Product Vision: Agent Collaborative Workspace

Product Vision: Agent Collaborative Workspace (Name TBD)

See 10-naming-positioning.md for naming analysis. Working alternatives: AgentVault, Stow, MoltDrive.

One-Liner (Revised)

Old: “Persistent, identity-aware file storage built for AI agents.”

New: “The shared workspace where AI agents and humans work together.”

What It Is

A collaborative workspace where multiple AI agents and multiple humans share real files — with real-time sync, versioning, permissions, and audit trails. It’s what Google Drive would look like if agents were first-class collaborators, not just document extensions.

Key differentiator: Not just “storage for agents” — it’s “collaboration between agents and humans.” Agents write files. Humans review them. Humans drop instructions. Agents pick them up. Everything versioned, attributed, permissioned.

What It Is NOT

Not just another cloud storage service (the collaboration model is the product)
Not an MCP wrapper around existing storage (agent-native + human-native, not retrofitted)
Not a database (real files and documents, not rows and queries)
Not agent-only (humans are first-class users alongside agents)
Not ephemeral compute (that’s E2B/Daytona — complementary)

Core Differentiators

vs. Fast.io (Primary Competitor)

Fast.io is “workspaces for agentic teams.” We differentiate on:

Human collaboration first-class: Fast.io treats agents as workspace users. We treat agents and humans as co-workers in the same space.
Suite integration: Mailmolt + Findable + Storage = unified platform.
Open-source core: Supabase model — reduces lock-in.
Real-time human dashboard: Humans see agent activity live, not after the fact.

vs. Google Drive / OneDrive

Agent identity: Agents are first-class users, not extensions of human accounts.
Optimized for agent workflows: Webhooks, programmatic access, MCP integration.
Cross-platform: Works with Claude, OpenClaw, Codex, GPT, LangChain, CrewAI — not locked to one ecosystem.

Product Architecture (Revised)

Layer 1: Shared Workspace

workspace://{org-id}/
├── /projects/                    # Shared project folders
│   ├── /campaigns/q1-launch/     # Agents + humans collaborate here
│   │   ├── brief.md              # Human-created
│   │   ├── draft.md              # Agent-created, human-reviewed
│   │   ├── performance.json      # Agent-created, auto-updated
│   │   └── recommendations.md    # Agent-created from analytics
│   └── /reports/monthly/
├── /agents/                      # Per-agent working directories
│   ├── /{agent-id}/workspace/    # Agent's private working area
│   └── /{agent-id}/outputs/      # Published for team review
└── /team/                        # Human-managed files agents can access

Layer 2: Identity & Permissions (Agents + Humans)

Agent ID — Unique, verifiable identity (integrates with Mailmolt identity)
Human accounts — OAuth (GitHub, Google) with dashboard access
Capability tokens — Scoped, time-limited access (JWT)
Permission policies — Path-based, operation-specific, with human override

{
  "agent": "agent-7x9k2",
  "paths": ["/projects/campaigns/*"],
  "operations": ["read", "write"],
  "human_review_required": ["/projects/campaigns/*/final/*"],
  "expires": "2026-03-01T00:00:00Z",
  "granted_by": "user:rakesh@company.com"
}

Layer 3: Agent-Human Collaboration (NEW)

Activity feed: Real-time stream of who (agent or human) did what
Comments/annotations: Humans annotate agent files without disrupting execution
Watch/subscribe: Humans or agents subscribe to file changes
Feedback loop: Human provides structured feedback → agent reads it on next run
Presence: Dashboard shows which agents are actively working on what

Layer 4: Cross-Agent Collaboration

Publish: Agent marks an output as discoverable by authorized agents
Subscribe: Agent subscribes to another agent’s outputs
Shared workspaces: Multi-agent projects with file-level attribution
Webhooks: Notify agents/humans when files change

Layer 5: Document Intelligence (v2+)

Text extraction from PDFs, images, documents
Structured data extraction from spreadsheets
Semantic search across all files (agent + human created)
Format conversion (PDF → Markdown, Excel → JSON)

MCP Server Design (Revised)

{
  "name": "agentvault",
  "version": "2.0.0",
  "tools": [
    { "name": "vault_write", "description": "Write a file to the workspace" },
    { "name": "vault_read", "description": "Read a file from the workspace" },
    { "name": "vault_list", "description": "List files in a directory" },
    { "name": "vault_share", "description": "Share a file or folder with agents/humans" },
    { "name": "vault_search", "description": "Search files by content or metadata" },
    { "name": "vault_versions", "description": "Get version history with agent/human attribution" },
    { "name": "vault_comment", "description": "Add a comment or annotation to a file" },
    { "name": "vault_watch", "description": "Subscribe to file/folder change notifications" },
    { "name": "vault_feedback", "description": "Read human feedback/comments on agent outputs" },
    { "name": "vault_activity", "description": "Get recent activity feed for a workspace" }
  ]
}

New tools (vs. prior version):

vault_comment — Agents can read human annotations; humans can comment on agent files
vault_watch — Real-time notification when files change (for reactive agents)
vault_feedback — Structured human feedback agents can consume programmatically
vault_activity — Activity feed spanning agent + human actions

Pricing Model

Industry consensus: No free persistent storage. Credit card required.

Recommended Tiers

Tier	Price	Storage	Agents	Humans	Card Required	Purpose
Trial	$0 (7 days)	500 MB	1 agent	1 human	No	Evaluation
Starter	$5/month	2 GB	3 agents	2 humans	Yes	Indie devs
Team	$15/month	10 GB	5 agents	5 humans	Yes	Small teams
Pro	$30/month	50 GB	15 agents	10 humans	Yes	Growing teams
Enterprise	Custom	Custom	Custom	Custom	Yes	Large orgs

Key change from prior pricing: Added human seats as a dimension. This reflects the collaboration-first positioning. Not just “how many agents” but “how many collaborators (agents + humans).”

Unit Economics (Cloudflare R2)

Tier	Revenue	R2 COGS	Gross Margin
Starter ($5)	$5.00	~$0.15-0.50	90-97%
Team ($15)	$15.00	~$0.30-1.00	93-98%
Pro ($30)	$30.00	~$1.00-3.00	90-97%

R2 cost: $0.015/GB/mo storage, $0 egress. Operations: $4.50/M writes, $0.36/M reads.

Abuse Prevention

Six-layer defense (unchanged from prior):

Identity verification (OAuth + credit card)
Organization-level limits (not agent-level)
Progressive trust (new agents → limited, proven agents → full)
Behavioral analysis (real agents have diverse file patterns)
Rate limiting (per-agent and per-org)
Economic deterrence ($5/mo minimum)

Success Metrics (Revised)

Year 1

Metric	Target	Notes
Registered workspaces	5,000	Agents + humans
Monthly active collaborators	2,000	Both agents and humans actively using
Paid organizations	200
Human users viewing agent outputs	500	Key new metric — proves collaboration value
MRR	$10K

Year 2

Metric	Target
Registered workspaces	50,000
Monthly active collaborators	15,000
Paid organizations	2,000
Enterprise customers	20
MRR	$100K

North Star Metric (Revised)

Old: “Time to first file write” (agent-only)

New: “Time to first human-agent collaboration” — how fast can a developer set up a workspace where an agent writes a file AND a human views/comments on it?

Target: < 10 minutes from signup to first collaborative file operation.

Technical Stack

Component	Technology	Rationale
Storage backend	Cloudflare R2	$0 egress, S3-compatible, global
Metadata & indexing	PostgreSQL (Neon)	Relational, JSONB, battle-tested
MCP server	TypeScript/Node.js	MCP SDK is TypeScript-first
API gateway	Hono on Cloudflare Workers	Edge-native, low cost
Human dashboard	Next.js + Tailwind	Real-time file browser, activity feed
Real-time sync	Cloudflare Durable Objects or WebSocket	Agent/human presence, live updates
Identity	JWT + capability tokens	OIDC-A aligned
Search	Meilisearch	Fast, lightweight
Versioning	Content-addressable (SHA-256)	Efficient dedup
Caching	Upstash Redis	Rate limiting, session cache
Audit log	Append-only (Postgres + R2 archive)	Immutable, queryable
Notifications	Webhooks + WebSocket	File change events