seed-agent to do strategy, creator sourcing, AI briefs (with mandatory #ad disclosure), scheduled publishing, view / GMV measurement, and kill-and-scale budget reallocation. The core agent is open source and owned by the community. The hosted $0.01 / verified view billing is an optional add-on for teams that don't want to run infra — see Fee model.
docker compose up# 1. clone git clone https://github.com/socialseed-collective/seed-agent cd seed-agent # 2. point it at YOUR Google Cloud / Vertex AI project cp config.example.yaml config.yaml $EDITOR config.yaml # set vertex.project_id, region, budget_cap_usd ... gcloud auth application-default login # 3. bring it up — orchestrator + sub-agents + mock TikTok + metrics simulator docker compose up # → ADK orchestrator :8080 (A2A) # → mcp-mock-tiktok :7001 (publish / shop endpoints, synthetic) # → mcp-metrics-sim :7002 (verified-view + GMV simulator) # → mcp-billing :7003 ($0.01/view ledger + hard budget cap) # → console UI :3000 http://localhost:3000
vertex: project_id: my-brand-gcp-project # YOUR project — nothing phones home region: us-central1 gemini_model: gemini-2.5-pro deploy: target: agent_engine # or: local | cloud_run campaign: budget_cap_usd: 5000 # agent HARD-STOPS at views*cpv == cap cpv_usd: 0.01 # cost per verified view disclosure: "#ad" # enforced in every creator brief fraud_filter: dedup: true min_watch_time_s: 3 source_anomaly_zscore: 3.5 require_attribution: true data: mode: synthetic # live TikTok API behind a feature flag — see roadmap
One ADK orchestrator coordinates six sub-agents over A2A; every external tool (mock TikTok, metrics simulator, billing ledger) is an MCP server so you can swap in your own connectors.
The agent itself costs nothing and is owned by the community. You only pay if you choose the hosted runner instead of self-hosting. Verified-view rules are in the repo, not a trade secret.
| Way to run it | What you pay | Who owns the agent |
|---|---|---|
Self-host (docker compose up, your Vertex project) | $0 to the project | You + the community |
| Self-host on Agent Engine | $0 + your GCP bill | You + the community |
| Hosted runner (we operate it for you) | $0.01 / verified view | Still community-owned code; you rent the ops |
Every billed view passes the fraud filter (dedup · watch-time threshold · source-anomaly · attribution) and the agent hard-stops when verified_views × $0.01 hits your deposit. The hosted add-on adds a small, line-itemed ops margin on top of the $0.01 pass-through — no markup hidden inside the CPV.
We want connectors, fraud-filter test fixtures, and brief-quality evals. Good first issues are labeled good first issue help wanted connector.
list_creators / creator_stats interface in mcp/ — see mcp/mock-tiktok as the reference.type: rfc). Maintainers triage weekly; 2 maintainer approvals to merge.make test (agent unit tests + simulator golden files) before opening a PR. CI must be green.# dev loop make dev # hot-reload orchestrator + sub-agents make test # pytest agents/ + simulator golden files make eval-briefs # Gemini brief quality + #ad-disclosure check
main.seed-agent is stewarded by the SocialSeed Collective — an open group of maintainers from multiple orgs. Decisions happen in public RFC issues; releases are cut by rotating release managers. The trademark and the hosted runner business are governed separately from the codebase, and any maintainer can fork if governance fails. Roadmap is on the public project board.
On the roadmap: live TikTok / TikTok Shop API behind a feature flag (currently approval-gated) · pluggable creator-payment rails · more regional disclosure presets · A2A interop with third-party measurement agents.
Influencer ops is full-time work small DTC brands can't staff. A hosted SaaS solves that but makes you rent your own growth engine and trust someone else's billing math. seed-agent gives you the autonomous workforce and the keys: read the code, run it on your cloud, audit every verified view, and contribute the connector you wish existed. Community owns the agent.