llms.txt
Compare with other sites' llms.txt files →# GreatArrow.ai
> Stop copy-pasting client context into your AI. Your clients' Gmail, Calendar, Drive and meeting notes, in Claude, ChatGPT or Cursor. One private workspace per client. Meeting notes become tasks on your board. Answers from your data show their source. Once the memory is full, Great Arrow is what a business owner asks to find out what is actually going on in their company — what is late, who is waiting on them, where the month's money went — answered from their own calendar, documents, connected tools and invoices rather than from a guess. It puts AI to work on the follow-ups, the scheduling and the tracking, and shows what that AI costs. Underneath: persistent cross-model memory, agent orchestration, a 732-tool MCP surface, OAuth connections for Claude, ChatGPT, Gemini and Grok on the web, and API-token connections for desktop, editor and terminal clients such as Cursor and Kiro, with a four-layer memory model (working, episodic, semantic, procedural) syncing across 48 connectable providers including Google Workspace, Microsoft 365, Slack, Notion and GitHub.
## Core Architecture
- Four-layer memory: Working (short-term), Episodic (conversations), Semantic (knowledge), Procedural (how-to)
- Multi-workspace isolation with RBAC (Owner, Admin, Member, Viewer)
- Workspace kinds: Home (your one everything space), Project (a client or initiative kept to itself), Container (holds nested spaces, e.g. a product built from several repos) and Repo (made when a coding agent connects from that checkout) — with parent/child hierarchy
- Per-workspace resource scoping for integrations (select specific repos, channels, projects)
- AES-256-GCM encryption at rest, OAuth-secured MCP connections
## How the Product Is Organized
The signed-in product has two axes, and the left rail shows them in that order.
First, the workspaces you belong to. A workspace is what everything else is
scoped by, so it is the top section of the rail; nesting (a repo inside the
application it belongs to) is shown as indentation. Opening one is a READ — it
does not change which workspace the rest of the app is pointed at. Changing
that is a separate, labelled switch.
Second, the rooms, named for what a person is trying to do rather than for the
feature that implements it. Most rooms read the workspace you currently have
open; a few are deliberately wider — People and Meetings read every workspace
you belong to, and Your AI Agents, inside AI Fleet, reads your own AI connections. Each room says
which above its title, so an assistant pointing a user somewhere should name
the room, not a legacy path:
- Today (/) — the day's briefing, what needs you, and live agent activity
- Ask AI (/ask) — one grounded conversation with every connected AI, cited back to the source
- Work (/work) — the task board across the workspaces you are reading
- Plan & Goals (/work/planning) — goals, pace and cost to complete, rolled up from real cards rather than a guess
- Money (/business) — revenue, spend and net per workspace, where every total can name its sources
- People (/contacts) — everyone you work with, read across every workspace you belong to, each row carrying the workspace it came from
- AI Fleet (/team) — the AIs you talk to, your specialist agents, the tool workbench, voice, and the war room
- Meetings (/work/meetings) — every meeting touching the work you are reading, with its own workspace named
- Structure (/workspaces) — how your workspaces nest, and what each one rolls up into
- Reports (/reports) — every saved selection of context: dashboards, client pages, decks and rollups
Everything else hangs off those rooms: revenue-facing surfaces (leads,
pitch, business dashboard) sit in a "Grow your business" drawer group, developer
and workspace-administration surfaces sit in an "Advanced" group beside it, and
account-level destinations (account, workspaces, notifications, usage and costs,
billing, invites, support) live behind the avatar menu.
Pre-2026 URLs still resolve — every one of them redirects into the room that
now owns it, so a stored link never breaks. They are deliberately NOT listed
here: an assistant reading a retired path in this file will hand it to a user
as a destination, which is the one thing the redirects cannot undo. Name the
room and its path from the list above.
## Capabilities
- Scoped external sharing: hand someone one project by link — a revocable, expiring join link that makes them a viewer of that workspace and nothing else, or a page published behind a shared password — and choose which categories travel with it. Email content, vault secrets, calendar details and contact records are off by default, and the filter runs server-side, so an MCP tool call is scoped exactly as the screen is
### Memory & Knowledge
- Semantic search and recall across all connected AI clients
- Bulk memory operations (create, update, tag, delete)
- Memory revision history and rollback
- Document ingestion from Google Docs, Sheets and Drive (all of Drive if you grant read access, otherwise only the files you pick), OneDrive and Dropbox
- Document intelligence: PDF/DOCX/XLSX/image ingestion with vision-model OCR fallback, passage-level semantic search
- Auto-memorization of important decisions and facts
- Entity knowledge graph: the people and organizations a memory names are linked to it with a typed relation (mentions, works at, attended, authored, discussed), and the graph is browsable as source clusters and entity hubs, as one memory's immediate links, or as a health view of contradictions, version chains and orphans
- Cross-client conversation persistence: transcripts from any connected AI feed the shared memory
- Memory for local models: LM Studio connects over MCP and reads the same workspace memory; the model runs on the user's machine, and on this path only its memory lookups reach GreatArrow.ai. Guide: /guides/give-your-local-model-a-memory. The separate gad CLI path for Ollama and LM Studio sends more, as described at /features/local-models
### Agent Orchestration
- Multi-step task decomposition with specialist routing
- 21 specialist agents for domain-specific tasks
- Custom agent creation and configuration
- Scheduled agent runs (cron-style recurring automation)
- Agent thread history, checkpoints, and rollback
- Todo/kanban management with agent assignment
### Integrations (48 providers)
- Google Workspace: Calendar, Docs and Sheets; Gmail and Drive reading when you grant it on Google's consent screen (otherwise Drive stays limited to the files you pick); mail is sent through your connected Gmail
- Microsoft 365: Outlook, Teams, OneDrive, SharePoint
- Dev tools: GitHub (issues, PRs, reviews, merges, Actions), GitLab, Jira, Asana, Linear
- Communication: Slack (search, threads), Discord, Notion, Confluence
- CRM: GoHighLevel (contacts, pipelines, opportunities, workflows, appointments)
- Analytics: Google Analytics 4, Microsoft Clarity
- Design: Canva — start a blank design or seed one with an uploaded asset, have an agent autofill a brand template with real data, then export to PDF, PNG or JPG (brand templates and autofill require a Canva Enterprise account)
- Other: Zoom, Vercel, Supabase, Dropbox
### Calendar & Meetings
- Availability queries from calendars synced in the background every 15 minutes
- Auto-scheduling with contact resolution and agenda generation
- Meeting templates and public booking links
- Post-meeting finalization with action item extraction
- Meeting briefings with per-participant context
### Communication
- Email sending via Gmail or Outlook
- Slack message search and thread replies
- GitHub issue/PR comments and reviews
### Development
- GitHub: create issues, PRs, reviews, merge PRs, dispatch Actions workflows
- GitLab: project management
- Code execution (sandboxed Python and JavaScript)
- Vercel project management and deployment
- Vibe coder software-building workspace: shared project memory, agent task board, GitHub/Vercel-aware workflows, cost guardrails, and SEO/agent-discovery launch pages
### Productivity
- Todo/kanban board with priorities, labels, checklists, due dates
- Accomplishment tracking with skill tagging
- Progress report generation
- Resume generation from accomplishments
- Lead pipeline: the same contact records carry a pipeline stage and a deal value, and the board totals the value sitting in each of the Lead, Qualified, Proposal and Won lanes — a lightweight CRM in the records the People room already reads, not a second system to keep in sync
## Pricing
Free 14-day trial — no card. After the trial, choose a paid plan to keep going. Free access may end after notice. Ending the trial does not delete saved records; your retention settings still apply. Nothing below is required to try the product.
- Free Trial: $0 — 14 days, no credit card, 5M embed + 100K chat tokens; every new account starts here
- Free: $0 — 2M embed tokens/month (standard included allowance). After the trial, choose a paid plan to keep going. Free access may end after notice. Ending the trial does not delete saved records; your retention settings still apply.
- Personal: $39/month — 50M embed + 700K chat tokens, all integrations
- Pro: $119/month — 200M embed + 1.8M chat tokens, full agent orchestration
- Enterprise: custom pricing — BYOK encryption, governed access across shared workspaces, custom token caps
## How Great Arrow Compares
Full comparison, including where each alternative is the better choice: https://www.greatarrow.ai/compare
- vs built-in AI memory (ChatGPT, Claude): built-in memory is free and needs no setup, but it belongs to that vendor and does not follow you to a second tool. Great Arrow works across 21 AI clients plus web, terminal and mobile, and the memory is an asset the company owns and can export.
- vs a memory library (Mem0, Zep, Letta): libraries are infrastructure you build a product on. Great Arrow is the running system — 48 providers, tenant isolation, permission enforcement and 732+ agent tools already wired.
- vs enterprise search (Glean and similar): enterprise search answers questions well and has strong permission-aware indexing, and leading products now offer agents that take actions in connected apps. Great Arrow is a shared memory that the AI apps you already use read and write over MCP. Teams often run both.
- What is distinctive: facts retire with a date and a replacement rather than accumulating silently, and a failed search is reported as failed rather than returned as an empty result.
## Security
- AES-256-GCM application-layer encryption for memories and integration credentials; uploaded files in encrypted storage
- Tenant isolation is enforced in the application layer: every workspace-scoped query resolves its tenancy server-side and is gated in CI; Postgres row-level security is enabled on every table as defence in depth
- OAuth 2.1 + PKCE with dynamic client registration for web AI clients
- API tokens stored as SHA-256 hashes, shown once at mint, revocable per client
- Prompt-injection screening of all integration-synced content before it enters memory
- Shared workspaces with Owner/Admin/Member/Viewer roles on every paid plan; Enterprise adds bring-your-own-key encryption
- Cost governance: an agent's model is chosen by assessed task complexity rather than fixed to the most expensive one, and a workspace can carry its own monthly token ceiling — the tighter clamp, applied before any granted credits are drawn — so one workspace cannot drain the account behind it. Your plan price is the most you pay
## Access
MCP (Model Context Protocol) endpoint: https://www.greatarrow.ai/api/mcp (Streamable HTTP; OAuth 2.1 with dynamic client registration, or Bearer API token). Setup: create an account at https://www.greatarrow.ai/sign-up first. The per-client install guides at /install are public; signing in there mints the personal install code each guide uses.
21 AI clients with first-class install paths:
- Claude Desktop — API token, one-click install
- Claude.ai (web) — OAuth 2.1 connector, one-click install
- ChatGPT — OAuth 2.1 connector, guided install
- Gemini (web) — OAuth 2.1 connector, guided install
- Grok (web) — OAuth 2.1 connector, guided install
- Cursor — API token, one-click install
- GitHub Copilot — API token, one-click install
- Claude Code (CLI) — API token, one-liner install
- OpenAI Codex (CLI) — API token, one-liner install
- Gemini CLI — API token, one-liner install
- Kiro — API token, guided install
- Kiro Terminal (Codex CLI) — API token, guided install
- Windsurf (Codeium) — API token, guided install
- Zed — API token, guided install
- JetBrains IDEs — API token, guided install
- Anti-Gravity — API token, guided install
- Hermes Agent (Nous Research) — API token, guided install
- OpenClaw — API token, guided install
- LM Studio — API token, guided install
- DeepSeek Harness — API token, one-liner install
- Muse (CLI) — API token, guided install
Any other MCP-compatible client can connect with the endpoint above.
## Guides
Practical setup guides (problem-space how-tos), one per buyer question. Hub: https://www.greatarrow.ai/guides
- [How to Run Multiple AI Agents on One Codebase](https://www.greatarrow.ai/guides/how-to-run-multiple-ai-agents-on-one-codebase): Five agents on one repo means overwritten branches and phantom claims. Here's the workflow — claims, isolated worktrees, resumable handoffs — that stops it.
- [Give Your Local Model a Memory (Ollama, LM Studio)](https://www.greatarrow.ai/guides/give-your-local-model-a-memory): Run open models on your own machine and keep what they learn: persistent memory for LM Studio, Ollama and any local MCP client. Inference stays local.
- [Give ChatGPT and Claude Company Data, With Citations](https://www.greatarrow.ai/guides/ai-access-company-data-with-citations): How to let ChatGPT and Claude search Gmail, Slack, Drive and Notion through a permission-aware layer that returns the source record behind every answer.
- [Share Context Across Claude, ChatGPT and Cursor](https://www.greatarrow.ai/guides/share-context-across-claude-chatgpt-cursor): Native memory does not travel between AI clients. Here is how a shared context layer over MCP gives Claude, ChatGPT and Cursor the same facts about your work.
- [Shared Memory Across Claude, ChatGPT, Gemini, Cursor](https://www.greatarrow.ai/guides/shared-memory-claude-chatgpt-gemini-cursor): Conversation history, personal memory and company context are three different things. Which ones can be shared across AI clients today, and what the limits are.
- [Stop Re-Explaining Your Company to Every AI Tool](https://www.greatarrow.ai/guides/stop-re-explaining-your-company-to-ai): A copyable context template that stops the re-briefing, where manual context files break down, and what it takes to keep the same facts current across AI tools.
- [What Your AI Stack Actually Costs](https://www.greatarrow.ai/guides/what-your-ai-stack-actually-costs): Adding up your AI invoices gets the number wrong in both directions. A three-bucket model for real AI spend, and the four traps that corrupt it.
- [AI Meeting Scheduling With Your Real Calendar](https://www.greatarrow.ai/guides/ai-meeting-scheduling-with-your-real-calendar): Let an AI client schedule real meetings on your calendar — previewing times before creating, checking invitee availability, and sending invites.
- [Connect ChatGPT to Your Business Data](https://www.greatarrow.ai/guides/connect-chatgpt-to-your-business-data): Connect ChatGPT to your own business data through a custom connector authenticated with OAuth 2.1 and PKCE, so it searches before it answers.
- [Connect GitHub So Your AI Knows Your Repos](https://www.greatarrow.ai/guides/connect-github-so-your-ai-knows-your-repos): Connect GitHub so any AI client can search your repos, with issues, PRs, and commits synced into a workspace scoped to that specific repository.
- [Connect Gmail to Claude](https://www.greatarrow.ai/guides/connect-gmail-to-claude): Connect GreatArrow.ai to Claude as an MCP connector, then link Gmail through OAuth so Claude can search and reference your inbox in conversations.
- [Connect Slack to Your AI Assistant](https://www.greatarrow.ai/guides/connect-slack-to-your-ai-assistant): Connect Slack to your AI assistant through OAuth so any connected client can search messages, follow threads, and reply — scoped to the channels you allow.
- [Give Cursor Persistent, Project-Aware Memory](https://www.greatarrow.ai/guides/give-cursor-persistent-memory): Give Cursor project-aware, persistent memory with a one-click deeplink install and a REPO workspace that auto-provisions from your repository URL.
- [One Memory Across Every AI Client](https://www.greatarrow.ai/guides/one-memory-across-every-ai-client): Every AI client you use keeps its own memory. Here is how MCP gives Claude, ChatGPT, Cursor, and your other AI tools one shared, persistent memory instead.
- [Persistent Memory for Claude Desktop](https://www.greatarrow.ai/guides/persistent-memory-for-claude-desktop): Give Claude Desktop persistent memory it can search across sessions, using the one-click .mcpb installer, the universal installer, or a manual config file.
- [Share Memory Between Claude and ChatGPT](https://www.greatarrow.ai/guides/share-memory-between-claude-and-chatgpt): Step-by-step setup to connect both Claude and ChatGPT to the same memory server, so a fact stored in one is instantly searchable from the other.
- [What Is MCP? A Guide for Teams](https://www.greatarrow.ai/guides/what-is-mcp-for-teams): An evaluator-level explanation of the Model Context Protocol — what it does, remote vs. local servers, OAuth vs. tokens, and what to check in a server.
## Software Builders
Marketing page for vibe coders and AI-assisted software founders: https://www.greatarrow.ai/vibe-coders
Use it to understand Great Arrow's software-building workflow: project memory, 21 specialist agents, 732 MCP tools, GitHub/Vercel workflows, kanban tasks, cost intelligence, report-to-board feedback, structured SEO metadata, and llms.txt/MCP agent discovery.
## ArrowDeck (macOS desktop app)
ArrowDeck by GreatArrow.ai is a native macOS app for running AI coding agents: https://www.greatarrow.ai/arrowdeck
Download page (artifact, install steps, release notes): https://www.greatarrow.ai/download
- ONE surface, not a set of rooms: a titlebar with a single ask bar, a filter list, one stream with a row per thing in flight, an inspector for the selected row, a terminal drawer that opens over the stream, sheets for everything configural (setup, account, agents, workspaces, storage, machines, pipeline, money, people, diagnostics), and a ⌘K command palette
- The stream is filtered by 6 lanes, which sort everything by whose move it is:
- ⌘1 Yours to decide — Nothing moves until you say so.
- ⌘2 An agent is on it — Running now. No input needed.
- ⌘3 Moving on its own — Checks and claims. No input needed.
- ⌘4 Waiting on someone — A person, not a machine.
- ⌘5 Ready to go live — Checks passed. Needs your go.
- ⌘6 Landed — Live and verified.
- A count with no completed read behind it renders as a dash, never a zero; "could not count" and "there are none" are different statements
- Runs any number of coding agents at once, each in its own git worktree on its own branch — two agents can never share a checkout, and the one you have open in your editor is never touched
- Snapshots an agent's uncommitted work to private git refs while it runs and on exit, death, or usage limit; recovery restores a snapshot into a fresh copy
- Tracks each task from claimed through pull request, checks, merged, and deployed-and-verified, offering only the action that is safe right now
- Installs and connects the agent apps themselves (Claude Code, Codex, Gemini CLI, Cursor and more), merging their settings rather than clobbering them
- Identity profiles keep multiple AI accounts, GitHub identities and workspaces separate; am