Context Harness MCP

An MCP server that exposes a Knowledge Graph (nodes, observations, relations) to Claude Code or any MCP-compatible client. Storage is Postgres + pgvector; semantic search runs locally via all-MiniLM-L6-v2 ONNX embeddings. Runs anywhere with Docker + Postgres+pgvector — local docker compose on a dev machine, or any container host (Railway, Render, Fly, Coolify, self-hosted, …) for team-shared deployments.

What it does

6 MCP tools to manage a knowledge graph — create_nodes, add_observations, create_relations, search_nodes, open_nodes, read_graph. See docs/mcp-tools.md.
Semantic search via 384-dim all-MiniLM-L6-v2 embeddings indexed with pgvector HNSW cosine. search_nodes("authentication patterns") returns the nodes whose observations are about auth, not just substring hits.
Content Filter on every write — three layers (size + junk denylist, secrets scan with gitleaks, taxonomy enforcement) reject payloads with secrets, oversized text, or out-of-taxonomy node/relation types before any DB transaction opens. Atomic reject — never partial writes.
MCP-protocol compatible. Standard MCP streamable-http transport; works with Claude Code and any other MCP-compatible client. JSON wire shapes are stable and documented in docs/mcp-tools.md.

Install

Option A — Local (Docker)

You need Docker and a reachable Postgres with the pgvector extension. The easiest target is a free Supabase project (no credit card), but any Postgres 16 + pgvector works.

git clone https://github.com/valianx/context-harness-mcp
cd context-harness-mcp
cp .env.example .env
# Edit .env and set DATABASE_URL to your Postgres DSN.

docker compose --profile migrate run --rm migrate   # apply schema (one-time)
docker compose up                                    # start the server

The server listens on http://localhost:7654/mcp. First build takes ~5 minutes (ONNX runtime + model download); subsequent ups are instant.

To use it from Claude Code, add to ~/.claude.json:

{
  "mcpServers": {
    "memory": {
      "type": "http",
      "url": "http://localhost:7654/mcp"
    }
  }
}

Full local runbook — DB options, migrations, troubleshooting, smoke tests — in docs/local-stack.md.

Option B — Cloud (any container hosting + any Postgres+pgvector)

Same Docker image, deployed to whatever container host you prefer (Railway, Render, Fly, Coolify, your own VPS, …) and pointed at any Postgres 16+ with pgvector (managed: Supabase, Neon, Railway Postgres, RDS, …; or self-hosted). Continuous deploy via the included GitHub workflows (goose up + optional deploy hook), weekly encrypted pg_dump backups, and a configurable keepalive cron for providers that auto-pause.

One-time setup runbook — principles plus equally-weighted per-platform examples — in docs/deployment.md.

Documentation

| File | What's in it | |---|---| | docs/mcp-tools.md | The 6 MCP tools — arguments, responses, examples, error codes. | | docs/local-stack.md | Local Docker runbook (Option A above, expanded). | | docs/deployment.md | Cloud deployment runbook (Option B above, expanded). | | docs/auth.md | Auth runbook — Supabase setup, dev flow, webhook config, revocation, troubleshooting. | | docs/knowledge.md | Durable decisions, constraints, patterns, stack notes. | | CHANGELOG.md | Release notes (Keep-a-Changelog format). | | CLAUDE.md | Repo conventions for AI-assisted contributors. |

Tech stack

Go 1.23 + mcp-go + pgx/v5 + pgvector-go + fastembed-go (ONNX). Migrations via goose. Tests via testcontainers-go. Operator tooling (khctl) is also Go — no Python/uv runtime required.

Repo conventions

Conventional commits (feat(area): …, fix(area): …, …).
Feature branches: feat/<kebab>, fix/<kebab>, docs/<kebab>, chore/<kebab>, refactor/<kebab>.
Never commit on main. Every change ships via pull request.
No secrets in the repo. Ever.

Full working agreements in CLAUDE.md §6.

License

Not yet declared. Treat as proprietary until a LICENSE file lands.

MCP Servers