oyhub 0.2.0

Local agent harness: central skill library, Obsidian + SQLite FTS5 memory, background curator — served over MCP.

OyHub

A local, always-on agent harness that fixes skill sprawl and context exhaustion.

Engineers using AI assistants hit two walls: skills scattered across projects with nothing organized in one place, and context windows that fill up and forget everything between sessions. OyHub is an MCP server you install once on your machine — Claude Desktop, Claude Code, and Cursor connect to it and gain:

• One central skill library with per-project activation. Skills live in ~/.oyhub/skills (agentskills.io-compatible SKILL.md format). Activate a project and only its relevant skills appear — your Kubernetes skills stay out of your React sessions.
• Persistent memory that survives any context window. Curated facts go to your Obsidian vault as plain markdown (human-readable, editable, graph-linked). Full conversation history goes to SQLite + FTS5 — full-text searchable recall at zero LLM cost. Instead of context running out, the assistant retrieves the five messages that matter.
• Loop engineering in the background. A curator loop runs on idle: archives stale skills (never deletes — archive is recoverable), dedupes memory entries, pins protect anything you care about, and every action is logged to a Curator Log.md note in your vault so the loop's work stays visible.
• Injection-scanned writes. Memory and skills persist into every future session's context, so anything flowing into them is scanned at write time — instruction-override phrasing and credential-exfil link shapes are refused with a reason. Legitimate engineering content (curl commands, env var names) passes.
• Safe under concurrent clients. Multiple assistants can run OyHub against the same state simultaneously: SQLite runs in WAL mode with busy timeouts, and all shared JSON/markdown writes go through file-locked atomic read-modify-write.
• CI/CD + Docker from day one. GitHub Actions runs lint + tests on a 6-way OS/Python matrix, builds the Docker image, cuts a GitHub release on every version tag, and publishes to PyPI via trusted publishing.

Zero runtime dependencies — Python stdlib only. No API key: the connected assistant is the intelligence; OyHub is the harness.

Install

git clone https://github.com/Asanovichoff/oyhub && cd oyhub
python3 -m oyhub doctor   # health check

Set where memory lives (your existing Obsidian vault):

export OYHUB_VAULT=~/Documents/MyVault

Connect Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "oyhub": {
      "command": "python3",
      "args": ["-m", "oyhub"],
      "env": {
        "PYTHONPATH": "/path/to/oyhub",
        "OYHUB_VAULT": "/Users/you/Documents/MyVault"
      }
    }
  }
}

Note: Claude Desktop ignores a cwd field — PYTHONPATH is what makes python3 -m oyhub importable. Alternatively, pip3 install -e /path/to/oyhub once, then use "command": "oyhub" with no args or PYTHONPATH.

Connect Claude Code

claude mcp add oyhub --env PYTHONPATH=/path/to/oyhub --env OYHUB_VAULT=~/Documents/MyVault -- python3 -m oyhub

Or after pip3 install -e /path/to/oyhub: claude mcp add oyhub -- oyhub

Docker

docker compose run --rm oyhub          # stdio MCP server
docker compose run --rm --entrypoint python oyhub -m oyhub doctor

The tools it exposes

Tool	What it does
project_activate	Scope skills + memory to a project (creates profile if new)
skill_add / skill_view / skill_list / skill_search	Central library; list shows the compact index (60-char descriptions), bodies load on demand
memory_add / memory_snapshot	Curated memory in your Obsidian vault, project-scoped or global
session_log / session_search / session_recent	FTS5-indexed history: log important turns, search them later with context windows
curator_run / curator_pin	Run maintenance now; pin skills the curator must never touch

A typical day

1. Start a session: Claude calls project_activate("webapp") and memory_snapshot — it now knows your conventions and this project's facts, without you repeating them.
2. Work normally. When something worth keeping comes up, Claude calls memory_add ("staging DB is read-only") or skill_add (a deploy procedure it just worked out).
3. Next week, in a fresh session: "how did we fix that CORS issue?" → session_search("CORS") returns the exact exchange. Nothing was lost when the old context window died.
4. Meanwhile the curator has deduped memory and archived skills untouched for 90 days — check Curator Log.md in Obsidian to see what it did.

Architecture

Claude / Cursor  ──MCP (stdio JSON-RPC)──►  oyhub/server.py
                                              │
                    ┌─────────────────────────┼─────────────────────┐
                    ▼                         ▼                     ▼
             skills.py                 memory.py               curator.py
      central library + index    Obsidian vault (curated)   idle-triggered loop:
      per-project activation     SQLite FTS5 (raw recall)   archive stale, dedupe,
      usage tracking             frozen-snapshot reads      log actions to vault

Design lineage: the frozen-snapshot memory, 60-char skill index, archive-never-delete curator, and idle-triggered maintenance are patterns from Hermes Agent (Nous Research, MIT), adapted for a keyless MCP harness.

Development

pip install -e ".[dev]"
ruff check oyhub tests && pytest tests/ -q   # what CI runs

22 tests, all offline — including concurrency tests (parallel writers, no lost updates) and injection-guard tests. CI matrix: Ubuntu + macOS × Python 3.10–3.12, plus a Docker build + smoke test. Tag v* to cut a GitHub release and publish to PyPI (trusted publishing — configure once under PyPI → Publishing).

Roadmap

• LLM-assisted curation (summarize/consolidate memory notes, improve skill descriptions)
• Skill import from agentskills.io and Claude skill bundles
• Session auto-logging middleware
• Vector search as an optional complement to FTS5

MIT license.