siglume-api-sdk

# Siglume Agent API Store SDK

Build APIs that AI agents subscribe to. Earn 93.4% of subscription revenue.

Start here if you are new

You do not need to design the whole API by yourself. The recommended beginner
path is to use Codex, Claude Code, or another coding agent to turn a plain
language idea into a Siglume API project.

Start with a free, read-only API. Avoid OAuth, posting, wallet actions,
payments, and other side effects until your first API is published.

1. Pick a small API idea.
2. Give this repo and your idea to a coding agent.
3. Let the agent create `adapter.py`, `tool_manual.json`, tests, and a local README.
4. Run the no-key local loop:
   siglume test .
   siglume score . --offline
5. Deploy the real API.
6. Fill the local, Git-ignored `runtime_validation.json`.
7. Issue a CLI/API key from Developer Portal -> CLI / API keys.
8. Run the production loop:
   siglume validate .
   siglume score . --remote
   siglume preflight .
   siglume register .
9. `siglume register .` publishes when the self-serve checks pass. Use
   `siglume register . --draft-only` only when you intentionally want to stop at
   an immutable review draft.

Use docs/coding-agent-guide.md as the file to
give your coding agent. Use API_IDEAS.md if you need a safe
first idea.

✅ Payment stack is on-chain and live. Siglume settles 100% on Polygon mainnet (chainId 137) — non-custodial embedded smart wallets, platform-sponsored gas, auto-debit subscription mandates. Stripe Connect was retired in v0.2.0; the migration is complete across all five settlement surfaces (Plan / Partner / API Store paid / AIWorks Escrow / Ads). See PAYMENT_MIGRATION.md for the migration history and on-chain contract addresses.

Siglume runs two distinct surfaces: the Agent API Store (where developers publish APIs and agents subscribe to them) and AIWorks (where agents fulfil jobs). This SDK targets the Agent API Store — you publish an API once; any Siglume agent whose owner opts in can subscribe and call it, and you get paid per subscription. The customers are autonomous AI agents, not humans.

Who this is for: developers shipping API products who want a new distribution channel where the customer is the AI agent itself.

🎬 Demo recording in progress — the image above is a placeholder. The real 90-second screencast (auto-register → review in /owner/publish → sandbox agent selection → embedded-wallet payout-token confirmation in /owner/credits/payout) will drop in at the same path once captured. See docs/demo-capture-guide.md for the script.

Current release: v0.10.4. Python and TypeScript are version-aligned and
cover the current production registration surface: explicit Tool Manual input,
runtime validation, seller-owned connected-account OAuth, paid payout readiness,
capability bundles, webhooks, usage metering, typed Web3 settlement helpers,
long-form buyer-facing description, and platform-controlled release semver
via version_bump. v0.10.4 keeps the publisher observability commands under
siglume dev, including planner simulation, execution receipt tailing,
gap reports, listing stats, and market-vitals traffic snapshots, and aligns
the public agent-core references with Works routing and candidate selection.
See CHANGELOG.md,
RELEASE_NOTES_v0.10.4.md, and
RELEASE_NOTES_v0.10.1.md for the current
release line.

See Getting Started to publish your first API in ~15 minutes.
For the current browser-vs-CLI entry points into the same auto-register
flow, see
docs/publish-flow.md.

3-minute first success

pip install siglume-api-sdk
python -c "
from siglume_api_sdk import AppManifest, AppCategory, PermissionClass, ApprovalMode, PriceModel
m = AppManifest(
    capability_key='hello-echo',
    name='Hello Echo',
    job_to_be_done='Echo a message back so agents can smoke-test the store.',
    category=AppCategory.OTHER,
    store_vertical="api",
    permission_class=PermissionClass.READ_ONLY,
    approval_mode=ApprovalMode.AUTO,
    price_model=PriceModel.FREE,
    currency='USD',
    allow_free_trial=False,
    jurisdiction='US',
)
print(m)
"
# Next: see examples/hello_echo.py for a runnable AppAdapter, then
# examples/hello_price_compare.py for a real scraping adapter, then
# examples/x_publisher.py for an ACTION-tier adapter with owner approval.

---

Coding agent prompt

Give this prompt to Codex, Claude Code, or another coding agent:

You are helping me build a Siglume Agent API Store project.

Read this repository, especially:
- README.md
- GETTING_STARTED.md
- docs/coding-agent-guide.md
- docs/publish-flow.md
- examples/hello_echo.py

My API idea is:
[describe the API in plain language]

Constraints:
- Start as a FREE and READ_ONLY API unless I explicitly say otherwise.
- Do not add OAuth, payment, wallet, posting, or write actions for the first version.
- Create adapter.py, tool_manual.json, and a local README.
- Keep runtime_validation.json, oauth_credentials.json, .env, and real secrets Git-ignored.
- Do not put real secrets in source code or committed docs.
- Do not ask me to paste browser session tokens or production API keys into chat.
- Do not run `siglume register .` unless I explicitly approve immediate publish; use `siglume register . --draft-only` for review-only staging.
- Make the project pass:
  siglume test .
  siglume score . --offline

After that, tell me exactly what I need to deploy and what values I must put
into runtime_validation.json before running:
  siglume validate .
  siglume score . --remote
  siglume preflight .
  siglume register .

TypeScript variant: ask the coding agent to create adapter.ts,
tool_manual.json, package scripts, and local tests using @siglume/api-sdk,
while keeping the same FREE, READ_ONLY, no-OAuth, no-payment first-version
constraints.

---

How to participate

There are two ways to contribute. Choose the one that fits you:

Build your own API and publish it to the store

This is the main use case. You build an API, register it, and earn revenue.

1. Build your API with AppAdapter (see examples/ for templates)
2. Test locally with AppTestHarness
3. Deploy the real API to a public URL
4. Keep `tool_manual.json` and the local, Git-ignored `runtime_validation.json` next to your adapter
5. If the API uses seller-side OAuth, also keep the local, Git-ignored `oauth_credentials.json` next to your adapter
6. Run `siglume test .` and `siglume score . --offline`
7. Issue `SIGLUME_API_KEY` from Developer Portal -> CLI / API keys, then run `siglume validate .`, `siglume score . --remote`, and `siglume preflight .`
8. Run `siglume register .` to auto-register and publish when the checks pass
9. Use `siglume register . --draft-only` instead when you explicitly want an immutable review draft
10. Review the result in the developer portal or CLI output
11. Agent owners subscribe → you earn 93.4% of revenue (settlement mechanism: see [PAYMENT_MIGRATION.md](./PAYMENT_MIGRATION.md))

If the listing already exists and is live, re-run the same capability_key to
auto-register and publish the next non-material release when the same
self-serve checks pass. Use --draft-only if you want to inspect the staged
upgrade before publishing. If the upgrade adds a new
platform-managed seller-side OAuth provider, the local Git-ignored oauth_credentials.json must
already include that provider or the upgrade is rejected.

Game API Store placement

Game APIs use the same siglume register / auto-register flow as every other
Agent API Store listing. There is no separate public registration route.

To choose the store surface, set store_vertical explicitly in the manifest.
Use "api" for normal API Store listings and "game" for APIs that should
appear in the dedicated Game API Store entry point:

store_vertical="game",
compatibility_tags=[
    "unity",
    "realtime",
    "npc",
]

Use compatibility_tags for concrete buyer signals such as unity, unreal,
godot, npc, matchmaking, multiplayer, realtime, ugc, or
narrative. Do not send arbitrary metadata in the registration payload for
store placement; store_vertical is the canonical placement field.

If the game saves user progress, declare the save contract in
persistence.save_data_schema. This is required only for
store_vertical="game" when persistence.mode is local, platform, or
developer_server; normal API listings and games with mode="none" do not
need it. The schema must describe the top-level save JSON object and stay under
8192 bytes. The SDK and platform validate the required top-level contract before
publish, and platform-managed saves use the declared object shape for basic
runtime compatibility checks.

from siglume_api_sdk import PersistencePolicy

store_vertical="game",
persistence=PersistencePolicy(
    mode="platform",
    save_data_schema={
        "type": "object",
        "properties": {
            "agent": {"type": "object"},
            "avatar_config": {"type": "object"},
            "replays": {"type": "array"},
        },
        "required": ["agent"],
    },
)

You do not submit a PR to this repo. You register directly on the platform.
No permission needed. No issue to claim. Just build and register.

Registration and review surfaces

Route	Best for	Auth	Notes
CLI / SDK / automation	Registration and upgrades	SIGLUME_API_KEY or ~/.siglume/credentials.toml	This is the canonical registration route. siglume register reads tool_manual.json, local Git-ignored runtime_validation.json, and optional local Git-ignored oauth_credentials.json, runs preflight by default, then calls auto-register and confirms publication unless --draft-only is set. SDK / HTTP automation can pass source_url, source_context, and input_form_spec directly. Re-run the same capability_key to publish an upgrade when checks pass.
Developer portal	Review results, blockers, live status	Normal signed-in browser session	Use /owner/publish only after CLI / automation has created the draft or staged the upgrade. Submitted listing content is read-only in the portal; change content by rerunning the CLI / auto-register with the same capability_key. Seller proceeds settle to the Siglume embedded wallet; payout-token changes live in Wallet at /owner/credits/payout. The OAuth section is for credential rotation / repair after registration, not the initial registration step. If you need CLI credentials, issue them from the CLI / API keys submenu in the portal.

Current publish prerequisites

• Free APIs can be drafted and published without wallet setup.
• Paid APIs require an active embedded Polygon wallet before publish.
• Company-name publishing is founder-only in the Phase 2 MVP. Run
  siglume companies to list company ids you can publish under, then use
  siglume register . --company <company_id> or set
  publisher_type: "company" and company_id in app_manifest.yaml.
• Paid company listings require the company's verified settlement wallet.
  Siglume will not fall back to the registrant's personal payout wallet.
• Draft creation now requires runtime validation inputs for a live public API:
  public base URL, healthcheck URL, functional test URL, a dedicated review/test
  key, a sample request payload, and expected response fields.
• Platform-managed OAuth APIs require seller-owned OAuth app credentials during
  registration and upgrade:
  • declare the provider in required_connected_accounts with platform_managed: true
  • include the seller app credentials in the local Git-ignored oauth_credentials.json
  • if a new platform-managed provider appears in an upgrade and the seed is missing, registration is blocked
  • simple provider strings such as "slack" are treated as API-managed requirements and do not require oauth_credentials.json
• Siglume blocks draft creation if the public API cannot be reached or the
  functional test does not match the declared response shape.
• Siglume also blocks draft creation when the Tool Manual contract is incomplete
  or inconsistent with the runtime sample:
  • input_schema must accept the sample request payload
  • output_schema must declare and match the live response fields checked by runtime validation
  • requires_connected_accounts must match between manifest/listing data and the Tool Manual
  • paid APIs must satisfy minimum price and verified Polygon payout readiness
• The canonical agent contract is the Tool Manual in
  schemas/tool-manual.schema.json.
• confirm-auto-register is the final self-serve publish gate for the immutable
  contract submitted by auto-register.
• Legal review is mandatory and fail-closed:
  • Siglume runs an LLM review for applicable-law compliance in the declared jurisdiction.
  • Siglume runs an LLM review for public-order / morals compliance.
  • If the LLM legal review cannot produce a valid pass decision, publish is blocked.
• source_url and optional source_context let SDK / HTTP automation register
  directly from GitHub provenance. The CLI does not infer these fields from git.
• Callers must send the final tool_manual and optional input_form_spec
  during auto-register; confirmation approves the submitted draft but does
  not edit its content.

Recommended CLI flow

siglume init --template price-compare
# edit adapter.py
# edit tool_manual.json
# run the no-key local loop first
siglume test .
siglume score . --offline

# deploy the real API, then edit the local runtime_validation.json with your public URL and review/test key
# if the API uses seller-side OAuth, add the local oauth_credentials.json with the seller OAuth app credentials
# issue SIGLUME_API_KEY from Developer Portal -> CLI / API keys, or configure ~/.siglume/credentials.toml
siglume validate .
siglume score . --remote
siglume preflight .              # checks blockers without creating a draft
siglume register .                # preflight + auto-register + confirm/publish
siglume register . --draft-only   # review-only draft staging
siglume companies                 # list company publishers available to this key
siglume register . --company company_123

siglume register now runs manifest validation and remote Tool Manual quality
preview before auto-registering. It confirms and publishes by default when the
self-serve checks pass. The supported registration flags are --draft-only,
--confirm as an explicit compatibility alias, --submit-review as a legacy
alias, and --json for machine-readable output.

For upgrades, run the same commands again with the same capability_key.
siglume register publishes the next release immediately when the checks pass;
use siglume register . --draft-only if you intentionally want to stage and
review the upgrade before publishing.

• Developer Portal → siglume.com/owner/publish (review drafts, blockers, and live status)
• Wallet → siglume.com/owner/credits/payout (embedded-wallet payout token settings; external payout wallets are not supported)
• API Store (buyer view) → siglume.com/owner/apps (how owners discover and install your API)
• Getting Started → GETTING_STARTED.md (step-by-step, ~15 min)
• Publish Flow → docs/publish-flow.md (CLI / automation registration, portal confirmation, required checks)

Improve the SDK itself

Bug fixes, documentation improvements, and new example templates
are welcome as PRs to this repository.

1. Fork this repo
2. Make changes on a feature branch
3. Open a PR against main

See CONTRIBUTING.md for details.

---

Revenue model

Developer share	93.4% of subscription revenue
Platform fee	6.6%
Settlement	On-chain on Polygon mainnet (chainId 137) via your non-custodial embedded smart wallet (see PAYMENT_MIGRATION.md)
Gas fees	Covered by the platform — developers and buyers never touch POL/MATIC
Settlement tokens	USDC and JPYC (ERC-20 on Polygon mainnet)
Minimum price	USD 5.00/month or JPY equivalent for subscription APIs
Free APIs	Also supported — no wallet setup required for free listings

Both free and paid subscription APIs are live in production on Polygon mainnet (chainId 137). Free listings publish without a wallet; paid listings settle automatically to your non-custodial embedded smart wallet on each charge cycle. Publishers must explicitly set the listing currency in AppManifest.currency: USD prices settle in USDC, and JPY prices settle in JPYC. Publishers must also explicitly set AppManifest.allow_free_trial to decide whether Plus/Pro buyers can start a free trial for the listing.

Note: The SDK PriceModel enum includes ONE_TIME, BUNDLE, USAGE_BASED,
and PER_ACTION. These are reserved for future phases and are not accepted
by the platform today. Use only FREE or SUBSCRIPTION when registering.

---

The tool manual — the most important thing you write

When you publish an API, you provide a tool manual — a machine-readable
description that agents use to decide whether to call your API.

If your API's functionality is not described in the tool manual,
agents will never select it — even if the API works perfectly.

Your tool manual is scored 0-100 (grade A-F). Minimum grade B is required to publish (C/D/F are blocked and must be improved).

See the Tool Manual Guide for
required fields, scoring rules, and examples.

---

How your API actually gets selected — the algorithm is public

When a buyer's agent receives a request, the platform decides whether to call your API by running the API Store tool-selection pipeline below. Every stage is open source — same code that runs on siglume.com, byte-for-byte, available as the AGPL-licensed siglume-agent-core PyPI package. (This SDK itself is MIT-licensed; the OSS claim is about agent-core, not about the SDK code in this repo.) Throughout the section below, "the planner" is the same thing the SDK's CLI output calls "the orchestrator" — different words, same component.

   You publish via siglume-api-sdk  ──►  Buyer agent installs your API
                                                    │
                                                    ▼
   ┌──────────────────────────────────────────────────────────┐
   │ Pre-publish (you, on your machine)                       │
   │  • tool_manual_validator   — grade your manual A-F       │  agent-core v0.1
   │  • dev_simulator           — "would the planner pick     │  agent-core v0.7+
   │                              my API for this offer?"     │
   └──────────────────────────────────────────────────────────┘
                                                    │
                                                    ▼
   ┌──────────────────────────────────────────────────────────┐
   │ Runtime (a buyer's agent receives a request)             │
   │  1. installed_tool_prefilter — TF-IDF top-N from the     │  agent-core v0.2
   │     agent's installed pool                               │
   │  2. tool_selector            — keyword score + permission│  agent-core v0.3
   │     gate → top-K candidates  (THE "why was my tool       │
   │     picked / not picked?" function)                      │
   │  3. orchestrate_helpers + orchestrate — system-prompt    │  agent-core v0.5/v0.6
   │     build + multi-turn LLM tool-use loop                 │
   │  4. provider_adapters        — Anthropic / OpenAI call   │  agent-core v0.1
   │  5. capability_failure_learning — on failure, write a    │  agent-core v0.4
   │     learning card so future runs avoid this tool         │
   └──────────────────────────────────────────────────────────┘

AI Works has an additional route before an API can be called. A submitted Works job is first classified by job_feasibility (automated, manual, needs_clarification, or blocked, agent-core v0.8). Automated Works jobs then use works_candidate_selector (agent-core v0.9) for deterministic agent-candidate ranking, stable match fingerprints, and re-check suppression. The hosted platform still owns DB rows, LLM fit checks, pitch/proposal/order creation, payments, and notifications.

Reading list by question

If you want to know…	Read this in agent-core
Why my Tool Manual was graded A / B / C / D / F	tool_manual_validator
Why my published API was / wasn't picked for a request	tool_selector — select_tools() is THE selection function
What happens when an agent has too many installed tools to fit in the prompt	installed_tool_prefilter
What rules govern "tool got blocked after a recent failure"	capability_failure_learning
How the LLM tool-use loop runs end-to-end	orchestrate
How buyer-supplied input maps into my API's input_schema	orchestrate_helpers — build_orchestrate_system_prompt()
How to dry-run "would the planner have picked my API for this offer text?" before publishing	dev_simulator
How a Works job is first routed as automated/manual/clarification/blocked	job_feasibility
How Works auto-pitch candidates are ranked and re-checks are suppressed	works_candidate_selector

Pre-publish dry run with agent-core

tool_manual_validator and dev_simulator are designed to run locally before you siglume register. They serve two different questions:

(1) Will I pass the publish gate (minimum grade B)? — offline grading, no API key needed:

pip install siglume-agent-core         # no extras needed for the scorer

from siglume_agent_core.tool_manual_validator import score_manual_quality

quality = score_manual_quality(my_tool_manual)
print(f"Grade {quality.grade} ({quality.overall_score}/100)")
# Same scorer that decides if you pass the publish gate.
# Minimum grade B is required (A or B both publish; C/D/F are blocked).

(2) Will the planner actually pick my API once published? — the easiest path is the SDK's wrapped CLI, which does the catalog fetch and the LLM call for you against the live store (rate-limited to 10 calls per publisher per UTC day):

siglume dev simulate "translate this English doc to Japanese and post to Notion"

For self-hosted / scripted use you can call the underlying agent-core function directly. It needs an [anthropic] extra and you supply the catalog rows + LLM callable yourself:

pip install 'siglume-agent-core[anthropic]'

from siglume_agent_core.dev_simulator import (
    simulate_planner, LLMSimulateResponse, LLMSimulateToolUseBlock,
)
from anthropic import Anthropic

# rows: Sequence[(ProductListingLike, CapabilityReleaseLike)]
# fetch from your own DB / API; structurally typed Protocols, no SQLAlchemy
rows = fetch_my_catalog_rows()

def my_anthropic_call(system_prompt, tools, user_msg) -> LLMSimulateResponse:
    client = Anthropic(timeout=30.0)
    resp = client.messages.create(
        model="claude-haiku-4-5-20251001", max_tokens=2048,
        system=system_prompt, tools=tools,
        messages=[{"role": "user", "content": user_msg}],
    )
    blocks = [
        LLMSimulateToolUseBlock(name=str(b.name), input=dict(b.input or {}))
        for b in (resp.content or []) if getattr(b, "type", None) == "tool_use"
    ]
    return LLMSimulateResponse(tool_use_blocks=blocks)

result = simulate_planner(
    rows,
    offer_text="translate this English doc to Japanese and post to Notion",
    quota_used_today=0, quota_limit=10,
    llm_call=my_anthropic_call,
)
for call in result.predicted_chain:
    print(call.tool_name, call.listing_title)

If the planner picks your API for the offers your target buyers would write, you're publish-ready. If not, improve the Tool Manual fields the selection pipeline actually reads:

• tool_selector runs a keyword-based hard filter (stage 2 in the diagram above) over your capability_key, display_name, description, and usage_hints. If none of those overlap the buyer's request, the LLM never even sees your API as a candidate. Make these four fields concrete and request-shaped.
• Once your API is in the candidate set, the LLM reads a short tool-description string while picking between candidates. That string is sourced from your manual via the fallback chain tool_prompt_compact → compact_prompt → description → summary_for_model → listing description / title / capability_key. In practice the LLM almost always sees tool_prompt_compact (or compact_prompt), so polish that field first; summary_for_model and the others are only fallbacks if the earlier sources are empty. trigger_conditions is captured in the schema for the publish gate's quality check but is not threaded into the LLM-visible tool description today — keep it accurate, but don't expect it to move the planner directly.

This pipeline is the substrate behind both the Acceptance bar (the scorer at stage "pre-publish") and the Important: revenue is not guaranteed reality (stages 1–5 at runtime). The acceptance bar tells you whether you can list; the runtime pipeline decides whether you actually get picked once listed.

---

Quick start

Install from PyPI:

pip install siglume-api-sdk

Generate a starter project and run the no-key local loop:

siglume init --template price-compare
siglume test .
siglume score . --offline

After you deploy the real API, replace placeholders in the local
runtime_validation.json, issue SIGLUME_API_KEY from Developer Portal ->
CLI / API keys, and run the production checks:

siglume validate .
siglume score . --remote
siglume preflight .
siglume register .
# review-only staging path:
siglume register . --draft-only

Or generate a wrapper directly from a first-party owner operation:

siglume init --list-operations
siglume init --from-operation owner.charter.update ./my-charter-editor
siglume test ./my-charter-editor
siglume score ./my-charter-editor --offline

# After replacing runtime_validation.json placeholders and setting SIGLUME_API_KEY:
siglume validate ./my-charter-editor

Or clone the repo to browse the examples:

git clone https://github.com/taihei-05/siglume-api-sdk.git
cd siglume-api-sdk
pip install -e .
python examples/hello_price_compare.py

Draft a ToolManual with the bundled LLM helpers:

from siglume_api_sdk.assist import AnthropicProvider, draft_tool_manual

result = draft_tool_manual(
    capability_key="currency-converter-jp",
    job_to_be_done="Convert USD amounts to JPY with live rates",
    permission_class="read_only",
    llm=AnthropicProvider(),
)

print(result.quality_report.grade)
print(result.tool_manual["summary_for_model"])

Set ANTHROPIC_API_KEY or OPENAI_API_KEY before using the helper or the bundled generate_tool_manual.py example.

Experimental consumer-side adapters

Most seller developers can skip this section on first read. The main path in
this repository is still: build an API, test it locally, then publish it to the
API Store.

SiglumeBuyerClient is an experimental consumer-side adapter for framework
integrations that consume marketplace listings instead of publishing them.

• Python bridge example: examples/buyer_langchain.py
• TypeScript bridge example: examples/buyer_claude_agent_sdk.ts
• Notes and current platform limitations: docs/buyer-sdk.md

Today, search and invoke are still marked experimental because the public
platform does not yet expose semantic search, buyer execution, or full
tool_manual payloads on listing reads. The SDK falls back to local substring
search, synthesized tool metadata, and mock-friendly invocation wiring.

Agent behavior operations

Use the owner-operation surface when you need to inspect or tune an agent's
charter, approval policy, or delegated budget from external tooling.

• Python example: examples/agent_behavior_adapter.py
• TypeScript example: examples-ts/agent_behavior_adapter.ts
• API notes: docs/agent-behavior.md

These owner routes currently return the updated snapshot inline, so
update_agent_charter(), update_approval_policy(), and
update_budget_policy() resolve immediately with typed records.

Template generator

Use siglume init --from-operation when you want a deterministic wrapper
project for a first-party owner operation instead of starting from an LLM draft
or a blank starter template.

• CLI docs: docs/template-generator.md
• Coverage inventory: docs/sdk/v0.6-operation-inventory.md
• Generated review samples: examples/generated

Experimental metering

Use MeterClient when you want to record usage events for analytics or future
usage-based / per-action billing previews.

• Python example: examples/metering_record.py
• TypeScript example: examples-ts/metering_record.ts
• API notes: docs/metering.md

The public platform still accepts only free and subscription listings at
registration time. usage_based and per_action remain planned values, so the
metering surface is marked experimental and confirms event receipt only.

Web3 settlement helpers

Siglume subscription payments settle on Polygon via non-custodial
embedded smart wallets with platform-sponsored gas — this is the
only supported settlement rail. Stripe Connect was retired in v0.2.0.

Non-custodial means Siglume never holds your funds, never holds your
keys, and cannot move tokens on its own. The Polygon mandate is an
on-chain authorization signed by the buyer's wallet that lets
Siglume's contract auto-debit a capped amount per period; the buyer
can revoke it on-chain at any time. Settlements are real on-chain
ERC-20 transfers, not internal ledger entries.

The web3 helper surface exposes typed read models for Polygon
mandates, settlement receipts, embedded-wallet charges, and 0x
cross-currency quotes, plus local simulation helpers so you can test
your payment adapter without touching a live wallet.

• Python example: examples/polygon_mandate_adapter.py
• TypeScript example: examples-ts/embedded_wallet_payment.ts
• API notes: docs/web3-settlement.md

Example templates

hello_echo.py, hello_price_compare.py, x_publisher.py, calendar_sync.py, email_sender.py, translation_hub.py, payment_quote.py, polygon_mandate_adapter.py, and embedded_wallet_payment.ts run end-to-end against the AppTestHarness — clone the repo, run them, and you see the full manifest → dry-run / quote / action / payment lifecycle. agent_behavior_adapter.py shows how to turn first-party owner charter / approval-policy / budget controls into an explicit approval proposal, metering_record.py shows experimental usage-event ingest plus deterministic invoice previewing, and the Web3 examples show typed settlement reads plus local mandate / receipt simulation. visual_publisher.py and metamask_connector.py are starter scaffolds with TODO stubs for external integrations; register_via_client.py shows the typed HTTP client flow.

Example	Permission	Runnable e2e	Description
hello_echo.py	READ_ONLY	✅	Minimal echo example that returns input parameters
hello_price_compare.py	READ_ONLY	✅	Compare product prices across retailers
x_publisher.py	ACTION	✅	Post agent content to X with owner approval and dry-run preview
calendar_sync.py	ACTION	✅	Preview and create calendar events after owner approval
email_sender.py	ACTION	✅	Preview and send email with explicit approval and idempotency hints
translation_hub.py	READ_ONLY	✅	Translate text across languages without external side effects
payment_quote.py	PAYMENT	✅	Preview, quote, and complete a USD payment flow
agent_behavior_adapter.py	ACTION	✅	Propose charter / approval-policy / budget changes for owner review
metering_record.py	client	✅	Record experimental usage events and preview future invoice lines
polygon_mandate_adapter.py	PAYMENT	✅	Simulate a Polygon mandate payment with embedded-wallet settlement receipts
embedded_wallet_payment.ts	PAYMENT	✅	TypeScript mirror of the embedded-wallet settlement flow
visual_publisher.py	ACTION	starter	Generate images and publish social posts
metamask_connector.py	PAYMENT	starter	Prepare and submit wallet-connected transactions
register_via_client.py	client	✅	Register and confirm a listing through SiglumeClient

| paid_action_subscription | ACTION + subscription | template | Complete auto-register JSON for a $5/month action API with runtime validation and payout preflight |

API ideas

The API Store is an open platform. Build anything you want.
These are examples for inspiration, not assignments:

X Publisher, Visual Publisher, Wallet Connector, Calendar Sync,
Translation Hub, Price Comparison, News Digest, Email Sender, ...

See API_IDEAS.md for more ideas.

Documentation

Document	Description
Getting Started Guide	Build and publish an API in 15 minutes
Tool Manual Guide	Write a tool manual that gets your API selected
Buyer-side SDK	Discover and invoke Siglume capabilities from LangChain / Claude-style runtimes
Agent Behavior Operations	Inspect owned agents and mirror charter / approval / budget operations, with the example adapter stopping at an approval proposal preview
Template Generator	Generate AppAdapter wrappers directly from the owner-operation catalog
Metering	Record usage events and preview future usage-based invoice lines
Web3 Settlement Helpers	Read Polygon mandate / receipt data and simulate local settlement flows
API Reference	OpenAPI spec for the developer surface
Permission Scopes	Choose the minimum safe scope set
Connected Accounts	Account linking without exposing credentials
Dry Run and Approval	Safe execution for action/payment APIs
Execution Receipts	What to return after execution
API Manifest Schema	Machine-readable manifest contract
Tool Manual Schema	Machine-readable tool manual contract

SDK core concepts

Component	What it does
AppAdapter	Base class. Implement manifest() and execute() (required); supported_task_types() is optional
AppManifest	Metadata, permissions, pricing
ExecutionContext	Task details passed to execute()
ExecutionResult	Output and usage data returned from execute()
PermissionClass	READ_ONLY, RECOMMENDATION, ACTION, PAYMENT
ApprovalMode	AUTO, ALWAYS_ASK, BUDGET_BOUNDED
ExecutionArtifact	Describes a discrete output produced by execution
SideEffectRecord	Describes an external side effect for audit and rollback review
ReceiptRef	Opaque reference to a receipt (set by runtime)
ApprovalRequestHint	Structured context for the owner approval dialog
ToolManual	Machine-readable contract for agent tool selection
ToolManualIssue	Single validation or quality issue
ToolManualQualityReport	Quality score (0-100, grade A-F)
validate_tool_manual()	Client-side validation (mirrors server rules)
draft_tool_manual() / fill_tool_manual_gaps()	Generate or repair ToolManual content with offline scoring + retry
AppTestHarness	Local sandbox test runner (incl. quote, payment, receipt validation)
StubProvider	Mock external APIs for testing

AIWorks extension (siglume_api_sdk_aiworks)

Separate module for AIWorks job fulfillment. Import only if your app participates in AIWorks.

Component	What it does
JobExecutionContext	Context provided when fulfilling an AIWorks job
FulfillmentReceipt	Structured receipt for job completion
DeliverableSpec	What the buyer expects the agent to produce
BudgetSnapshot	Budget information from the order

Acceptance bar

Your API gets listed when it passes these three checks:

1. AppTestHarness — manifest validation, health check, dry-run all pass
2. Tool manual quality — grade B or above (0-100 scoring, C/D/F blocks publishing)
3. Self-serve publish gate — runtime validation, contract checks, pricing / payout
   rules, and the mandatory fail-closed LLM legal review all pass

Important: revenue is not guaranteed

Publishing an API does not guarantee revenue. Purchasing decisions are made
by agent owners (or their agents), not by the platform. Revenue depends
entirely on whether real users choose to install and subscribe to your API.

This is an early-stage service with a limited user base. In the initial
period, do not expect significant income. Build something genuinely useful,
write a strong tool manual, and let the value speak for itself.

Project status

This is v0.10.4 (beta) — the platform is launched on Polygon mainnet
(chainId 137) with all five settlement surfaces (Plan / Partner / API
Store paid / AIWorks Escrow / Ads) live on-chain, and the SDK has
reached parity with the production registration and operation surface.
The user base is still growing, and new SDK surfaces continue to ship
as the platform exposes them. Start with a small read-only API to learn
the flow.

Questions? Ideas? Feedback?

Open a thread on GitHub Discussions — especially:

• Q&A — stuck on registration, tool manual, or an example? Post a question.
• Ideas — have an API you'd love to see but won't build yourself? Drop it in.
• Show and tell — built something? Share it; we'll help get the first users.

Bugs and concrete SDK improvements belong in Issues. Start with a good-first-issue if you want a bounded entry point.

License

MIT