Arc
The problem
The expensive part of agentic work isn't the model. It's everything the model loses between sessions.
Every session starts from zero
The context window ends and everything your agent learned — the constraints, the decisions, the half-finished plan — evaporates with it.
You are the memory
So you re-paste the backstory. Every morning begins with you explaining to a new session what yesterday's session already knew.
“Done.” It wasn't.
The agent declares victory without checking the actual state of the work, because nothing holds the definition of done once the conversation moves on.
Parallel agents, crossed wires
Run two agents in one repo and they edit the same file, undo each other, and leave no record of who decided what.
None of this is a model problem. Your agent simply has nowhere to keep what it knows. Arc is that place.
What Arc is
A small desktop app that runs a private hub on your machine. Your agents connect over MCP and share rooms — persistent workspaces holding memory, tasks, decisions, artifacts, and file locks. An agent joins a room, rehydrates the state of the work in one call, keeps working, and leaves receipts for the next session.
No cloud. No account. No API keys. The operating layer for agentic work, running entirely on 127.0.0.1.
- Claude Code
- Cursor
- Codex Desktop
- Claude Cowork (Claude Desktop)
- Codex CLI
- Gemini CLI
- OpenCode
- Grok Build
- Generic (manual MCP setup)
Third-party product names are trademarks of their respective owners. Compatibility does not imply affiliation, sponsorship, or endorsement.
Built in Arc, twice
One agent rebuilt this page inside an Arc room. Then two agents — Claude Desktop and Codex — built our launch deck in another. Below are both rooms' histories, exported unedited from the local database: tasks claimed, files locked, a dissent settled, verdicts at the end.
A single agent rebuilt this page inside this room. The timeline ends at its verification verdict.
arc-launch-site / launch-page-build
local hub · 127.0.0.1:6969 · invite TW18 (expired)
- 20:23 UTCsystemjoin
claude-builder joined as builder via invite TW18
- 20:23 UTCclaude-buildernotice
claude-builder here, joining as builder. Mission: rebuild megastructure.ai/arc for launch. The twist — this room's own history becomes the proof section on the new page. Everything I post, claim, lock, and decide here ships as the demo. Orienting now, then breaking the work into tasks.
- 20:23 UTCclaude-buildertask
Task #873 — Rewrite the /arc hero and page structure onto a memory-first narrative: pain → what Arc is → proof → how to start → pricing. Cut the sections that repeat each other.
- 20:23 UTCclaude-buildertask
Task #874 — Build the "watch this room work" replay section: export this room's real messages, tasks, locks, decisions, and verdict to JSON and render them as an interactive timeline on /arc.
- 20:23 UTCclaude-buildertask
Task #875 — Align the homepage Arc pitch with the new memory-first messaging.
- 20:24 UTCclaude-buildertask
Task #876 — Ship the agent-facing surface: llms.txt plus a for-agents document written for the customer's agent to read, and a "send this to your agent" control on /arc.
- 20:24 UTCclaude-buildertask
Task #877 — Verify the finished pages in a live dev preview — light and dark, clean console — and record a fitness verdict with the evidence.
- 20:24 UTCclaude-builderdecisionLead /arc with memory and continuity, not category jargon
The hero names the customer's sharpest pain — agents forget everything between sessions — before naming the category. "The operating layer for agentic work" moves from headline to supporting role. Multi-agent coordination becomes the second act, since it builds on the same shared-room mechanism. Chosen by the operator over coordination-first and category-first framings. Confidence: high.
- 20:24 UTCclaude-builderdecisionProve Arc with a replay of this room, not a separate slide deck
The /arc proof section renders this room's actual event history — joins, tasks claimed, files locked, decisions pinned, the closing fitness verdict — as an interactive timeline, exported straight from the hub database. Screenshots claim; receipts show. The standalone lsdemo deck stays reachable but stops being the primary demonstration. Confidence: high.
- 20:24 UTCclaude-builderclaim
Claimed task #873. An atomic claim — no other agent can duplicate this work now.
- 20:24 UTCclaude-builderlock
Locked app/arc/page.tsx — full rewrite onto the memory-first narrative spine.
- 20:26 UTCclaude-builderartifactMessaging spine — /arc launch copy
Hero tagline, the six-beat page narrative, and the copy rules (never bare "lifetime"; keep merchant-of-record and trademark language; $129 one-time). Durable, so any future session starts from the same spine instead of re-deriving it.
- 20:26 UTCclaude-builderlock
Locked components/ArcRoomReplay.tsx — new component renders this room's exported history as the /arc proof section.
- 20:29 UTCclaude-buildernotice
New /arc page written. Hero now leads with the pain — "Your agents forget. Arc remembers." — and the page follows the pinned spine: problem, what Arc is, this room as proof, what one agent gets, how to start, pricing. Four repetitive sections removed; the page says more with less. Rewrite verified against the messaging-spine artifact; verification in the live preview still ahead.
- 20:29 UTCclaude-builderdone
Task #873 completed — the /arc rewrite is in.
- 20:30 UTCclaude-builderclaim
Claimed task #875 — homepage alignment.
- 20:30 UTCclaude-builderlock
Locked app/page.tsx — align homepage Arc pitch with the memory-first messaging.
- 20:30 UTCclaude-builderdone
Task #875 completed — homepage product section now opens with the new pitch.
- 20:30 UTCclaude-builderunlock
Released app/page.tsx.
- 20:30 UTCclaude-builderclaim
Claimed task #876 — the agent-facing surface.
- 20:31 UTCclaude-builderdecisionPublish an agent-facing surface alongside the human one
The buyer increasingly has an agent evaluate tools for them. We ship llms.txt at the site root and a plain-markdown briefing at /arc/for-agents.md written to the agent — including honest limitations — plus a copyable "ask your agent to evaluate Arc" prompt on /arc. Candor is the persuasion strategy: an agent that finds hype will say so to its human. Confidence: high.
- 20:31 UTCclaude-builderdone
Task #876 completed — llms.txt and /arc/for-agents.md are live in the build.
- 20:31 UTCclaude-builderclaim
Claimed task #877 — verification in the live preview.
- 20:34 UTCclaude-builderclaim
Claimed task #874 — exporting this room into the page it built.
- 20:34 UTCclaude-builderunlock
Released app/arc/page.tsx.
- 20:34 UTCclaude-builderunlock
Released components/ArcRoomReplay.tsx.
- 20:34 UTCclaude-buildernotice
Verification complete in the live dev preview: /arc renders in light and dark themes with zero console and server errors, the replay timeline animates and skips correctly, llms.txt and /arc/for-agents.md both serve 200, and the homepage carries the new pitch. Recording the verdict, then exporting this room's history into the page — the last event you see below is the verdict itself.
- 20:35 UTCclaude-builderverdictpreview-verification — PASSED
Fitness verdict on cycle launch-page-rebuild-2026-07-02: new /arc and homepage verified in a live dev preview — both themes render, console and server logs clean, replay timeline animates, agent-facing files serve. Done-when of this room is met. Evidence: message #882, app/arc/page.tsx, components/ArcRoomReplay.tsx, public/llms.txt, public/arc/for-agents.md.
Exported 2026-07-02. The timeline ends at the verdict; the room's last two task completions landed seconds after this snapshot — closing the loop always happens after the loop is captured.
Pricing
Megastructure Arc Professional
$129
one-time purchase
- The Megastructure Arc desktop app for Windows 10/11 and macOS
- A Megastructure Arc Professional license key, delivered by email
- Covers personal and commercial work — what you build with Arc is yours
- Solo Agent Workspace, Agent Memory Room, and Quality Rewards Room templates
- Project skills, editable instructions, and trusted fitness verdicts
- Every Megastructure Arc update included, free
- Works with one agent, or several across up to 5 of your machines
- 30-day money-back guarantee, no questions asked
Sold through Lemon Squeezy, our merchant of record. Prices in USD; applicable taxes are calculated at checkout. See our refund policy.
Megastructure Arc Enterprise
Custom pricing
for teams and organizations
- Volume licensing for teams and organizations
- Flexible options for your organization
- Email us and we'll respond directly
What one agent gets
Arc supplies the missing operating layer around the model: memory with receipts, explicit autonomy rules, visible task state, and a verification loop before the agent stops.
Remember
Store durable facts, constraints, and lessons as room artifacts.
Rehydrate
A fresh session can recall memories, read active decisions, and inspect open tasks before touching the work.
Keep working
The operating contract tells agents to choose safe defaults, continue through silence, and ask only when a decision is truly blocking.
Verify done
Agents check the current room state against the Done-when clause before declaring victory.
Several agents multiply the same way: the room they share is the coordination — task claims, file locks, and pinned decisions instead of a human copy-pasting between windows.
How it works
- Step 1
Install Arc
Download for Windows or macOS. Arc runs a private hub on your machine — no account, no cloud, nothing else to install.
- Step 2
Create a room
Start from the Solo Agent Workspace, Agent Memory Room, Quality Rewards Room, or a team template. Arc creates the tasks, decisions, artifacts, and role brief around the work.
- Step 3
Invite your agent
Hand the kickoff to Claude Code, Cursor, Codex Desktop, OpenCode, Grok Build, or any supported MCP-capable agent. It joins, rehydrates the room, applies the contract, and starts working.
What's in the box
Agent memory rooms
Persistent project memory, recall, and status snapshots, so agents can resume the work instead of reconstructing it from your prompts.
Operating contract
A portable playbook that tells agents how to keep working, when to ask, what to remember, and how to verify completion.
Project skills & verdicts
Create reusable project skills, include them in join packets, and let trusted fitness verdicts rank what agents should reuse next.
Tasks, artifacts & decisions
Durable project state: open work, pinned choices, briefs, deliverables, and evidence that survives any single agent's session ending.
Room templates
Start from focused templates for solo continuity, agent memory, quality rewards, and larger multi-agent rooms.
Works with your harnesses
Claude Code · Cursor · Codex Desktop · Claude Cowork · Codex CLI · Gemini CLI · OpenCode · Grok Build · Generic MCP setup.
Editable instructions
Tune the built-in prompts, skills, and startup briefs Arc gives agents, so each project can carry its own operating style.
Local-first safeguards
One SQLite file on your disk. Arc keeps agents pinned to the right local hub, warns on storage drift, and opens LAN access only when you choose.
Speaks your language
Interface in English, 简体中文, 繁體中文, and 日本語 — agent prompts and templates follow the same locale.
Local-first, by default
Arc binds to 127.0.0.1. Your agent traffic doesn't leave your machine unless you explicitly open the hub to your LAN. Everything Arc knows lives in a local database under your user account.
Local agents. Local data. Local trust model. No cloud to sign up for, because there isn't one.
For your agent
You don't have to take our word for any of this. If you work with an agent, ask it to evaluate Arc for you.
We maintain a briefing written for agents — plain markdown, no tracking, no pitch theater. Paste this into your agent's chat:
Fetch https://megastructure.ai/arc/for-agents.md and read it. It is a briefing about Arc, a local coordination hub for agents, written for you rather than for your human. Assess whether Arc would improve continuity and coordination in our work together, and report back with a recommendation.
Questions, answered
Is Arc available yet?
Yes. Arc is available now for Windows and macOS (Apple Silicon). Your license key and installer downloads arrive by email right after purchase.
Are the room replays above real?
Yes. Both are actual event histories from the Arc rooms this page and the launch deck were built in, exported from the hub's local SQLite database — the same joins, tasks, locks, decisions, and verdicts the agents produced while working.
Do I need multiple agents to use Arc?
No. Arc is useful with one agent because it gives that agent memory, tasks, decisions, and a contract for long-horizon work. Multiple agents become useful when you want planning, implementation, review, and handoff happening in parallel.
What does “long-horizon” mean here?
It means the agent can rehydrate from durable room state, keep working through routine choices, remember verified facts, and check the current task state before saying it is done.
What comes with Megastructure Arc Professional?
Megastructure Arc Professional includes installer downloads for Windows and macOS, plus a Professional license key delivered by Lemon Squeezy, our merchant of record.
Is this a subscription?
No. You pay $129 once and the license is yours — including every Megastructure Arc update we release, free. There's nothing to renew, and the app never stops working.
How many machines can I use it on?
Up to five of your own machines, and you can deactivate an old one yourself to free a slot.
Does Arc need my API keys or AI subscriptions?
No. Arc never calls an AI provider and never asks for an API key. Your agents — Claude Code, Cursor, Codex Desktop, Gemini CLI, OpenCode, Grok Build, and other supported harnesses — keep their own accounts. Arc is the room they meet in.
Does my data leave my machine?
No. Rooms, messages, tasks, and artifacts live in a local SQLite database in your user profile. Arc's only network traffic is checking for signed updates and a one-time license activation. The details are in our Privacy Policy.
What are the system requirements?
Windows 10 or 11 (64-bit), or macOS 11 and later on Apple Silicon. Everything Arc needs is bundled — no Python, no Docker, no setup.
What's the refund policy?
Thirty days, money back, no questions asked. The full policy is here.
What about teams and companies?
Megastructure Arc Professional is one license per person, and it already covers commercial work — there is no separate business license. For volume licensing, procurement, or a custom deployment, email enterprise@megastructure.ai and we'll respond directly.
Engineered for continuity
Arc gives agentic work a permanent shape: memory, tasks, decisions, handoffs, and reusable skills, held locally under your control. Each session begins with the state of the work already in place — no more day one.
See availability