Comparison
Both products independently landed on the same elegant rail — one MCP endpoint, two tools: search capabilities, execute capability. Great minds, same plumbing. The difference is what flows through it. OpenWork shares capabilities: skills, plugins, connections, published to a marketplace. vahue shares memory: what the work taught the room, owned by the channel. One ships your team better tools. The other stops your team from re-learning everything. Guess which one compounds.
OpenWork is an open-source control surface for agentic work — an Electron app around the OpenCode engine, with an org tier for teams. vahue is a shared workspace where the room itself accumulates knowledge. Different bets, both respectable. Here's the tape.
| Axis | vahue | OpenWork |
|---|---|---|
| What the team shares | Memory — distilled knowledge, scoped and owned by the channel | Capabilities — skills, plugins, MCP connections, providers, via org marketplaces |
| The agent | Any ACP-speaking agent; the workspace is agent-agnostic by design | One embedded engine (OpenCode); you bring your own LLM key or client, not your own agent runtime |
| Protocols | ACP (control) + MCP (capability, memory) + HTTP/WebSocket (record) | MCP only — OpenWork itself is an MCP server you add to the agent you already have |
| Memory model | Channel-scoped shared store; the agent distills at end of turn; provenance and trust are mandatory; retrieval is pushed before every prompt | Per-user memory bank; every write is human-verified; recall is explicit search only — deliberately no auto-recall |
| Collaboration surface | Channels with a shared record: messages, run steps, artifacts, attribution | No rooms or shared chat; collaboration is grant-based access to capabilities and shared hosted workers |
| Rail scoping | Structural — one endpoint per channel, the channel is in the URL | Policy-based — org membership and grants, checked server-side |
| Enterprise surface | Hosted SaaS; SSO/SAML and always-on cloud agents in the Enterprise tier; workspaces isolated by PostgreSQL row-level security | SSO/SCIM, RBAC, Helm charts for three clouds, air-gapped installs, managed inference gateway |
| Maturity | Hosted SaaS, free for teams up to 5 · open-source self-host | Active v0.x product with a large user base and an enterprise tier |
A credible comparison admits what the other side does better. We're not monsters.
The desktop app is optional: add OpenWork's MCP to Claude Code, Codex, or Cursor and your skills and connections follow you across tools. Adoption starts with a pasted prompt — a genuinely lovely onboarding trick.
Org marketplaces for skills, plugins, and connections — with grants, kill switches, and admin curation — solve "how does the team standardize its agent setup" today. Not someday. Today.
SSO/SCIM, desktop policies, air-gapped deployment, managed inference. Unusually deep for a v0.x open-source project — credit where it's due.
Same rail, different cargo. We think the cargo is the point.
OpenWork's memory is per-user and manually saved — knowledge stays with the person, and walks out the door with them. In vahue the channel owns what the work produced, so it outlives any person's session, machine, or employment. Your team's knowledge shouldn't have a notice period.
Channels carry messages, run steps, and artifacts with attribution. You can see what ran, what it cost, and what came back — the visibility layer capability-sharing simply doesn't have. Who did what is a fact, not a Slack thread you'll never find.
vahue pushes the channel's knowledge into every prompt, trust-ordered, before the agent starts. OpenWork deliberately never auto-recalls — the agent only knows what it remembers to search for. Agents, like people, rarely remember to search.