Open-source foundation for an agent-native services marketplace.
The core idea: many Fiverr-style freelance jobs can become structured agent workspaces. Buyers should be able to hire a service agent, give it a brief, receive a quote, track the work, inspect the evidence, request revisions, and receive a verifiable deliverable.
This repository is not a production marketplace yet. It is a working foundation: service taxonomy, workspace templates, order runtime, provider scaffolds, quality gates, escrow scaffolds, eval fixtures, and local verification.
Freelance marketplaces organize human labor around service listings.
Agent-native marketplaces will organize repeatable knowledge work around service workspaces:
- A clear service spec instead of an open-ended chat.
- A structured brief instead of back-and-forth clarification.
- Tool and API permissions instead of blind account access.
- Quality rubrics instead of vague "looks good".
- A trust ledger instead of untraceable delivery.
- Human escalation for risk, taste, regulation, and edge cases.
The long-term thesis is Agent as a Service: in the future, you will hire other people's agents to create value for you.
- 14 Fiverr top-level category baseline.
- 20 MVP service agents across writing, design, coding, data, marketing, audio, video, and business work.
- 800 generated long-tail service specs with 160 saleable candidates.
- Reusable service workspace template.
- Buyer brief schemas and deliverable schemas.
- QA rubrics, policy files, revision rules, and handoff rules.
- Local order lifecycle runtime.
- Provider/API matrix for OpenAI, GPT Image 2, Banana, Seedance, Kling, Renoise, Suno, ElevenLabs, GitHub, Vercel, WordPress, Shopify, and Google Ads.
- Provider adapter scaffolds with credential gates and dry-run traces.
- Mock escrow plus Stripe Connect scaffold.
- Human QA queue, reviewer pool scaffold, and decision records.
- Trust ledger export for replaying order evidence.
- Document, table, and widget deliverable packaging.
- 30 pilot sample orders, 200 Phase 2 simulated orders, and 100 eval fixtures.
- Static marketplace alpha UI under
web/. - One-command local verification and GitHub Actions CI.
agent_fiverr/ Local runtimes for catalog, orders, providers, QA, escrow, trust, costs
data/ Service catalog, provider matrix, pilot samples, generated fixtures
docs/ Plan, acceptance evidence, runtime notes, article drafts
schemas/ JSON schemas
scripts/ Generators and verification gates
services/ 20 generated MVP service workspaces
templates/ Reusable service workspace template
tests/ Unit and simulation tests
web/ Static buyer-facing alpha UI
python3 scripts/run_all_gates.pyExpected output:
All local gates passed.
Run individual examples:
python3 -m agent_fiverr.cli discover --category Data
python3 -m agent_fiverr.cli quote data-cleaning-formatting '{"dataset_file":"contacts.csv","target_schema":"email,name","dedupe_rules":"email","missing_value_rules":"blank","output_format":"csv"}' --package standard
python3 -m http.server 8127 --directory webThen open:
http://127.0.0.1:8127/
scripts/run_all_gates.py runs:
- catalog validation
- 30-order pilot simulation
- 200-order Phase 2 simulation
- 100-order alpha metrics simulation
- MVP cost gate
- long-tail catalog generation
- eval fixture generation
- full unit test discovery
git diff --check
Latest local evidence:
Ran 68 tests
OK
All local gates passed.
Actual API keys are never committed. Copy .env.example to .env or use a
secret manager.
The scaffold already names the expected keys:
OPENAI_API_KEYBANANA_API_KEYSEEDANCE_API_KEYKLING_API_KEYRENOISE_API_KEYSUNO_API_KEYELEVENLABS_API_KEYGITHUB_TOKENVERCEL_TOKENWORDPRESS_API_TOKENSHOPIFY_ACCESS_TOKENGOOGLE_ADS_DEVELOPER_TOKENSTRIPE_SECRET_KEY
Until real keys and live adapters are enabled, provider and payment calls stay in dry-run or planned mode.
This repo proves the local foundation. It does not yet prove production marketplace acceptance.
Remaining real-world gates:
- real provider keys and API calls
- generated media assets from real providers
- live payment/escrow flow
- staffed QA reviewer operation
- 100 real alpha orders
- observed cancellation, refund, speed, and cost metrics
- human-approved golden outputs
The X Article draft is here:
docs/articles/agent-as-a-service.md
MIT. See LICENSE.