What's new in the workspace: connectors, wiki, skills, DLP, quotas, MCP, synthesis

One corpus, many new doors

Since v0.22 the platform grew one full-depth capability per area of the operator's work - connecting channels, organizing knowledge, instructing assistants, shielding sensitive documents, capping usage and spending, integrating external tools, and combining sources honestly. This post tours the six areas at reading depth; each links its guide page (the complete operator reference) and its technical post (the design deep dive) here on the blog.

The facts ride the platform's claims ledger: every number in this post cites its census row; anything lacking a traceable source stays out of the page.

Connectors — the channels where your customers already are

External channels (WhatsApp, Telegram, Discord, Slack) connect to the same knowledge base the internal chat uses. The connector's adapter contract = one platform, one secret token, one workspace - bound at creation and immutable after (the trust model); the archive binding determines the grounded corpus per channel. The v0.28.1 release added the edit arm: name, archive binding, welcome/fallback messages and both limits are now mutable via a single endpoint call, with the mutation audited.

• The guide: The operator guide · The technical post(/en/blog/connectors-adapter-contract)

The wiki knowledge engine — knowledge on written rails

Each archive gains a three-layer instruction model: hard rules enforced by code (the filesystem guard + an always-on prompt block), an advisory per-archive schema prompt carried as data, and the structured fields. The precedence chain reads: code > advisory > structured - and the two built-ins (wiki_query/wiki_write) run without a model call, so the guarantees don't depend on the model's judgment.

• The guide: The operator guide · The technical post(/en/blog/wiki-schema-layer)

Custom skills — instructions you can call

A custom skill is a reusable prompt template invoked as a slash command. The compile pipeline is the safety story: whitelisted parameters against a schema, one replacement pass, marker-shaped text rejected at creation, masking before compile, and the compiled output arriving as one spotlighted untrusted data entry - a custom skill can add context but never tools or permissions. The freemium fact: 3 custom skills in Community, unlimited in Enterprise.

• The guide: The operator guide · The technical post(/en/blog/custom-skill-authoring)

Document DLP — masks that hold

An optional Enterprise feature per workspace: the async scan runs a 3-tier detection (admin regex patterns; checksum-validated fixed-format identifiers; contextual LLM detection for the residual classes), rewrites every chunk into document-wide numbered placeholders, re-embeds the masked text (no unredacted vector remains searchable), and keeps the only clean copy AES-256-GCM encrypted. The clean copy is reachable only via two explicit gates (chat re-composition with the dlp:unmask permission; the preview unmask ladder) - and widget visitors never receive re-composed text.

• The guide: The operator guide · The technical post(/en/blog/dlp-document-pipeline)

Quotas and costs — ceilings before the spend

Three kinds of ceilings (tokens, storage, cost) resolve through one chain: the per-user override, the install preset, then fail-open unlimited (the opt-in shape - "not configured" never silently becomes "blocked"). The cost arm is the two-phase hold-and-settle reservation: a row-locked reserve BEFORE the LLM call (concurrent requests serialize rather than race a stale check), the settle at the usage seam refunding estimate minus actual. Model pricing is per-row snapshot-frozen; an unpriced model never bypasses the cost quota (the fallback rates apply, warn-once).

• The guide: The operator guide · The technical post(/en/blog/usage-cost-quota-engine)

Bidirectional MCP — both directions

The connect direction: external MCP servers install from the curated marketplace and their tools become agent skills (collision-free names, per-chat pinning that only ever adds tools). The expose direction: the built-in MCP server serves the workspace's RAG search (rag_query) to external IDE clients via the stateless Streamable HTTP transport - with the loopback-only fallback when no API key is configured. The marketplace's write arms are admin-gated.

• The guide: The operator guide · The technical post(/en/blog/bidirectional-mcp-integration)

The synthesis pipeline — where documents disagree, it says so

The multi-document pipeline combines an archive's sources into combined wiki pages - contradiction detection via a similarity pre-filter and a per-pair LLM judge, three explicit budgets (tokens, calls, pages), and selective approval: the deterministic preview precedes every write, and only approved modifications persist. The model proposes; the human ships.

• The guide: The operator guide · The technical post(/en/blog/synthesis-pipeline-design)

License activation — paste, don't touch the server

The Enterprise edition activates in the admin panel by pasting the instance key (probe-first verification; a rejected key never persists; AES-256-GCM ciphertext at rest; runtime application without a restart). The flag set also carries the per-feature gates named throughout this post (connectors_enabled, max_skills, ...).

• F-LIC-1 in FACT-CENSUS.md

Reading on

Facts in this post (all rows in FACT-CENSUS.md): F-VERSION-1 (v0.28.1, released 2026-10-02), F-CONN-1/2, F-WIKI-1, F-SKILL-1, F-DLP-1, F-QUOTA-1, F-MCP-1, F-SYNTH-1, F-TIERS-1, F-LIC-1. Provenance: the platform's docs at the pinned state v0.28.1 (SHA aca5b8ff) - the source of each row. The guide's index: /en/docs · The docs: The operator guide → /it/docs · The operator guide

• Request a demo
• The plain-language version
• /it/docs · The operator guide - the same reading, in-language.