knowledge-dlc MCP server
Commit Nov 9, 2026Converts documents, wikis, and cloud drives into governed, provenance-bearing knowledge bases for AI agents with citation-exact retrieval.
Also known as aithinkers/knowledge-dlc
- Stars
- 0
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Nov 9, 2026
- License
- MIT
What is knowledge-dlc MCP?
knowledge-dlc converts documents, wikis, and cloud drives into governed, provenance-bearing knowledge bases for AI agents with citation-exact retrieval.
Summary generated from the project's README
About knowledge-dlc MCP server
Transforming scattered documents into structured, verifiable knowledge for AI agents is the primary function of knowledge-dlc. It converts diverse file formats including PDFs, Word documents, spreadsheets, emails, and diagrams into a governed knowledge base using the Open Knowledge Format. This system is designed for developers and teams who need to ensure their AI agents provide answers backed by exact citations rather than model hallucinations.
Users can connect the server to remote sources like Confluence, SharePoint, OneDrive, and Google Drive using read-only connectors. The setup process is managed through a CLI that supports integration with Claude Code, Codex CLI, Kiro, and any standard MCP client. Once configured, the system follows a strict lifecycle of ingestion, claim generation, and human-in-the-loop review to ensure every published concept is accurate and traceable to its original source hash and locator.
What you can do with it
- 1Ingest diverse document formatsA developer can feed a variety of files like Excel, PowerPoint, and Visio into the system to create a unified knowledge base.
- 2Connect remote cloud storageTeams can sync their existing documentation from Google Drive or SharePoint to keep their AI knowledge up to date.
- 3Verify AI answer provenanceUsers can query the knowledge base to receive answers that include specific citations for every claim made.
- 4Enforce human review gatesOrganizations can prevent unverified information from being used by requiring manual approval of concept proposals.
Written from the project's README. Every use case is checked against a quote from it.
Automated checks
Ran Nov 16, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachableaithinkers/knowledge-dlc is public and reachablegithub.com/aithinkers/knowledge-dlc ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-11-09github.com/aithinkers/knowledge-dlc ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubK-DLC — turn your documents into agent-consumable knowledge
K-DLC (Knowledge Development Lifecycle) turns the documents your team already has — wikis, PDFs, Word docs, emails, spreadsheets, decks, diagrams, web pages, and pages living in Confluence, SharePoint, OneDrive, or Google Drive — into durable, provenance-bearing knowledge that AI agents can actually trust: curated, linked Markdown concepts in the Open Knowledge Format (OKF) (schemas vendored at core/schemas/okf-0.2/), produced through a governed lifecycle with explicit human approval gates. Think of it as a governed knowledge base builder for the agent era: every answer cites the exact page, paragraph, or shape it came from, and staleness, conflicts, and access rules are first-class instead of afterthoughts.
# get started (Node 22+, from a checkout)
npm ci && npm link
kdlc setup claude-code ~/my-project # or codex | kiro | kiro-ide | mcp
# then, inside your AI tool: "init a knowledge base and ingest ./docs"
Once the package is published to npm, the same works with zero install:
npx knowledge-dlc init — the CLI ships under both the kdlc and
knowledge-dlc bin names.
One harness-neutral core, rendered natively for Claude Code, Codex CLI, Kiro CLI, Kiro IDE, and any MCP client — the same governed engine, the same versioned envelope, everywhere.
Sources -> Evidence -> Claims -> Concepts -> Published knowledge
\ | |
\------ provenance + review receipts
[!NOTE] This repository is in pre-release MVP development against specification version 0.2.0. It is not a supported release. Conformance is declared module-by-module; no capability is considered implemented until its linked issue, tests, and independent review evidence are complete.
[!IMPORTANT] Generative AI can make mistakes. K-DLC never publishes stable knowledge on model confidence alone — review the evidence, claims, and diffs its gates put in front of you.
Why K-DLC
Retrieval over raw documents works until the corpus gets real. Then agents cite stale pages, contradictions surface mid-answer, and nobody can say where a "fact" came from. K-DLC puts structure around knowledge the way CI puts structure around code: every claim traces to a source hash and locator, every concept records who generated and who verified it, conflicts are retained instead of silently resolved, and publication is a reviewed, atomic transaction. Search indexes, embeddings, and graphs stay rebuildable projections — plain files remain the durable contract.
Key features
- Evidence-first ingestion — bounded, sandboxed normalization of thirteen formats — Markdown, text, CSV, PDF, Word (.docx), Excel (.xlsx), PowerPoint (.pptx), Visio (.vsdx), draw.io, GIF, email (.eml and Outlook .msg, attachments inventoried by hash), and HTML — into anchored, quality-labeled evidence units (spec §12)
- Remote sources with real provenance — deterministic read-only connectors
for Confluence, SharePoint, OneDrive (Microsoft Graph), and Google Drive
(native Docs/Sheets/Slides exported to formats the pipeline normalizes);
every fetch records the provider's revision identity and a content hash, so
staleness checks can tell you when the original changed. A guided
connector-setupagent walks you through credentials — read-only scopes only, secrets live in environment variables and are refused anywhere else - Claim-to-source provenance — every assertion carries its source ID, version hash, and locator; inferred claims are never dressed up as explicit statements
- Governed lifecycle — fifteen core stages across Define → Acquire → Understand → Integrate → Govern, with review packets, receipts bound to exact review hashes, and transactional publication
- Nine-role agent roster — conductor, curator, source-analyst, integrator, librarian, trust-reviewer, retrieval-agent, maintainer, governance-reviewer — with runtime-enforced read/write capabilities, not prompt-text promises
- Federation — mount multiple knowledge bases with explicit modes, locked versions, and routing; retrieval rank never grants write authority
- Access-intersection answers — a query result is allowed only when the requester may see the concept and all evidence it discloses
- Deterministic governance — sensors, audit trails, traceability, and a release evaluation that runs with zero live model calls
Step 1 — add K-DLC to your AI tool
K-DLC is designed to be used inside your AI tool: the agents drive the lifecycle (ingest → claims → concepts → review → publish) and stop at the review gates for your decisions. One setup command installs any harness surface into your project, with runner paths resolved against this installation — the installed files work from any directory:
kdlc setup <claude-code|codex|kiro|kiro-ide|mcp>[,...] <project-directory>
| Harness | Setup | Invoke |
|---|---|---|
| Claude Code | kdlc setup claude-code . prints the marketplace add + claude plugin install kdlc@kdlc commands |
/kdlc:<operation>, kdlc:<role> agents |
| Codex CLI (≥ 0.145) | kdlc setup codex <project> writes .codex/ (skill + agents) into your project; surface source in distribution/codex/ |
$kdlc skill, .codex/agents/<role> |
| Kiro CLI (≥ 2.6) | kdlc setup kiro <project> writes .kiro/ (skills + agents) into your project; surface source in distribution/kiro/.kiro/ |
/kdlc-<operation>, .kiro/agents/<role> |
| Kiro IDE | kdlc setup kiro-ide <project>; surface source in distribution/kiro-ide/.kiro/ |
/kdlc-<operation>, .kiro/agents/<role> |
| Any MCP client | kdlc setup mcp <project> writes a stdio config; HTTP packaging in distribution/mcp/ |
kb_search, kb_fetch, proposal_create, … |
Adapters differ only in packaging: stage requirements, security policy, state transitions, and artifact contracts are identical everywhere (spec §25), and every distribution tree is generated from the authored core — CI fails on drift.
Step 2 — talk to it
Inside the harness, work conversationally; the agents run the engine and surface decisions:
you> /kdlc:init, then ingest everything under ./docs and our Confluence ENG space
kdlc> 12 sources normalized (494 evidence units). The curator proposes 9 for
adoption and asks: is the vendor pricing page in scope?
you> no — defer it. Draft concepts for the rest.
kdlc> 7 concept proposals ready for review — packet hash a41c… Approve?
you> approved
kdlc> Published 7 concepts. Try: /kdlc:query ["what is our failover timeout?"]
Remote sources (Confluence, SharePoint, OneDrive, Google Drive) connect
through the guided connector-setup agent — read-only scopes, secrets stay
in environment variables (guide).
The five verbs
Day to day you only need these — everything else is automation or inspection:
kdlc # start or resume: where you are + the next step
kdlc ingest <files> # feed it (folders, PDFs, Office, email, HTML, …)
kdlc query "…" # ask it — answers carry citations
kdlc publish # your gate: list what awaits you; <id> --approve "reason" lands it
kdlc revisit # ratify auto-published drafts into default answers
Automation & CI — the engine directly
The same kdlc CLI the harnesses invoke is a stable surface for scripts and
pipelines — no agent in the loop:
kdlc ingest docs/spec.md # queues a background job; evidence, not answers
kdlc lint # structural/policy findings for CI
kdlc refresh # re-check published knowledge against sources
kdlc publish # governed publication in a release pipeline
Heads up if you drive it by hand: query answers only from published
knowledge. Evidence produced by ingest becomes queryable after
proposal → review → publish — the full sequence is the walkthrough in
docs/getting-started.md, executed end-to-end by a
test on every change. In a harness the agents run that middle stretch for you.
Inspecting your project
Read-only commands you can run anytime, even mid-conversation:
kdlc status # where everything stands
kdlc jobs # background work and results
kdlc sources # remote-source receipts and connector readiness
kdlc trace kb://<kb>/<concept> # full provenance chain
kdlc doctor # diagnosis and safe repair
Repository layout
Schemas and profiles live under core/; stage, agent, sensor, tool,
normalizer, and server sources live under packages/; generated harness
output lives under distribution/<harness>/. The complete mapping from the
specification's reference layout is recorded in
ADR-002.
Development status and governance
- Specification baseline (full spec)
- MVP milestone and issue backlog
- Traceability index
- Agent development contract — issue → plan → implementation → tests → independent review → release evidence
- Release-readiness gates
- Machine-readable pre-release conformance (status: not-ready)
- Recorded pre-release evaluation — zero live model calls
- Contributing · Security policy · Support · Code of Conduct
Local governance checks
node scripts/verify-governance.mjs
node --test tests/governance/*.test.mjs
npm run check:supply-chain
License
knowledge-dlc MCP questions
What sources can K-DLC ingest?
It supports Markdown, text, CSV, PDF, Word, Excel, PowerPoint, Visio, draw.io, GIF, email (.eml and .msg), and HTML, as well as Confluence, SharePoint, OneDrive, and Google Drive.
How does K-DLC ensure AI reliability?
It uses a governed lifecycle with human approval gates and ensures every answer cites the exact page, paragraph, or shape it came from to prevent hallucinations.
Is the knowledge published automatically by the AI?
No, K-DLC uses a governed lifecycle where the agents stop at review gates for human decisions before publishing stable knowledge.
What kind of access do the remote connectors have?
The connectors use read-only scopes, and all secrets are stored in environment variables.
Can I use K-DLC in my existing CI/CD pipeline?
Yes, the kdlc CLI provides a stable surface for scripts and pipelines to perform ingestion, linting, and governed publication without an agent in the loop.