TL;DR
- WorkflowCam is an open-source starter for turning timestamped meeting/workflow moments into operator-ready actions
- It is API-free by default so anyone can run it locally without VideoDB keys, LLM keys, or private Agent88 infrastructure
- It emits Agent88-style outputs: MemPalace note, CRM payload, and Kanban task suggestion
- The architecture is designed so VideoDB CaptureSession / RTStream / search can be added later behind a clean adapter
- The goal is simple: give workflow agents eyes, ears, and auditable memory instead of generic summaries
Most AI agents can read text. Very few can observe real work.
But real business workflows do not happen inside neat prompts. They happen during Zoom calls, screen shares, browser sessions, product walkthroughs, dashboards, approval loops, and messy operating conversations where the important detail is buried somewhere between minute 3 and minute 38.
That gap is exactly what WorkflowCam is built to address.
WorkflowCam is Agent88’s open-source starter project for the VideoDB hackathon theme: give agents “eyes and ears.” The project is intentionally shipped in an API-free, open-source-safe form first. No private keys. No hidden integrations. No dependency on a closed internal stack. Just a clean local codebase that demonstrates the workflow-memory thesis properly.
The core job is straightforward:
Take timestamped workflow moments, search them locally, and convert them into evidence-backed business actions.
That means outputs like:
- searchable memory hits
- client pain summaries
- next-step recommendations
- CRM update payloads
- proposal notes
- Kanban task suggestions
Not “AI magic.” Just useful operator infrastructure.
The problem WorkflowCam solves
In most companies, the valuable context disappears almost immediately after the meeting ends.
A client explains the real workflow pain once. Someone says they will send the Drive folder later. A scope boundary gets mentioned casually. A key objection appears in passing. Then the call ends, the notes are partial, and everyone works from memory.
This creates three problems:
1. Important moments are hard to retrieve.
People remember that “the client mentioned something about reports,” but not when, or exactly how they framed it.
2. Business actions drift away from source evidence.
The CRM note, proposal scope, and team understanding slowly diverge from what was actually said.
3. Agents stay trapped in text-only workflows.
Without perception and memory over media, the agent can only summarize what a human has already cleaned up.
WorkflowCam is a small but concrete answer to that problem.
What the open-source starter does
The current release is local-first and API-free.
It ships with sample timestamped memory and a CLI that lets you:
- run an offline demo
- search workflow moments using natural language
- generate a simple operator brief
- export Agent88-style files for later human review
Example questions:
- “Where did the client describe the reporting pain?”
- “What did they ask us to send next?”
- “What should go into the proposal scope?”
Example outputs:
- a markdown operator note
- a CRM-style JSON payload
- a Kanban task suggestion JSON file
This matters because it makes the business output shape clear even before any live API integration is added.
Why it is API-free first
We deliberately kept this release free of VideoDB credentials, OpenAI calls, Anthropic calls, or private Agent88 backend requirements.
That decision does three useful things.
First, it makes the repo truly open-source ready. Anyone can clone it and run it immediately.
Second, it keeps the architecture honest. The public version proves the workflow structure without hand-waving about hidden APIs.
Third, it gives us a clean adapter boundary. When we add VideoDB later, it plugs into a defined memory interface instead of being tangled into the app.
This is the right order for an open-source launch: show the operating model first, then wire in perception APIs second.
How it fits the Agent88 stack
WorkflowCam does not expose private Agent88 systems. But it mirrors Agent88 operating patterns in a safe form.
1. MemPalace-style note output
The project exports a markdown note containing:
- the source query
- timestamped evidence
- operator interpretation
- suggested next step
- explicit human approval requirement
2. CRM-ready payload
The project exports a JSON object representing a structured workflow-memory capture event.
That means a future private deployment can map it into a lead note, opportunity update, or client follow-up system.
3. Kanban-ready task suggestion
The project also exports a JSON task recommendation shaped for triage.
That is important because good workflow agents do not stop at summarizing. They route work.
In other words, WorkflowCam is designed to sit naturally beside:
- meeting intelligence
- CRM update flows
- proposal generation
- task orchestration
- human approval checkpoints
That is why it is relevant to Agent88 rather than being a random hackathon demo.
Where VideoDB fits later
In a production or hackathon-live version, the local JSON provider can be replaced with a VideoDB-backed provider.
The intended flow is:
CaptureSession / RTStream
→ transcript + scene indexing
→ semantic search over workflow moments
→ timestamped evidence
→ WorkflowCam operator outputs
→ Agent88 private workflow actions
That gives the agent access to what happened during the call or session, not just what someone manually typed afterward.
We like this architecture because it keeps the roles clear:
- VideoDB provides perception, indexing, search, and playable evidence
- WorkflowCam provides business interpretation and operator outputs
- Agent88 provides private workflow orchestration, approvals, and downstream action
That separation is cleaner, safer, and easier to maintain than trying to collapse everything into one giant app.
Why this is more useful than a flashy demo
A lot of hackathon projects impress for five minutes and then disappear because they are not attached to a real workflow.
WorkflowCam is intentionally narrower than that.
It is not trying to be a full autonomous video agent platform on day one. It is proving one specific thing:
If an agent can search what happened in a workflow session, it can create better CRM notes, better proposal scope, and better follow-up actions.
That is a useful wedge.
It is also a credible wedge for:
- agencies that live on meetings and reports
- consultants handling discovery calls
- operators reviewing product walkthroughs
- workflow teams that need auditability over what was actually discussed
The simplest useful product is often the right first one.
Open-source repo structure
The codebase includes:
- a CLI entrypoint
- a local memory/search layer
- templates for action generation
- Agent88-style export helpers
- adapter stubs for future providers like VideoDB
- sample demo data
That makes it easy for someone to:
- fork it
- swap in their own provider
- adapt the export layer for their own CRM or ticketing system
- use it as a base for a real-time workflow memory product
What comes next
The obvious next step is a real VideoDB adapter.
That would allow WorkflowCam to:
- ingest recordings directly
- search indexed transcript moments
- optionally use visual scene search
- return playable evidence clips
- support live workflow-memory sessions instead of only offline examples
But the point of this release is that the workflow already makes sense before that step.
That is usually a good sign.
Closing thought
The future of useful agents is not just better prompting. It is better grounding.
If agents are going to operate in real business workflows, they need to do more than generate polished text. They need to observe work, remember what mattered, and produce actions that humans can verify.
WorkflowCam is a small open-source step in that direction.
If you want the codebase, start with the repo. If you want the private deployment path, that is where Agent88 comes in.