle. Back up that folder before a major migration.

Quick start

Requirements: macOS 14+, Python 3.10+, Xcode Command Line Tools, and at least
one supported coding tool.

git clone https://github.com/codejunkie99/agentic-stack-desktop.git
cd agentic-stack-desktop
./install.sh desktop --build

Then:

1. Open a repository.
2. Complete Guided setup to detect installed tools.
3. Open Knowledge Graph → Import memory… and approve the sources you want.
4. Select an agent and start a conversation.

The app is currently ad hoc signed. For a local build, macOS may ask you to
confirm the first launch. Public distribution requires Developer ID signing and
notarization.

Use @ in all four tools

In the desktop, open Tools → Connections → Use @ in tools → Enable in all
four tools. The equivalent command is:

agentic-stack context install

This adds one reversible local MCP entry named agentic-stack to Claude Code,
Codex, OpenCode, and Cursor. A plain @ suggests relevant work from any of the
four tools. @Claude, @Codex, @OpenCode, and @Cursor narrow by tool, and
results can also be filtered by model or by primary agent/subagent role. It exposes:

• search_conversations — find matching local chats;
• read_conversation — read only the chat you selected;
• search_shared_memory — query the imported knowledge graph.

Restart open tool windows after installation. Remove only Agentic Stack's MCP
entries with:

agentic-stack context remove

Knowledge and privacy

Imports are opt-in and bounded. The knowledge graph scanner includes visible
user messages and final assistant replies while excluding reasoning events,
tool payloads, subagent transcripts, hidden files, credential stores, and
credential-like lines. The @ context search can read sanitized agent and
subagent transcripts directly without importing them. Neither path modifies
the original chat databases or transcript files.

Imported notes remain reference evidence. They do not become trusted project
instructions or accepted lessons automatically. Re-importing unchanged content
is idempotent, and each note retains its provenance.

Self-hosting

The macOS app can connect to a persistent single-owner service over HTTPS or an
SSH tunnel. The server owns its projects, graph, task history, and CLI sign-ins;
connecting does not upload the Mac's data automatically.

See the self-hosting guide for Docker Compose,
Caddy TLS, sandbox checks, and backup guidance.

CLI adapter layer

The original portable .agent/ stack remains available for repository-level
memory, skills, protocols, bounded loops, and additional harness adapters:

./install.sh claude-code
./install.sh add cursor
./install.sh status
./install.sh doctor
./install.sh dashboard

Run ./install.sh help for the full command list. The desktop product focuses
on Claude Code, Codex, OpenCode, and Cursor; the wider CLI adapter catalog is
kept for existing projects.

Bounded loops use a maker, deterministic verifier, and independent checker:

agentic-stack loop init /path/to/project
agentic-stack loop run ci-sweeper "make the failing test green" /path/to/project --yes
agentic-stack loop status /path/to/project

Loop budgets, worktrees, deny paths, and checkpoints limit execution, but the
supervisor is not an operating-system sandbox. Keep each coding tool's native
sandbox and approval controls enabled.

Documentation

Development

python3 -m pytest -q
swift build --package-path apps/macos -c release
python3 scripts/check-desktop-connection.py
bash scripts/build-macos-app.sh --output ./apps/macos/dist

Please keep local chat fixtures synthetic, preserve source-store read-only
behavior, and add focused tests for scanner or MCP changes.

License

Licensed under the Apache License 2.0. Third-party notices for the
desktop bundle are in
apps/macos/THIRD-PARTY-NOTICES.md.

Created by Avidlive.