CliDeck MCP

An open-source framework for version-aware, source-backed knowledge that AI agents query through MCP. Known answers use deterministic PostgreSQL retrieval; incomplete answers create durable research demands. Isolated Codex runs discover and extract evidence; the core validates and publishes immutable revisions. CliDeck provides guidance and never connects to devices or executes commands.

Capabilities

• Network Knowledge: device context, commands, workflows, snapshots and redaction, change review, verification, upgrades, topology and expert tasks. Exact context ranks first; broader guidance is labelled and unsupported questions remain unknown. Coverage varies by vendor, platform and version.

• Domain Packs: subject-specific schemas and deterministic validation over the shared revision/release core. Engineering Measurements demonstrates exact decimal values, units and tolerances outside networking.

• Continuous research: 1–8 enabled isolated executors, independent Luna/High models and official price comparison in /admin/models; downloads, conversion, OCR, chunking, indexing and publication are mechanical. AI receives bounded tasks through the researcher bridge. See model policy.

• WebMCP: six typed browser tools share a versioned case with the engineer. Complete files stay local; selected evidence is redacted before transmission. Browser-agent evidence access requires explicit opt-in. Manual controls work without WebMCP. See contract and walkthrough.

• Operations demo: /admin and /demo use the same frontend and real production counters. Server-side projections redact private data; demo controls cannot mutate production.

Live totals are available in the demo and /public/v1/stats; historical counts and test results are not current service guarantees. Codex was the primary engineering environment. Pipeline use through a local authenticated Codex installation needs no separate model API integration and remains subject to plan and usage limits.

Related MCP server: mcp-rag-bridge

Run locally

Requirements: Node.js 24, pnpm as pinned in package.json, PostgreSQL 16 (Docker Compose is provided).

cp .env.example .env
set -a
source .env
set +a
pnpm install --frozen-lockfile
docker compose up -d postgres
pnpm db:migrate
pnpm db:seed
pnpm knowledge:reindex-applicability -- --resume --verify
pnpm build
pnpm dev:api

Edit .env for your local database before sourcing it. MCP listens at http://127.0.0.1:8787/mcp. Worker/researcher entry points and all supported commands are in package.json. Production uses the canonical deployment workflow.

To connect the hosted service with the Codex CLI:

codex mcp add clideck --url https://mcp.clideck.com/mcp
codex mcp list

Other Streamable HTTP MCP clients can use the same endpoint.

Develop

pnpm check
pnpm test
pnpm eval
pnpm build

PostgreSQL integration tests and the 250-case evaluation need a migrated, seeded test database. A run that skips DB tests is not full release validation; the deploy script provisions its own disposable database and applies grants.

Create a pack:

pnpm domain:create -- --id marine-science --name "Marine Science"
pnpm install --lockfile-only
pnpm domain:validate -- --id marine-science
pnpm --filter @clideck/domain-marine-science test

See agent rules, documentation map, architecture, pack authoring and security.

License and privacy

Apache-2.0 covers code and project-authored fixtures, not production knowledge, third-party manuals or user data. Internal evidence and secrets stay private. The dedicated provenance tool exposes only allowlisted metadata for active revisions; the operations demo retains its stricter redaction boundary. See DATA-NOTICE.md.