OWD Platform

Make one AI's work the next AI's head start.
Owner-controlled policy autopilot and continuity for Obsidian and independent AI agents.

Deploy Community · See how it works · Policy autopilot · Request managed alpha · Trust model · OWD Sync

OWD gives independent AI agents one durable, cited Project record—and lets routine research and coding continue under standing policy without dashboard babysitting. Connect once, set the boundary, and step back in only when the work needs new authority or an owner Decision.

It turns the Obsidian vault you choose into bounded Project context without becoming another model subscription, taking over the agent runtime, scraping hidden conversations, or silently deciding what is true.

Every capable AI starts strong and then forgets the work trapped in another tool. Copying chats forward loses sources, mixes claims with decisions, and makes access nearly impossible to audit.

flowchart LR
    V["Your Obsidian vault"] <--> S["OWD Sync"]
    S <--> O["Your OWD workspace"]
    O <--> A["Agent A"]
    O <--> B["Agent B"]
    A --> H["Cited Handoff"]
    H --> B
    B --> R["Independent Review"]
    R --> D["Your Decision"]
    D --> O
    O --> N["Next agent's head start"]

1. Pair the vault you choose. OWD Sync connects only the open, explicitly approved Obsidian vault.
2. Publish current context. OWD automatically builds a searchable library from the latest durable sync state.
3. Connect independent agents. Each compatible client receives its own revocable, read-only vault and folder grant through remote MCP.
4. Prepare one Project handoff. Choose the agent, Project name, vault, and folder once during guided onboarding.
5. Say Connect this project to OWD. The working agent creates, joins, rejoins, or resumes the exact Project on the same connection.
6. Carry only the work you approve. Share cited contributions, request an independent Review, and record the owner Decision that should survive.

No prompt is copied back into the agent. The prepared first Project requires no second MCP authorization and no daily packet renewal. A changed boundary or a later unmatched Project still returns to one exact owner consent.

OWD Sync does not authorize agents or create Projects. An MCP grant does not grant local Obsidian CLI, shell, or filesystem authority. The human always remains the vault owner.

The authenticated workspace is a focused folder-style application:

• How OWD works — guided onboarding first; advanced controls afterward
• Vaults — install OWD Sync, pair vaults, and see exact sync state
• Projects — inspect durable activity, consent, participants, and provenance
• Agents — connect, scope, inspect, and revoke MCP clients
• Notes — search, browse, create, and safely edit Markdown text
• Backup & restore — create, verify, export, and restore encrypted recovery points
• System health — compatibility, deployment identity, and redacted diagnostics

OWD derives one next action from authoritative state. A user should not have to understand materialization jobs, internal IDs, raw scopes, or packet rotation to finish setup.

OWD automates the routine coordination around agent work while the execution harness keeps planning, models, tools, terminals, retries, schedules, and worktrees. OWD is the durable policy and continuity layer—not another agent supervisor.

• Owner-authored standing policy. A lead cannot write or approve the policy that judges its own work.
• Deterministic completion gates. Research and coding Runs pass only from bounded accepted evidence, independent Review, current budgets, integrity, and fenced continuity state—never model confidence or hidden reasoning.
• Exceptions instead of babysitting. Routine allowed work continues; authority expansion, destructive actions, protected paths, conflicts, exhausted budgets, and integrity failures return to the owner explicitly.
• Continuity that is measured. Scheduled checkpoints and disposable drills report RPO, RTO, continuity age, recovery quality, and runtime independence.
• Recovery without restored power. A fresh Community installation can restore quarantined evidence, but never grants, leases, credentials, OAuth state, policy authority, or scheduler authority.

Start with policy autopilot, the continuity plan, or the security model.

OWD exposes one authenticated remote MCP server using Streamable HTTP and OAuth 2.1 with S256 PKCE. It implements the current MCP 2026-07-28 protocol and stateless 2025-11-25 compatibility. OWD-specific behavior is expressed through ordinary MCP Tools, Resources, and Prompts—not a custom transport or a client-side protocol fork. See the frozen MCP compatibility matrix.

The portable baseline is designed for compliant remote MCP clients. This is a protocol-compatibility claim, not certification of every client. A named client and version is described as independently validated only when a dated acceptance receipt for that exact combination exists.

• Obsidian Mind receives a versioned compatibility profile, MCP Resource, setup Prompt, and script-free skill while preserving its native note layout and local qmd/om server.
• Eve.dev receives a source-pinned, user-scoped connection module, MCP Resource, setup Prompt, and script-free skill. Its sessions and subagents remain native to Eve while OWD supplies portable Project continuity and independently attributable collaboration identities.
• Albatross receives a source-pinned setup kit, workspace prompt, MCP Resource, setup Prompt, and script-free skill. A pinned temporary bridge connects its stdio-only MCP client to OWD without changing OWD's standard remote endpoint or OAuth authority.
• Hermes Agent uses the universal MCP endpoint and remains an optional client, never an OWD runtime dependency.
• Orca ADE can host compatible clients without giving OWD control of Orca tasks, worktrees, terminals, or permission settings.

Compatibility profiles improve setup; they never widen authority. See Agent access, release compatibility, and the Obsidian Mind, Eve, and Albatross profiles.

The reviewed upstream versions, source commits, critical paths, and package integrities live in compatibility/upstreams.json. A daily, low-noise monitor opens one rolling GitHub issue only when a newer upstream contract needs human review; it never silently advances an OWD compatibility claim.

Obsidian Mind turns an Obsidian vault into an active local knowledge system with graph search, scoped recall, reasoning, and capture. OWD gives selected work a durable, owner-approved Project record that can move between independently authorized agents without exposing the rest of the vault or flattening local memory into shared truth.

• Keep Obsidian Mind's native qmd and om servers. OWD is added beside them as the remote md-evolved MCP connection; it never replaces or proxies the local tools.
• Use Mind's search, expand, recall, and reason for fast local cognition. Use OWD for the current Work Packet, cited Artifacts, Handoffs, Reviews, owner Decisions, provenance, and recovery.
• Detect an Obsidian Mind vault from its versioned manifest, preserve its native note layout, and derive narrow concrete content roots without treating local profile data as authorization.
• Keep memories/, private notes, configured never-expose files, and runtime infrastructure outside ordinary OWD Project context. The server enforces the same ceiling across reads, search, Project discovery, citations, repair, and resume.
• Treat Mind's record_work and remember as direct vault writes. They obey OWD's primary-writer guidance so a second agent does not silently create overlapping edits or sync conflicts.
• Carry Mind's consultation result into the OWD Artifact or Handoff before committing to an approach, including contrary evidence or an explicit “nothing recorded” result.

The same profile ships through the shared client pack, the portable Obsidian Mind skill, the standard MCP Resource owd://compatibility-profiles/obsidian-mind/v1, the connect-obsidian-mind Prompt, and the dashboard's additive setup command. The profile source is source-verified against Obsidian Mind 8.3.1 at upstream commit 538522e4ea660cdc1265f8ef71ef43966e1d9a96; live client acceptance remains explicit rather than assumed. See the complete Obsidian Mind compatibility guide.

Eve runs durable agent sessions, subagents, sandboxes, channels, and schedules. OWD gives that execution an owner-approved Project record that can survive the Eve session and become a cited head start for another independently authorized agent.

• Copy one generated agent/connections/owd.ts module from the OWD dashboard.
• Authenticate through Eve's native user-scoped Vercel Connect flow—no static token, proxy, custom transport, or OWD-specific tool-call protocol.
• Discover the standard OWD tools as owd__<tool> and use the same open_project → wait-if-needed → resume_project lifecycle.
• Give every genuinely independent Eve reviewer a distinct connector identity, preserving real attribution instead of treating a renamed child as another participant.

The profile is source-verified against Eve 0.29.4 and @vercel/connect 0.6.0; live Eve acceptance remains explicit rather than assumed. See the complete Eve.dev compatibility guide.

Albatross is a local coding harness with open-weight model support, operator modes, checkpoints, resets, and parallel path exploration. OWD gives those runs a cited Project record another independently authorized agent can resume.

• Copy one generated setup kit from the OWD dashboard.
• Pre-authorize the exact OWD client before Albatross starts, avoiding its 30-second MCP initialize limit.
• Keep Albatross's native agent.config.json, tools, approvals, /reset, and /path fork; add only one MCP entry and one marked workspace-prompt block.
• Use qualified mcp__owd__<tool> names and resume .owdignore after every fresh task or reset.
• Give a genuinely independent writer or reviewer a different participant ID and OAuth authorization, preserving real attribution.

Albatross 2.0.3 is currently stdio-only, so the profile pins the temporary mcp-remote 0.1.38 bridge. OWD remains ordinary remote Streamable HTTP MCP with OAuth; there is no Albatross-only server or custom tool protocol. The profile is source-verified, while live Albatross acceptance remains explicit rather than assumed. See the complete Albatross compatibility guide.

• Single owner, multiple vaults. Team accounts and shared administration are intentionally outside the current authorization model.
• Read-only agent context. Every MCP vault call rechecks the live client, audience, vault, folder, scope, expiry, and revocation boundary.
• No hidden-conversation ingestion. OWD stores durable contributions and provenance, not provider transcripts or chain-of-thought.
• Owner Decisions remain distinct. Agents can append contributions; they cannot approve their own work or silently promote it into owner truth.
• Restored data is not silently trusted. Restored note paths remain separately identified and excluded until explicitly approved.
• One advisory local writer. The first Project agent becomes the default primary writer for owner-requested bounded local tasks. Later agents are warned to remain read-only. A same-client restart keeps the role after resume_project; a different authorization cannot be promoted from the global Agents screen. Any future responsibility handoff must be explicitly Project-scoped. This guidance is not presented as an operating-system lock.
• Recovery excludes authority. Snapshots omit sessions, OAuth tokens, credentials, live grants, and protocol secrets.

Read the full Security Model before moving beyond synthetic alpha data.

The managed model is not a shared multi-tenant vault database. Every owner cell has separate runtime and storage resources. A public OWD Cloud account, billing system, and service-level offering are not generally available.

See Deployment Modes and current alpha status.

The deployment runs in your Cloudflare account. Review the alpha limits, security model, and deployment boundary first. OWD never receives your Cloudflare credentials or recovery private key.

The alpha is for learning whether OWD makes real cross-agent work materially easier while hardening Community deployment and upgrade behavior.

• Use synthetic or disposable vaults; personal and production-vault adoption remains a later safety gate.
• OWD Sync 0.1.7 is not yet in Obsidian Community Plugins. The OWD desktop installer is the primary alpha path; BRAT is a disclosed fallback.
• Web mutation covers Markdown text and note creation. Rename, delete, attachment writes, and arbitrary .obsidian mutation are out of scope.
• Snapshot attachment and .obsidian sections remain disabled.
• Local synthetic policy-autopilot and disposable continuity-drill acceptance is complete. The current external gate remains one unassisted, two-agent Handoff → Review → owner Decision → recovery run using disposable data.
• A standards-compatible path does not claim that every client has completed the full acceptance exercise.

Already use Obsidian and more than one AI client? Deploy Community or request managed alpha access.

• Current alpha status
• Agent-first quick start
• OWD Sync installation and pairing
• Deployment modes
• Public roadmap

• Product definition
• Architecture
• Security model
• Security reporting
• Backup and recovery

• Agent access and MCP
• Release compatibility
• Obsidian Mind compatibility
• Eve compatibility
• Albatross compatibility
• Development contract
• Public quality gates
• Complete documentation index

OWD Platform is a TypeScript monorepo built around one integrated Cloudflare Worker and a separately released Obsidian companion.

pnpm install
pnpm types
pnpm check
pnpm test
pnpm test:e2e
pnpm build
pnpm deploy:dry-run

pnpm dev builds the authenticated web application and starts Wrangler with local bindings. These commands are for Community operators and contributors; managed alpha users receive a pre-provisioned workspace instead.

OWD Platform is licensed under the Apache License 2.0. Vendored upstream components retain their original notices and licenses.