frames-examples

A reference repo showing how to maintain multiple live datasets with frames + an agent harness (opencode) on a daily cron.

Inspired by the blindspot.news pipeline — same pattern, but the writes happen through MCP tools instead of JSON blobs, so every fact has a source, every change is an event, and the git history doubles as an audit log.

What's here

Datasets, each in its own directory under datasets/:

Dataset	What it tracks
mcp-servers	Every notable MCP server on GitHub. Recursive: frames cataloging the protocol it speaks.
ai-promises	Public roadmap commitments by AI labs, with a shipped? column.
ai-models	Every notable model release, every vendor, every modality.
elon-tweets	Recent posts from @elonmusk on X, classified by topic.
blindspot-briefs	Daily 8-section analytical briefs on yesterday's news.
blindspot-threads	Non-obvious connections across each day's briefs.
agent-networks	Open questions on the agent / machine-network economy, with rolling evidence.
agent-wallets	Live catalog of agent wallets, registries, and payment protocols + the services each one exposes.

Each dataset is a self-contained frame: schema.yml (the contract), prompt.md (the maintenance loop), events.ndjson (the append-only log), and a regenerated SQLite index under .frame/.

How it works

.github/workflows/tick.yml   # daily cron, matrix over datasets
opencode.json                # opencode config: provider + 3 frame MCP servers
.mcp.json                    # same MCP wiring for Claude Code / other clients
datasets/<name>/prompt.md    # per-dataset agent loop
datasets/<name>/schema.yml   # field contract
datasets/<name>/events.ndjson# append-only event log (the source of truth)

Every day:

1. GitHub Actions starts a job per dataset.
2. opencode boots with the dataset's frame MCP server attached, plus the frames registry for paid discovery APIs (Exa, GitHub, Hugging Face).
3. The agent reads prompt.md, queries current state, fetches sources, writes facts via add_entity_with_facts / set_fact / attach_evidence.
4. The index is regenerated; the commit is pushed.

No bespoke pipeline code — the loop logic lives in plain markdown prompts.

Running locally

You need OPENROUTER_API_KEY and an agentwallet for the registry. Then:

# tick one dataset
opencode run "$(cat datasets/mcp-servers/prompt.md)"

# inspect any dataset
npx -y @frames-ag/frame query datasets/ai-models --all

Or attach Claude Code (it will pick up .mcp.json automatically) and ask:

Read the schema and prompt under datasets/ai-promises and run one tick.

Adding a new dataset

npx -y @frames-ag/frame init datasets/<name>
# edit datasets/<name>/{schema.yml,README.md,prompt.md}
npx -y @frames-ag/frame init-mcp     # rewrites .mcp.json
# add <name> to the matrix in .github/workflows/tick.yml

That's the whole authoring surface: a schema, a scope description, and a loop prompt. The engine is the same.

Credits

• frames — the protocol and CLI
• opencode — the agent harness
• blindspot.news — the pipeline shape that inspired this layout