agience-flare-dkg-integration 0.4.4

Governed authoring layer above DKG v10: commit-gated Agience artifacts, policy-controlled projection, FLARE confidentiality, typed agience: RDF Knowledge Assets.

agience-flare-dkg-integration

The governance layer above dkg mcp setup. OriginTrail's official dkg mcp setup (shipped 7 May 2026) makes MCP-to-DKG a two-command commodity. This package is what sits upstream — committed Agience Core artifacts, optional FLARE confidentiality, and typed agience: RDF Knowledge Assets — so what reaches Working Memory is governed, attributable, and SPARQL-queryable, not raw LLM output.

See the head-to-head comparison and design brief.

Features

• MCP stdio server (agience-dkg-mcp) — exposes agience_wm_write, agience_share (agience_promote alias), agience_search tools for Claude Desktop, Cursor, Claude Code, and any MCP host
• Writes committed Agience artifacts to DKG v10 Working Memory as typed JSON-LD Knowledge Assets with the agience: RDF vocabulary — SPARQL-queryable by type, author, collection, and memory layer
• Promotes eligible assets to Shared Memory (SHARE) via the daemon's POST /api/knowledge-assets/{name}/swm/share (or dkg-create with privacy=public on MCP transport)
• Searches across memory layers via the daemon's POST /api/query (or dkg-sparql-query on MCP transport)
• Default transport: direct HTTP to the local DKG v10 daemon at http://127.0.0.1:9201; bearer token auto-read from ~/.dkg/auth.token. Alternative transport: MCP Streamable HTTP at POST /mcp with SSE stream handling
• Distinguishes transport success from blockchain anchoring state (status: anchored vs status: pending)
• FLARE optional path: when policy_class = "internal-confidential", only derived projections reach DKG; raw content stays encrypted

Install

pip install agience-flare-dkg-integration

MCP Server (Claude Desktop, Cursor, etc.)

Default config uses the local DKG v10 daemon — no token is needed because the daemon transport auto-reads ~/.dkg/auth.token:

{
  "mcpServers": {
    "agience-dkg": {
      "command": "agience-dkg-mcp",
      "env": {
        "DKG_BASE_URL": "http://127.0.0.1:9201"
      }
    }
  }
}

Tools: agience_wm_write, agience_share (agience_promote alias), agience_search.

To target an MCP-fronted DKG node instead, add "DKG_TRANSPORT": "mcp", set "DKG_BASE_URL": "http://localhost:8083", and add "DKG_TOKEN": "<mcp-bearer>".

CLI

# Default: local DKG v10 daemon. Bearer token auto-read from ~/.dkg/auth.token.
# DKG_BASE_URL defaults to http://127.0.0.1:9201 if unset.

# Write to Working Memory
agience-dkg wm-write \
  --title "Architecture Decision: use DKG v10" \
  --artifact-type decision \
  --artifact-id art-001 \
  --content "We will use DKG v10 Working Memory as the shared knowledge substrate." \
  --context-graph-id agience-demo \
  --collection-id my-project

# Share to Shared Memory (SHARE)
# Pass the Knowledge Asset NAME from the wm-write output (e.g. "art-001-Architecture-Decision-..."),
# NOT the v10.0.1 turnUri (its trailing revision index does not contain the KA name).
agience-dkg share <ka-name> --context-graph-id agience-demo
# agience-dkg promote is still accepted as a backward-compatible alias.

# Search
agience-dkg search "architecture decisions" --context-graph-id agience-demo

Resolved invocation block. Every wm-write call prints a # agience-dkg wm-write — resolved invocation (copy to replay): block to stdout before the HTTP call fires, showing all fully-resolved flags with tokens truncated to 8 chars. Useful for testers and reviewers to copy and replay the exact command.

For MCP-fronted DKG nodes, add --transport mcp --base-url http://localhost:8083 --token <mcp-bearer>.

Python API

from agience_dkg_integration import DkgDaemonClient, MemoryTurnRequest
from agience_dkg_integration.formatter import artifact_to_markdown, session_uri_for_collection

# Default: local DKG v10 daemon. Bearer token auto-resolves from ~/.dkg/auth.token.
client = DkgDaemonClient(base_url="http://127.0.0.1:9201")

# Alternative: MCP-fronted DKG node
# from agience_dkg_integration import DkgHttpClient
# client = DkgHttpClient(base_url="http://localhost:8083", bearer_token="<mcp-bearer>")

result = client.memory_turn(MemoryTurnRequest(
    contextGraphId="agience-demo",
    markdown=artifact_to_markdown(title="My Note", content="...", artifact_type="research-note", artifact_id="art-001"),
    layer="wm",
    sessionUri=session_uri_for_collection("my-project"),
))
print(result.turn_uri)

Links

• Full documentation & design brief
• Agience Core
• FLARE Index · Paper

License

MIT