# Eranos – the two-sided marketplace

Spare tokens for good. Your unused AI capacity, doing real work for nonprofits.

Eranos ([live site](https://eranos.ta-amps.workers.dev) · [3-minute film](https://agora.ta-amps.workers.dev/)) is a two-sided marketplace for spare AI capacity, built for the Meta Global AI Developer Hackathon 2026. People who want to help donate unused tokens, and their AI agents build and review agent kits and cited public-information reports. Nonprofits and public-good groups describe a need in plain words and get a reviewed result, such as an intake checklist, a local document workflow, or a cited summary of tenant-deposit laws. A different volunteer's agent checks every deliverable before anyone can use it (`drafter != reviewer`), and people keep every decision about a person.

An **agent kit** is a seed pack: task objects (`nouns`) and tasks (`verbs`) generated for one NGO's agent. One task object can hold an agent's whole setup, including the harness, the model and the files in its working directory.

- **"I want to help":** point your agent at an open task (`verb`). It drafts a kit or public-source report, or independently reviews one that a different volunteer's agent built, using your spare tokens. [Help build a kit](#help-build-a-kit).
- **"I'm an NGO looking for agentic solutions":** describe the job in plain English on [`/nonprofits`](https://eranos.ta-amps.workers.dev/nonprofits) or the public [request form](https://github.com/ta-amps/eranos-public-marketplace-for-public-good-token-burn/issues/new?template=request-a-kit.yml) (no internal jargon required). A volunteer checks it, and a need that passes becomes a task brief that volunteers' agents can claim. Hotlines, legal-aid desks and small nonprofits then get a kit or cited report that a second agent, and for sensitive work a steward, has reviewed.

## Push the Frontier + Build for Public Good

- **Why we built this (problem specificity):** AI models are smart, yet everyday agent use is frustratingly dumb: models miss post-cutoff facts, rewrite non-novel code, and stall on multi-step queues. At the same time, resource-constrained nonprofits lack the AI budget and engineering hours to build safe, auditable tools, while developers and subscribers leave paid token quotas unused every billing cycle. Eranos pairs civic token donation with a shared atomic task and task-object engine ([Tessera](https://github.com/ta-amps/tessera-infrastructure-work-and-storage-engine)) and context/model shim ([Herma](https://github.com/ta-amps/herma-infrastructure-shim-for-intelligent-agents)) so agent work compounds instead of starting from scratch.
- **Target users:**
  1. **Resource-constrained nonprofits, legal-aid clinics, hotlines and mutual-aid teams** using [Eranos](https://eranos.ta-amps.workers.dev/nonprofits) in plain language, with zero AI setup needed for public-research requests and self-contained local kits for file-touching workflows.
  2. **Open-source agent builders and token donors** connecting Muse Code, OpenCode, Claude Code or Pi via [Agora](https://agora.ta-amps.workers.dev/), [Tessera](https://github.com/ta-amps/tessera-infrastructure-work-and-storage-engine) and [Herma](https://github.com/ta-amps/herma-infrastructure-shim-for-intelligent-agents).
- **Where your data goes (privacy and contributor-tier refusal):**
  - **Client files stay with the nonprofit:** `eranos/privacy.py` screens and refuses personal identifiers on public intake. Questions answerable from official public records become cited reports; anything touching internal NGO records becomes an **agent kit** that runs on the nonprofit's own computer so client files never go to Eranos or volunteer workers.
  - **Hard block on training-contributor tiers:** Standard mode is enforced in code (`ContributorTierRefusedError` across Herma and seed-pack runtimes, plus `--contributor-mode off` in Muse Code), refusing `contributor` and `*-contributor` model IDs so sensitive or nonprofit prompts never enter provider training pools.
  - **Local open-weights option:** Nonprofits and volunteer workers can run open-weights **Llama 3.2 1B and 3B Instruct** (with local PEFT LoRA adapters) offline on a single consumer GPU or laptop, keeping inference 100% local.

## Which Meta models do what

Across the `ta-amps` ecosystem ([Eranos](https://github.com/ta-amps/eranos-public-marketplace-for-public-good-token-burn), [Agora](https://github.com/ta-amps/agora-public-overview-architecture-setup-and-contribution), [Tessera](https://github.com/ta-amps/tessera-infrastructure-work-and-storage-engine), [Herma](https://github.com/ta-amps/herma-infrastructure-shim-for-intelligent-agents), and [seed-packs](https://github.com/ta-amps/seed-packs-library-of-nouns-and-verbs)), Meta open-weights models, fine-tuned PEFT/QLoRA adapters, Meta Model API endpoints, and multimodal vision/audio pipelines handle distinct, measured roles:

| Meta model / tool | Role in the system | Measured result & verification |
| --- | --- | --- |
| **Muse Glimmer 30B (QLoRA fine-tuned, open weights)** | Specialist typed MCP tool-call adapter trained with `peft` + `bitsandbytes` 4-bit QLoRA on two free Kaggle T4 GPUs (synthetic template data only; unadmitted candidate) | Pre-registered 150-row fresh holdout exact match rose **`93.3% -> 99.3%`** (`+6.0%`, 95% CI `[+2.0%, +10.7%]`), valid JSON **`100.0%`**, 12/12 promotion checks passed ([evidence](https://github.com/ta-amps/agora-public-overview-architecture-setup-and-contribution/blob/main/evidence/glimmer-30b-2026-10-29/README.md)) |
| **Llama 3.2 1B & 3B Instruct (PEFT LoRA fine-tuned, open weights)** | Local continuous tool-call tuning (`torch`, `transformers`, `peft`, `accelerate` on 12 GB RTX 3060), multi-LoRA parametric memory (`load_adapter` / `set_adapter`), and local Herma shim routing for offline NGO kits | Frozen unseen-family holdout exact match rose **`12.0% -> 50.7%`** (1B) and **`52.0% -> 72.7%`** (3B), cutting stringified typed arguments from `17.5%` to `0.0%` ([method & receipts](https://github.com/ta-amps/tessera-infrastructure-work-and-storage-engine/blob/main/docs/continuous-tuning.md)) |
| **Muse Spark 1.3 (`api.meta.ai`)** | High-capability worker reasoning for drafting multi-step NGO agent kits, decomposing requests on `eq01`, and auditing source citations on `eq03` (`--contributor-mode off`) | Exercised in authorized worker runs with strict `ContributorTierRefusedError` guardrails ([providers & compute](https://github.com/ta-amps/agora-public-overview-architecture-setup-and-contribution/blob/main/docs/providers-and-compute.md)) |
| **Muse Wonder 1.0 (`api.meta.ai`)** | Fast, low-latency worker turns for structured JSON schema validation, synthetic fixture generation, and bounded classification | Configured in Herma's Meta Model API adapter with per-attempt receipt logging ([providers & compute](https://github.com/ta-amps/agora-public-overview-architecture-setup-and-contribution/blob/main/docs/providers-and-compute.md)) |
| **Muse Code & nanoMuse** | Native coding-agent harnesses claiming leased verbs over Tessera MCP (`/mcp`) and HTTP (`/v1/*`) alongside OpenCode, Claude Code and Pi | Native Muse Code -> Meta lifecycle completed 12 provider calls end to end against `workerd` ([readiness evidence](https://github.com/ta-amps/agora-public-overview-architecture-setup-and-contribution/blob/main/evidence/readiness-2026-10-29.json)) |
| **Meta SAM (`facebook/sam-vit-base`) + `wav2vec2` (`facebook/wav2vec2-base-960h`)** | Multimodal video/audio quality assurance: SAM UI-region mask segmentation (`video-qa`) and local `wav2vec2` CTC speech-to-caption verification | Verified all 4 rendered showcase cuts locally on CPU (`CER 0.013` and `0.016`, 16 cryptographic task receipts, zero paid API calls) ([video QA receipts](https://github.com/ta-amps/agora-public-overview-architecture-setup-and-contribution/blob/main/evidence/video-production-2026-10-30/README.md)) |

## How it works: tasks (things to do) and task objects

It is not a new model architecture; it is how agent work is organized so results are checked and reusable. On the wire, **verbs are tasks (things to do)** and **nouns are task objects** -- while the `/nonprofits` web page keeps plain everyday words so nonprofit teams never need to learn internal terms. Everything in `ta-amps` is made of four parts:

- **task objects (`nouns`)** are versioned, reusable things: knowledge, read-only MCP tool contracts, tested code, feeds, practice datasets, LoRA adapters, agent kits and verified reports.
- **tasks (things to do, `verbs`)** are leased units of work with acceptance checks. An authorized worker claims one, runs it and gets a cryptographic receipt.
- **Surfaces** are where people and agents come in. There are exactly two public surfaces: **Eranos**, this marketplace, and **[Agora](https://agora.ta-amps.workers.dev/)**, the single overview of architecture and setup, interactive task & task-object explorer, use cases, connecting any agent and contributing.
- **Infrastructure** is the narrow shared engine underneath. **[Tessera](https://github.com/ta-amps/tessera-infrastructure-work-and-storage-engine)** keeps storage, permissions, compare-and-set versions, leases and receipts; a lease means only one agent works on a task at a time. **[Herma](https://github.com/ta-amps/herma-infrastructure-shim-for-intelligent-agents)** sits between an agent's harness and its model to retrieve task objects as context or answer directly.

Everything else is nouns and verbs, bundled in seed packs so they are easy to share. Packs live in one
[pack library](https://github.com/ta-amps/seed-packs-library-of-nouns-and-verbs), carry no brand name and stay separate from the
surfaces. Workers, meaning anyone's AI agents, do the thinking. The server keeps the to-do list, the results and
the permissions. Home is not a surface; its address sends people to Agora.

![People and AI agents use the two surfaces, which lead to nouns and verbs. Workers claim verbs and make nouns. Seed packs bundle nouns and verbs. Infrastructure supports them.](docs/images/nouns-and-verbs.svg)

In Eranos, the nouns are the tracks, the task briefs, the agent kits, their synthetic test cases and the review
records. The verbs are checking a brief, building and self-testing a kit, and reviewing it. A person signs off on
sensitive kits. [More detail](docs/nouns-and-verbs.md).

## Try it on your computer

You need [uv](https://docs.astral.sh/uv/). The web app also needs Node 22.13 or newer and pnpm. No AI key or
cloud account is needed.

```bash
git clone https://github.com/ta-amps/eranos-public-marketplace-for-public-good-token-burn.git && cd eranos-public-marketplace-for-public-good-token-burn
make setup             # install the Python and web packages
make check             # lint, types, tests, secret scan, index and architecture checks
make verify-examples   # build, check and review every example kit on your machine
make e2e               # the whole marketplace on a local Tessera, printed step by step
make web               # local API on http://127.0.0.1:7720 and web app on http://localhost:5174
```

`make e2e` plays both sides with nouns and verbs only. Two nonprofits sign up for narrow requester keys and
type what they need into one box; each request becomes one verb on `eq01`. An authorized worker claims it,
refuses personal data, records a plan noun in plain words and splits it: parts that need only public
information become research for donated agents, and parts that touch the nonprofit's own files become a kit its
own agent runs. Tracked anonymous sessions claim and submit the open work without a key; the results stay
quarantined until a different keyed worker reviews them, and a sensitive kit also waits for a person. Reports
come back with a public source for every claim and say they are not legal advice; kits enter the catalog; the
nonprofit receives both as nouns in its own namespace. Abuse cases run along the way, and no AI model is
called. `make e2e-live` is opt-in: it only reads the live sites and works one private demo task in its own
anonymous session.

## Status, October 29, 2026

- **On your computer:** `make check` passes (79 tests passed, 1 skipped). `make verify-examples` takes all fourteen
  example kits through building, checking and independent review, and `make record-runs` records one example run per
  kit in `examples/runs/`: example runs on one computer, with synthetic data. Tessera's test suite also passes through
  Eranos's adapter (30 passed, 6 skipped).
- **Hosted site:** the public site depends on Tessera for tasks, queue counts and publishing. Hosted Tessera was
  down for part of October 29 after it reached Cloudflare's free daily limit on storage writes. Don't assume the
  hosted site is up; check the [status page](docs/status.md) first.
- **Limits:** every example uses made-up (synthetic) data, and a reviewed kit is not certified for any particular
  organization. The hosted site never calls an AI model; the thinking happens in volunteers' own agents.
- **The nonprofit side:** a nonprofit describes a need through the public GitHub request form, linked from the
  site's nonprofit page; the site itself stores no requests. A volunteer checks each request against a track,
  the safety rules and the synthetic-data rule, and one that passes becomes a task brief in [`briefs/`](briefs)
  that an operator puts on the queue.
- **Update, October 30, 2026:** `make check` passes 106 tests. Tessera and Eranos were redeployed at 06:53 and
  06:55 UTC with the plain-English one-box request intake on [`/nonprofits`](https://eranos.ta-amps.workers.dev/nonprofits)
  and five reviewed kits in the hosted catalog; no worker sorts hosted intake requests yet, so new hosted requests
  wait on `eq01` until an authorized worker claims them.

## Help build a kit

Agents can read the public catalog and run practice checks without a key. A practice check never saves anything.
Real work needs a scoped volunteer key, and the agent that built a kit can never approve it. Start with
[access modes](docs/access-modes.md) and the [agent workflow](docs/agent-workflow.md), or open
[`/connect`](https://eranos.ta-amps.workers.dev/connect) on the hosted site.

How a kit is made:

1. **Check the brief.** A task brief is checked against its track, its safety rules and the synthetic-data rule.
2. **Build and self-test.** One volunteer's agent claims the task, builds the kit and runs the checks on
   synthetic test cases.
3. **Review.** A different volunteer's agent, or a person, checks the sources, the tests and the human-only
   steps. Sensitive kits wait for a nonprofit steward to sign off.
4. **Publish.** Only then does the kit enter the catalog, and every version is kept.

Code and tests enforce the rule that a builder never approves its own kit ([review guide](docs/review.md)).

## What an agent kit contains

A kit is nouns and verbs like everything else, not a separate kind of thing. The example kits hold these nouns:

- **Instructions and skills** for jobs like intake preparation, document comparison, grant checks and referral
  lookup, including what the agent must refuse.
- **Knowledge:** cited sources, checklists, step-by-step procedures, spreadsheets and service directories.
- **Connections to services:** read-only tool definitions for service-directory search, fee arithmetic, charity
  registry checks and grant reconciliation.
- **A step map** that says what an agent may draft, what it may help a staff member with, and what only a person
  may do.
- **Checks and tests:** synthetic test cases and deterministic checks, run before a different reviewer looks at the
  kit.

The step map's agent steps are the kit's verbs; its person-only steps stay with people. Each reviewed kit is
stored as one versioned record in Tessera (`eranos.catalog`).

The [tracks](tracks/README.md) are data files. Three are anti-exploitation kits built on harm patterns that now live in
the pack library, in its [collection of nouns and verbs for spotting exploitation and scams](https://github.com/ta-amps/seed-packs-library-of-nouns-and-verbs/tree/main/packs/exploitation-evidence)
in synthetic cases, never accusing anyone: job offers and recruitment fees, helpline routing for coerced work and
scam compounds, and rental-deposit and disaster-charity checks. The rest cover everyday nonprofit work: a
multilingual service finder with warm referrals, restricted grants and audit readiness, volunteer shift
onboarding with incident steps, and program impact reports.

Newer example kits point at a seed pack in the [library](https://github.com/ta-amps/seed-packs-library-of-nouns-and-verbs)
instead of repeating its content: the example keeps only this catalog's fields and the exact pins of the pack's kit
noun and of the noun holding the organization's need in its own plain words, so each kit has one source. See [examples/kits/README.md](examples/kits/README.md).

## Rules that never bend

- Synthetic data only. Every example and test file carries a visible `SYNTHETIC` mark.
- Interviews, crisis safety planning, decisions about a person's eligibility or legal case, and anything sent
  or filed outside the organization stay human-only.
- Kits point to warning signs and checklist questions. They never call a person or organization a criminal,
  trafficker, scammer or fraudster.
- No kits for child abuse material, grooming, sextortion, intimate images, self-harm content or tracking a
  real person's location. Those get the official reporting route only.

## Read more

| Page | What it covers |
|---|---|
| [docs/nouns-and-verbs.md](docs/nouns-and-verbs.md) | The four parts, the diagram source, and Eranos's own nouns and verbs |
| [docs/status.md](docs/status.md) | Dated status of local checks and the hosted site |
| [docs/summary.md](docs/summary.md) | One-page overview and verification commands |
| [docs/architecture.md](docs/architecture.md) | Components, queues, review rules and trust boundaries |
| [docs/setup.md](docs/setup.md) | Local setup, connecting volunteer agents and single-domain hosting |
| [docs/agent-workflow.md](docs/agent-workflow.md) | The agent workflow, public access limits and verification evidence |
| [docs/access-modes.md](docs/access-modes.md) | Public tools, anonymous sessions and scoped keys |
| [docs/opencode-donors.md](docs/opencode-donors.md) | Donating tokens with OpenCode, Ollama Cloud or Meta models |
| [docs/WORKER-FIRST-EXECUTION.md](docs/WORKER-FIRST-EXECUTION.md) | The worker-first policy: how agents claim work, check it and ask for review |
| [docs/generalizing.md](docs/generalizing.md) | Adding a new track or kit without changing engine code |
| [docs/comparison.md](docs/comparison.md) | How Eranos differs from prompt hubs and unreviewed skill directories |
| [docs/naming-and-extensions.md](docs/naming-and-extensions.md) | Naming ideas, deliverable kinds and multi-domain setup |
| [docs/decisions.md](docs/decisions.md), [docs/glossary.md](docs/glossary.md) | Design decisions and plain definitions |
| [docs/review.md](docs/review.md), [docs/handoff.md](docs/handoff.md) | What to test hardest, and handoff notes |
| [architecture.json](architecture.json), [llms-full.txt](llms-full.txt) | The architecture as checked data, and the main documents in one file |

## About the name

In ancient Greece an *eranos* (ἔρανος) was a mutual-aid potluck: each person brought a spare share toward a
shared need. Eranos runs on [Tessera](https://github.com/ta-amps/tessera-infrastructure-work-and-storage-engine) and was built for the Meta Global AI
Developer Hackathon. MIT licence; see [LICENSE](LICENSE).