Odin

A gated, multi-engine research orchestrator for Claude Code.

Odin turns a single research question into a planned, budgeted, multi-engine research flight. It scopes GTM questions with a senior product-marketing advisor (PMM Sherpa), routes each sub-question to the engine best shaped for it, fans out parallel "ravens" (subagents) across your internal sources, accumulates all evidence in a persistent provenance-tagged vault, and ships a synthesized report to both the vault and your notes system.

Named for the Norse god who sends his ravens Huginn and Muninn (Thought and Memory) across the worlds each morning and learns everything they saw.

What it does

/odin <your research question>

1. Classify: GTM-shaped, technical-shaped, or mixed. Technical questions never touch the PMM advisor; GTM questions get its full treatment.
2. Intake (Gate 1): external / internal / hybrid mode, depth budget, deliverable shape. Checks the vault first for prior coverage.
3. Scope: facet decomposition validated by ask_sherpa, one scope_pmm_research call per GTM facet, then merges sub-questions under strict rules (cap 5 to 7, park the rest).
4. Route (Gate 2): a routing table with per-question engine, justification, and cost estimate. Long/expensive engine runs require explicit per-row approval before launch.
5. Execute: parallel legs across three engines:
   • hyperresearch for deep, contested, or academic questions
   • Deep-Research-skills for structured items-by-fields comparisons (competitive matrices)
   • Raven subagents for internal sources (Slack, email, docs, meeting transcripts, trackers)
6. Synthesize (Gate 3): blends external and internal evidence, hunts internal-vs-external tension, pressure-tests the thesis once with get_feedback, then writes the report to the vault and your notes folder. Optionally drafts a named PMM artifact (battlecard, comparison matrix, GTM brief, positioning, etc.).

Every evidence note is provenance-tagged. Internal-origin evidence never enters an externally-shareable artifact without explicit per-item sign-off. Every flight writes a resumable run manifest.

Prerequisites

Odin is an orchestration skill: it assumes three underlying systems are installed and working. Set them up in this order.

0. Claude Code

Odin runs inside Claude Code (CLI, IDE extension, or desktop). Install it and sign in first.

1. hyperresearch (deep research engine + vault)

Requires Python 3.11–3.13 (3.14 is NOT supported; the install will fail on it).

# macOS, via Homebrew + pipx (recommended):
brew install python@3.13 pipx
pipx ensurepath
pipx install hyperresearch --python /opt/homebrew/opt/python@3.13/bin/python3.13

# Create your research workspace and install into it:
mkdir -p ~/research-workspace && cd ~/research-workspace
hyperresearch install

This initializes the vault, injects the pipeline skills/agents into .claude/, and (if crawl4ai is detected) configures free local web fetching. Verify with hyperresearch status. Full docs: jordan-gibbs/hyperresearch.

2. Deep-Research-skills (comparison matrix engine)

Installed by copying files into the SAME workspace:

git clone --depth 1 https://github.com/Weizhena/Deep-Research-skills /tmp/drs
cd ~/research-workspace
cp -R /tmp/drs/skills/research-en/* .claude/skills/
cp /tmp/drs/agents/web-search-agent.md .claude/agents/
cp -R /tmp/drs/agents/web-search-modules .claude/agents/
python3 -m pip install pyyaml   # its only dependency
# (on Homebrew Python add: --break-system-packages)

Full docs: Weizhena/Deep-Research-skills.

3. PMM Sherpa MCP (GTM judgment layer)

A remote MCP server (OAuth sign-in, no self-hosting). Add it to Claude Code:

claude mcp add --transport http pmm-sherpa https://pmmsherpa.com/api/mcp

Then authenticate when prompted. Setup guides and the calling-guide skill: boommark/pmmsherpa-mcp. Sherpa is metered by credits; Odin enforces strict call discipline (at most 2 ask_sherpa + 1 scope_pmm_research per facet + 1 get_feedback per flight).

Sherpa is only exercised for GTM-shaped questions. If you only ever run technical research, Odin works without it (those flights skip the advisor entirely).

4. Internal connectors (optional, for internal/hybrid mode)

Odin's ravens use whatever MCP connectors your Claude Code environment has: Slack, Microsoft 365 (Outlook/SharePoint), Granola, Atlassian, etc. Connect the ones you use; edit references/internal-source-map.md to match your toolset. External-only flights need none of these.

Installing Odin

Clone this repo into your research workspace's skills folder:

cd ~/research-workspace
git clone https://github.com/boommark/odin .claude/skills/odin

Then edit the Configuration section at the top of SKILL.md (notably OBSIDIAN_REPORTS_DIR, the folder where final reports are mirrored; point it at your Obsidian vault or any notes folder).

Open Claude Code in the workspace folder and run:

/odin <your research question>

Repo layout

SKILL.md                              # the orchestrator (phases, gates, rules)
references/internal-source-map.md     # Sherpa source classes → your internal surfaces
references/run-manifest-template.yaml # resumable per-flight state
references/note-frontmatter.md        # provenance tagging convention + firewall rule

Design notes

• Engines are folder-scoped: Odin, hyperresearch, and Deep-Research-skills must live in the same workspace, and Claude Code must run from it.
• Cost control is structural: light tier is the default, full-tier hyperresearch (~$60-120 of model spend, 1.5-2.5 hr) requires explicit per-row approval at Gate 2.
• The manifest's template flag exists so a completed, approved flight can later be re-run headlessly on a schedule. Odin never creates templates on its own.

License

MIT