Skip to main content

Changelog

Status: Reference Date: 2026-04-01

All notable changes to Agience are documented here.

Format follows Keep a Changelog.

[1.0.0-mvp] — 2026-04-01​

Added​

Core platform

  • Artifact-based knowledge model with durable IDs, content, context metadata, and provenance fields
  • Workspaces (ArangoDB) as ephemeral draft staging areas with card-based UI for humans and artifact-based API surface for agents
  • Collections (ArangoDB) as versioned, immutable committed knowledge stores
  • Explicit commit lifecycle — workspace draft to collection canonical — with commit preview, diff, and review
  • ArangoDB as the sole database for both workspaces and collections

Search

  • Hybrid BM25 (lexical) + kNN (semantic) search with Reciprocal Rank Fusion (k=60)
  • Embeddings via the configured embeddings provider; encrypted IVF cells in S3
  • Aperture control — statistical elbow method for semantic neighborhood filtering
  • Field boost presets (description-first, balanced, content-heavy)
  • Unified search across workspaces and collections via POST /artifacts/search

AI-assisted workflows

  • extract_units agent — LLM-based extraction of candidate artifacts from raw content
  • Operator artifacts (vnd.agience.transform+json) — reusable multi-step workflow definitions
  • Agent dispatch via unified POST /artifacts/{id}/invoke endpoint
  • WorkspaceEventHandler — event-driven automations triggered by artifact lifecycle events (artifact_created, artifact_updated, artifact_deleted)
  • Chat artifacts (vnd.agience.chat+json) — agentic multi-turn sessions with full tool-call history

Ingestion

  • Browser upload for documents and media with direct-to-S3 presigned PUT flow
  • PDF text extraction via Astra document_text_extract tool
  • File URL ingestion via Astra ingest_file tool

MCP integration

  • Agience as MCP server — 12 purpose-built tools exposed at POST /mcp (Streamable HTTP)
  • Agience as MCP client — workspace-scoped proxy to external MCP servers registered as vnd.agience.mcp-server+json artifacts
  • VS Code and Claude Desktop compatibility via MCP client connection artifacts
  • Desktop host relay — local MCP server bridge via WebSocket for development workflows
  • Official-first integration rule — vendor MCP servers registered as artifacts, not reimplemented

First-party MCP servers

  • Aria (port 8083) — response formatting, chat turn execution, presentation artifacts
  • Astra (port 8087) — file ingestion, text extraction
  • Mantle (port 8081) — provenance tracing, conflict detection, contract enforcement
  • Iris (port 8086) — email/message delivery, MCP server management, sandboxed shell execution
  • Sage (port 8084) — hybrid search, artifact lookup, Azure AI Search projection
  • Unified host mount (_host, port 8082) — all persona servers on a single port for deployment simplicity

Content types (Agience-owned)

  • application/vnd.agience.chat+json — chat conversation artifacts
  • application/vnd.agience.view+json — configurable live workspace views
  • application/vnd.agience.stream+json — RTMP stream source configuration
  • application/vnd.agience.mcp-server+json — external MCP server registration
  • application/vnd.agience.mcp-client+json — MCP client connection configuration
  • application/vnd.agience.host+json — Agience host configuration

Platform MIME renderers (built into the frontend)

  • text/markdown, text/plain, application/json, application/pdf
  • image/*, audio/*, video/*

Authentication and access control

  • Multi-provider OAuth2: Google, Microsoft, Auth0, custom OIDC, password
  • RS256 JWT tokens with JWKS endpoint (/.well-known/jwks.json)
  • Scoped API keys — resource|tool|prompt : mime : action format
  • Server credential client_credentials grant for MCP server identity
  • Collection grants (CRUDEASIO 9-flag model) for shared knowledge access
  • Allowed-email / allowed-domain / allowed-ID access control

Deployment

  • Docker Compose stack: all services, infra (ArangoDB, Postgres, MinIO), and MCP persona servers
  • Self-hosting path with documented environment variable configuration
  • Hosted preview environment at agience.ai
  • Frontend runtime config injection (public/config.js) for zero-rebuild environment switching

Architecture​

  • Three-layer model — Core (kernel), Handlers (MCP servers and drivers), Presentation (frontend shell) with enforced separation
  • Type blindness in Core — no MIME constants, no artifact.context parsing in Core services or Presentation components
  • Registry-driven viewer dispatch — all type-specific viewer wiring flows through src/facet/src/registry/; Presentation never imports handler code directly
  • MCP Apps pattern — content type viewers served as ui:// resources from MCP servers and rendered in the frontend's iframe sandbox (McpAppHost.tsx)
  • Unified entity design — all artifacts use a single Artifact entity with one to_dict() method. Collection is an alias for Artifact. Container artifacts are distinguished by content_type.
  • Fractional indexing — workspace artifact ordering uses lexicographic base-62 keys to avoid renumbering on reorder
  • Artifact reference model — cross-artifact references are always a single artifact_id string (UUID); no embedded content, no workspace-scoped references