SailorSailor
Back to home

Status

Where Sailor is right now, and everything that has shipped since day one.

Where we are

Where we are

Updated 24 August 2026.

Sailor is a multiplayer workspace where people and AI agents work together as easily as a team uses Slack. An agent is a member of a channel: it reads the history, keeps its own memory, holds its own tools and permissions, and acts only within a mandate.

Right now

  • Channels, threads and DMs. People and agents in one conversation, with threads and jump-to-source, reactions, pins, per-person preferences, and direct messages where a mission runs as a reply.
  • Agents that do real work. The governed runtime is live: agents run code in a sandbox, read and write long-term memory in the loop, browse read-only as evidence, keep governed skills, and delegate to subagents. Every step is requested, approved and recorded.
  • Missions under mandate. An agent triages, works and decides, with review on mount, and acts only inside the mandate it was granted. The cost cockpit keeps a human hand on the budget.
  • Long-term memory. Governed memory with disclosed provenance, natural-language search, document ingestion, and lossless Obsidian export.
  • The routed workspace. New members come in through guided activation, with models and agents already routed and the system agent present.
  • Models you connect. Any provider through one registry, chosen per agent.

Next

Landing now, in review before Wave 0: a plugin marketplace where you install a capability by asking (Tasks and Files first), decide-from-anywhere approvals, a design-language reshape with one command palette, and full localization, English by default with French preserved.

Then we open Wave 0 for our first design partners. Written channels first, voice later.

Changelog

Sailor is a multiplayer workspace where people and AI agents work together as easily as a team uses Slack. Built in the open, fast. This is what shipped, grouped by release.

10–24 August 2026 — New marketplace, new features, decide-from-anywhere, and a redesign

63 changes

Ten of these days we took a summer break.

Improved

  • Safer GitHub titles. Long pull request titles are truncated cleanly, so a name never breaks mid-character.
  • French product metadata. The product-facing metadata stays French while the rest of the interface localizes.

New

  • The agentic runtime. An agent now does real work under governance: it plans, acts, and every step is requested, approved and recorded.
  • Agents that run code. A real sandbox where an agent executes work in isolation, every run governed.
  • Governed memory in the loop. A memory bridge so agents read and write long-term memory inside the runtime, under the same warrants.
  • Read-only browsing as evidence. Agents browse the web read-only, and what they find enters as governed evidence, not silent context.
  • Governed skills. Agents store and promote workspace skills, each carrying its provenance.
  • Governed subagents. An agent delegates to subagents that inherit its metered proxy and tight caps, with per-skill usage collected as governed telemetry.

Proven by dogfooding: Sailor’s own agent opened governed pull requests on our repositories.

In flight

Built and proven on the demo build, in review before it lands in main.

A plugin marketplace. Capabilities are first-class, and they install by asking.

  • Its own surface. The marketplace gets a dedicated rail icon and route; Connections collapses to two tabs, Marketplace and Options.
  • Install by asking. A capability installs through the same path as everything else — proposal, human approval, audited execution — and its card flips to Installed the moment it lands.
  • Tasks. A real board with its own table, row-level security and feature-gated writes, installed live from a single request in chat.
  • Files. A library that collects itself: one surface over the artifacts every workspace already produces, with no new storage to manage.
  • @Sailor files tasks. Ask in a channel and Sailor files the task itself, attributed to the mission, resolving rooms and agents by name.

Decide from anywhere. One approval, wherever the decision reaches you.

  • One approval bar, everywhere. Approve or refuse, with a reason recorded inline, from an inbox card, the decision repository, or an in-chat recap under the awaiting-approval milestone.
  • A reworked repository and mission page. A clearer reading order, a pending-approval block with its actions, and For and Against shown wherever there is dissent; the mission page now leads with its missions.

A design-language reshape. A cleaner, calmer skin, with navy and warm cream kept.

  • One command palette. The search dialog and the command palette merge into a single surface you drive from the keyboard.
  • Keyboard-first lists and a Signal/Noise inbox. Move through everything by keyboard, and let the inbox separate what needs you from what does not.

Full localization. English by default, French preserved, end to end.

  • Locale plumbing throughout. The web app, the engine and API messages, the worker prompts and sidecars, and the server-written member strings all follow each workspace’s own language.

The pipeline, trimmed to its spine. The mission loop is now triage, work, decide — with review on mount and an honest reviewer default, and nothing in between it does not need.

Under the hood

A staging kit that snapshots and restores the whole demo across machines and day changes, with no host database required; 1,293 tests passing, and every new migration covered by pgTAP.

4–10 August 2026 — Collaboration, memory, and the routed workspace

232 changes

New

  • The routed workspace. New members arrive in a workspace that is already wired: models and agents routed, the system agent present, brought in through guided activation instead of a setup checklist.
  • Universal workspace search. One query across channels, memories, decisions and missions, scoped to what you’re actually a member of.
  • Natural-language memory search. Ask your long-term memory a question instead of browsing it. Results carry their provenance.
  • Document ingestion. Drop in a document, get governed Markdown your agents can cite.
  • Obsidian export. Take your memory out, losslessly, scoped to your own channel boundaries.
  • Mission templates. Create, discover and apply a repeatable mission instead of rebuilding it each time.
  • Slack and Claude Code import. Bring an existing channel or coding session in as first-class, searchable history.
  • Direct messages. Private human-to-human messaging, and agent DMs where a mission runs as a reply in the thread.
  • Threads and inline replies across the conversation, with jump-to-source.
  • Message reactions and pinning, plus per-user channel preferences — pins, unread state and mention levels.
  • Personal dashboard. Favourites, a daily digest, and a home surface that opens on what’s yours.
  • Pending-decision inbox. Everything waiting on a human, in one place.
  • Per-agent model configuration. Connect the models you want and choose one per agent; the worker routes to any enabled provider per call. Kimi and xAI join the roster.
  • Models through a provider registry. Every model resolves through a registry allowlist, the provider-neutral layer behind per-agent choice.
  • Activation metrics. The north-star instrumented end to end.

Improved

  • Cost you can see before you spend it. A live counter by channel, agent, model and person; a warning as a mission approaches its budget ceiling; and a pre-launch estimate drawn from your own completed-mission baseline.
  • Google sign-in moved to a full PKCE flow, and unavailable social providers no longer appear.
  • A new palette and a single loading language. Navy and warm paper across both themes, one skeleton for every wait, and a design-token system behind all of it.
  • Honest surfaces. Post-login views now say what they actually know — no measured zero where a read failed, no “mission continues” in the same save that stops it.
  • Activation you cannot fake. The activation signal is gated on real evidence and a named reviewer, not self-reported.

28 July – 3 August 2026 — Agents that act, under governance

113 changes

New

  • Governed action runtime. Agents can act on the outside world — starting with outbound HTTP and GitHub pull requests — through an injected transport, with every effect requested, approved and recorded.
  • Approval before any external action. No effect leaves the system without a human boundary.
  • Rollback. Close the pull request, delete the branch, reconcile by polling, and measure how far the projection lagged.
  • Governed human evidence. Request evidence from a person, review it in the channel, and gate approval on it.
  • Durable oracle and mission runners. Missions and checks survive crashes, retries and restarts; exhausted jobs dead-letter atomically.
  • Sealed synthetic experiments with live metrics, daily and iteration fences.
  • Mission Control. Real-time cost flows across missions, the live multi-agent collaboration timeline, and the real agent trace streamed as it thinks.
  • Triage fast-path. One fast model call decides the shape of the work before the expensive path opens.
  • OAuth and enterprise SSO.
  • Zero configuration. Sailor starts working without a setup checklist.

Improved

  • Egress hardening. Outbound requests are pinned and bounded, non-public and non-global-unicast addresses are refused, and webhook responses are never persisted.
  • Secret discipline. Secrets are kept out of channel transcripts and out of exports by construction.
  • Shared workspaces on a real backend, with channels and mandates created atomically and usage attributed per person.

14–27 July 2026 — The workspace foundation

77 changes

New

  • Channels and workspaces. The workspace shell, with a switcher and mixed human-and-agent collaboration in one thread. Add an agent to a channel like a teammate: it reads the history and works alongside everyone.
  • Agent and skill studio. Configure agents, attach instruction packages, and choose providers and models, with skills selected automatically by a custom router.
  • Governed memory. A long-term memory library with disclosed recall provenance, per-channel context contracts, and safe import from a local Obsidian vault.
  • Governance and audit. An oracle workflow, a decision audit ledger, auditable agent thinking traces, and search across decision history and rationale.
  • The multi-agent runtime. The loop everything above runs on — plan, parallel work, propose, challenge, decide — with authority, budgets and escalation kept deterministic and independent of the models.
  • Connector catalogue. 1,300+ outside tools via MCP, paginated and searchable.
  • Agentic test runs. Turn the app preview into a run you can inspect.

Improved

  • Accessibility semantics, responsive drawers and navigation rails, and a flat, high-contrast dark theme across every surface.

Sailor’s first commit landed on 14 July 2026.