OpenClaw

OpenClaw is an agentic pipeline that converts meeting transcripts into structured action items, tasks, and follow-ups. It runs on a dedicated M4 Mac Mini node and integrates with the team's communication channels.

Pipeline Architecture

The meeting-to-action flow:

Meeting Recording
  -> Fireflies.ai (transcription)
    -> n8n (workflow automation)
      -> OpenClaw webhook (processing)
        -> Discord / Signal (output)

Processing Stages

1. Transcript ingestion -- Fireflies.ai captures meeting audio and produces timestamped transcripts
2. Webhook delivery -- n8n workflow triggers on new transcript, sends to OpenClaw endpoint
3. AI processing -- Cloud Claude APIs (Sonnet for extraction, Opus for complex reasoning) parse the transcript into structured data
4. Action generation -- Extracted items are formatted as actionable tasks with owners and deadlines
5. Channel delivery -- Results posted to Discord and Signal channels

Model Selection

The pipeline uses a tiered model approach to balance cost and quality:

Task	Model	Why
Heartbeat/health checks	Ollama llama3.2:3b (local)	Zero cost, no API dependency
Action extraction	Claude Sonnet	Good structured extraction at lower cost
Complex reasoning	Claude Opus	Required for ambiguous context resolution
Classification	Claude Sonnet	Sufficient for categorization tasks

Monthly API cost estimate: ~$35-90 depending on meeting volume.

Communication Channels

Discord

Primary output channel for the development team. Actions are posted as structured embeds with:

• Task description
• Assigned owner
• Priority level
• Meeting source and timestamp

Signal

Operates in allowlist mode for sensitive discussions. Only pre-approved numbers receive action items. Used for community-facing coordination where Discord is not appropriate.

Infrastructure

Hardware

A dedicated M4 Mac Mini serves as the OpenClaw node:

• Local Ollama instance for zero-cost heartbeat checks
• n8n workflow engine for automation orchestration
• Persistent storage for transcript history and action tracking

API Configuration

Cloud Claude APIs require API keys (not OAuth subscriptions -- Anthropic restricts subscription OAuth in third-party tools). The pipeline uses direct API key authentication.

Key environment requirements:

• ANTHROPIC_API_KEY -- For Claude Sonnet/Opus processing
• FIREFLIES_API_KEY -- For transcript retrieval
• DISCORD_BOT_TOKEN -- For channel posting
• SIGNAL_CLI_PATH -- For Signal message delivery

Integration with Green Goods

OpenClaw generates action items that feed into the Green Goods development workflow:

• Meeting decisions become GitHub issues or plan files
• Sprint commitments are tracked against grant deliverables
• Community feedback from meetings is routed to the appropriate agent workflow

The pipeline does not directly create code changes. Its output is structured input for human review and agent task assignment.

Operational Notes

• Transcripts are stored locally on the Mac Mini, not in cloud storage
• The pipeline is designed for the team's meeting cadence (weekly syncs, ad-hoc standups)
• Action items include confidence scores -- low-confidence items are flagged for human review
• The system runs independently of the main Green Goods infrastructure

Next best action

See how Gemini is used alongside other AI tools in the project.

Gemini

• Codex
• MCP Guide