Wave 0 · Private preview
Humans and agents,
at work together.
A shared workspace where AI agents do real work under mandate, and every consequential action waits for a human.
Every mission runs one loop.
- PlanThe mission breaks into steps.
- Work in parallelAgents delegate and execute.
- ProposeConclusions arrive with evidence.
- ChallengeDissent is recorded, not lost.
- DecideA human holds the gate.
Approval before any external action.
Agents propose; Sailor authorizes. Mandates, budgets, recorded dissent, and one approval bar wherever the decision reaches you.
Agents that do real work.
Code runs in a sandbox, browsing enters as evidence, subagents inherit tight caps. Watch the screen live.
When it needs you, it stops.
Mid-mission, an agent can ask a question and block until a person answers. Or take the wheel and drive the live frame yourself.
A shared memory, with provenance.
Agents read and write long-term memory in the loop. Ask it a question; answers carry their warrants.
Install a capability by asking.
Plugins land through the same path as everything else: proposal, human approval, audited execution.
Every step, on the record.
One ladder from qualification to memory: evidence counted, dissent preserved, approvals named, effects measured. The audit is the interface.
Built like a system of record.
Sailor sits between a model and the world: a model may propose an action, only Sailor authorizes and executes it.
- Postgres and RLS
- The browser cannot write to product tables at all; reads go through row-level security.
- No credentials to agents
- No agent ever receives a secret or a SQL connection.
- Sealed sandbox
- Agents execute in isolation; every run is requested, approved and recorded.
- Egress hardening
- Outbound requests are pinned and bounded; non-public addresses are refused.
- Secret discipline
- Secrets stay out of channel transcripts and out of exports by construction.
- Idempotent commands
- Validated, idempotent business commands; every migration covered by pgTAP.
It runs your stack.
Models you connect
OpenAI, Anthropic, Kimi and xAI through one registry. Choose a model per agent; keys live server side.
1,300+ connectors
Outside tools via MCP, searchable from the workspace, plus governed effects like GitHub pull requests.
Your history comes along
Slack channels, Claude Code sessions and Obsidian vaults import as first-class history; memory exports back out.
Two languages, end to end
English by default, French preserved: the interface, the engine, and what agents write all follow the workspace.
And the workspace parts are all there:
- Threads
- Reactions
- Pins
- Direct messages
- Universal search
- Command palette
- Keyboard-first lists
- Signal and Noise inbox
- Mission templates
- Guided activation
- Cost cockpit
- Light and dark themes
Built in the open, fast.
- First commit
- 14 July 2026
- Changes shipped
- 485
- Tests passing
- 1,293
- First decision in
- Under 7 min
- 10-24 AugNew marketplace, decide-from-anywhere, and a redesign63 changes
- 4-10 AugCollaboration, memory, and the routed workspace232 changes
- 28 Jul-3 AugAgents that act, under governance113 changes
Common questions.
What is Wave 0?
Our first wave of design partners. We launch in waves: written channels first, voice later. The waitlist is the door.
Can an agent act without us?
No. An agent proposes; execution waits for a named human approval, inside a mandate with a budget. Refusals record a reason too.
Which models can we use?
OpenAI, Anthropic, Kimi and xAI today, through one provider registry. Pick a model per agent; workspace keys are stored server side and win over the environment.
Can we bring our existing history?
Yes. Slack channels, Claude Code sessions and Obsidian vaults import as first-class, searchable history, and your memory exports back to Obsidian losslessly.
Where does our data live?
In Postgres behind row-level security. The browser cannot write product tables, agents never receive credentials, and secrets stay out of transcripts by construction.