marketerloop

Marketer-in-the-loop agentic workflows. Open source, BYOK, and every output waits for a human before it goes anywhere.

A workflow turns a source signal into drafts, then puts every external write behind review. Accepted edits become few-shot examples for later drafts. The runtime and catalog live here. This is the parent project, not the name of one template.

git clone https://github.com/DeepanshuPal/marketerloop
cd marketerloop
pip install -e .
cp .env.example .env

marketerloop run meeting-to-content --sample
marketerloop queue
marketerloop show <draft-id>
marketerloop edit <draft-id>

The shared runtime

The common path is deliberately small and inspectable:

source -> normalize evidence -> qualify/extract -> draft -> approval queue -> outcome

• Templates are folders with a readable template.yaml, versioned prompts and offline evals.
• Runs, source evidence, prompt/model versions, costs, decisions, edit diffs and outcomes are append-only in local SQLite.
• API keys stay in environment variables. Mock mode runs CI and the sample without keys or cost.
• Budgets fail closed before model calls.
• Nothing posts, sends or touches a social account automatically.

There is no visual builder and no hosted account. Clone it, bring your own keys, and keep your data in your own SQLite file.

Workflow catalog

Templates that fit the shared runner stay as folders. Specialized workflows ship as focused repos while their interfaces settle, but they are catalog entries under Marketerloop, not separate product bets.

workflow	status	what it does
meeting-to-content	native template	Transcript or notes -> timestamped ideas -> LinkedIn post, X thread and newsletter blurb -> approval queue. Confidential lines are flagged and excluded.
fresh-intent-reply-queue	standalone workflow	Reddit RSS + HN intent signals -> ICP match -> scored reply drafts -> human approval.
linkedin-visitor-conversion	standalone workflow	Manual LinkedIn visitor/follower CSV -> ICP qualification -> connection-note drafts -> human approval and export.
citation-queue (spec)	spec, not built	am-i-cited citation feed -> durable cited threads -> participation drafts -> human approval. The durable-signal companion to fresh intent.

All three share the same operating rules: local state, BYOK models, append-only run evidence, explicit budgets, and no external write without human approval. As their contracts stabilize, they can move behind the common runner without breaking their focused CLIs.

Native template contract

templates/meeting-to-content/
  template.yaml
  prompts/*.v1.md
  evals/
  README.md

template.yaml declares inputs, connector scopes, trigger, DAG, approval gates, budgets and state. The runner reads that contract. There is no hidden workflow behavior.

The meeting-to-content template accepts a transcript through manual paste, a file or stdin, or from Granola. It extracts timestamped ideas, excludes confidential material, and drafts three formats: LinkedIn post, X thread and newsletter blurb. Every draft stops in the queue.

Connectors

connector	status	cost
manual paste/file/stdin	live and verified	free
Granola	implemented against the public API docs; live account path not verified	Business-plan API key required
Firecrawl	interface stub	free tier exists
Spider Cloud	interface stub	free tier exists
Browserbase	interface stub	paid
Exa	interface stub	paid

A missing connector key means “not configured”, not a broken run. Stubs are labelled as stubs so the architecture is visible without pretending unfinished integrations work.

Models and keys

One OpenRouter key can route all model calls, or use OpenAI or Anthropic directly.

env var	purpose
OPENROUTER_API_KEY	recommended model router
OPENAI_API_KEY / ANTHROPIC_API_KEY	direct provider keys
GRANOLA_API_KEY	Granola Business API
MOCK_LLM=1	deterministic local mode used by CI and evals

The default model is configurable with MARKETERLOOP_MODEL. Limits are declared in the template (weekly_llm_usd, max_runs_per_day) and checked before calls.

Verification

pip install -e ".[dev]"
MOCK_LLM=1 pytest -q
MOCK_LLM=1 python templates/meeting-to-content/evals/eval.py
MOCK_LLM=1 marketerloop run meeting-to-content --sample
marketerloop queue

The suite covers the pipeline, approval state, template loading and CLI. The offline eval checks extraction, evidence coverage, confidentiality exclusions and all three draft formats. Granola still needs a real Business-plan key for a true live connector test; the README will not call that verified until it is.

Develop

marketerloop templates list
marketerloop connectors
marketerloop runs
marketerloop budgets

Python 3.11+ is recommended; mock mode works on 3.10. MIT licensed.