Swarm At

io.github.Mediaeater由 GitHub 帳號 Mediaeater 發布v0.7.1更新於 Oct 24, 2026

Settlement protocol for AI agent swarms — hash-chained ledger, trust, 48 blueprints, 18 tools

已驗證SSE可網頁執行開發者工具安全與監控AI 與機器學習

概覽

AI 產生的概覽

把 AI 代理的動作結算進雜湊鏈公開帳本，並提供信任查詢、回執、藍圖、額度與著作權記錄。

可以做什麼
  • 將動作與提案結算進雜湊鏈帳本，支援單筆或批次，並可在執行前先結算以「守護」動作
  • 依任務或雜湊查詢帳本狀態、結算狀態與公開回執
  • 查詢與瀏覽代理信任等級、檔案與聲譽，並註冊新的代理身分
  • 瀏覽、檢視並派生工作流程藍圖，接著執行與追蹤工作流程步驟與分子
  • 管理額度、Webhook 訂閱，以及著作權與寫作工作階段的溯源記錄
適用情境
當你希望代理的產出留下可驗證、防竄改的稽核記錄，或需要查詢某個代理的信任等級、依雜湊查找結算回執時適用。若想執行帶角色分工與額度成本的預先驗證工作流程藍圖，也可考慮。
執行需求
可作為 api.swarm.at 的遠端端點使用，也可作為本機 Python 程序執行（swarm-at-sdk；獨立驗證腳本需 Python 3.8+）。結算、額度與 Webhook 等需驗證的端點需要 SWARM_API_KEY 中的 API 金鑰；公開帳本、回執、信任與徽章查詢無需驗證。可用 SWARM_API_URL 覆寫基礎位址。存取託管 API 需要網路。
安裝前請注意
結算、加值額度、註冊 Webhook 與註冊代理都會寫入資料，並可能消耗額度（一次結算一個額度；新代理據稱獲贈 100 個）。SWARM_API_KEY 是金鑰，不應外洩。README 說明公開帳本中的條目是合成的種子資料，並非真實代理互動記錄，不要當作實際活動證據。Webhook 會把結算事件傳送到你註冊的位址。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 Swarm At，將其新增到工作區。
  2. 為需要使用其工具的對話啟用該服務。

Web executable，透過 SSE。 遠端服務在工作區中設定後即可從網頁執行環境執行。

交給代理程式安裝

把這段提示詞貼上到 Claude Code、Codex、Cursor 等代理程式，或 SourceWeft 對話中。代理程式會讀取該伺服器的安裝指南，先向你展示發布者、端點和登入方式，經你同意後再把它新增到你的客戶端。金鑰不會寫進任何命令。

閱讀 https://sourceweft.com/mcp/io.github.Mediaeater%2Fswarm-at/install.md，並按其中說明把這個 MCP 伺服器接入我的客戶端。修改任何設定之前，先告訴我它的發布者、端點和登入方式，等我確認。不要把 API 金鑰寫進命令或我會提交的檔案；問我金鑰應從哪裡取得。

檢視代理程式所依據的安裝指南

其他 MCP 客戶端

把它新增到你客戶端的 mcpServers 設定中。

{
  "mcpServers": {
    "swarm-at": {
      "type": "sse",
      "url": "https://api.swarm.at/"
    }
  }
}

README

swarm.at Public Ledger

Hash-chained public record of verified agent settlements. Append-only, tamper-evident, auditable by anyone.

Every entry in ledger.jsonl is a SHA-256 hash-chained settlement that passed the full verification pipeline: integrity check, confidence threshold, and shadow audit. Modifying any entry breaks all subsequent hashes.

Ledger Format

Each line in ledger.jsonl is a JSON object:

json
{  "timestamp": 1770480398.05,  "task_id": "fingerprint-pg84",  "parent_hash": "000000...000000",  "payload": { "type": "text-fingerprint", "title": "Frankenstein", "..." : "..." },  "current_hash": "453eaa...178b287"}
  • parent_hash links to the previous entry's current_hash (genesis = 0 x 64)
  • current_hash = SHA-256 of the entry with current_hash set to ""
  • The chain is tamper-evident: modifying any entry breaks all subsequent hashes

Canonical Hash Algorithm

To verify any entry:

  1. Parse the JSON line into a dict
  2. Set current_hash to "" (empty string)
  3. Serialize with json.dumps(entry, sort_keys=True).encode() (UTF-8)
  4. Compute hashlib.sha256(serialized).hexdigest()
  5. Compare against the stored current_hash

The genesis (first entry's parent_hash) is always "0" * 64 (64 zero characters).

Data Provenance

Entries in this ledger are synthetic seed data generated from public domain sources (Project Gutenberg texts, scientific constants, geographic facts). They demonstrate the settlement protocol's hash-chaining and verification pipeline. They are not records of real agent interactions.

Verification

Three ways to verify the chain:

Standalone (zero dependencies, Python 3.8+):

bash
python verify.py# OK: 1107 entries, chain intact

SDK:

python
from swarm_at.settler import Ledgerledger = Ledger(path="ledger.jsonl")print(ledger.verify_chain())  # True

API:

bash
curl https://api.swarm.at/public/ledger/verify# {"intact": true, "entry_count": 1107}

Settlement Receipts

Every settled entry produces a receipt that any agent can look up by hash. No authentication required.

bash
curl https://api.swarm.at/public/receipts/{hash}# {"status": "SETTLED", "hash": "...", "task_id": "...", "timestamp": ..., "parent_hash": "..."}

Receipts let third parties verify that a specific settlement happened, when it happened, and where it sits in the chain.

Trust Verification

Check an agent's trust level without authentication:

bash
# Does this agent meet a trust threshold?curl "https://api.swarm.at/public/verify-trust?agent_id=X&min_trust=trusted"# {"agent_id": "X", "meets_requirement": true, "trust_level": "trusted", "reputation_score": 0.95}
# How many agents at each trust level?curl https://api.swarm.at/public/trust-summary# {"total_agents": 5, "by_trust_level": {"untrusted": 0, "provisional": 0, "trusted": 4, "senior": 1}}

Trust Badges

Embeddable SVG badges show an agent's trust level at a glance. No authentication required.

https://api.swarm.at/badge/{agent_id}

Colors: untrusted (red), provisional (yellow), trusted (green), senior (blue).

MCP Server

29 tools available via the Model Context Protocol:

bash
mcp add swarm-at -- python -m swarm_at.mcp
ToolDescription
settle_actionValidate an action against institutional rules
check_settlementQuery ledger status for a task or hash
ledger_statusCurrent chain state (latest hash, entry count, integrity)
guard_actionSettle before acting — propose + settle + receipt in one call
list_blueprintsBrowse workflow blueprints by tag
get_blueprintFull blueprint detail with steps and credit cost
get_creditsCheck an agent's credit balance
topup_creditsAdd credits to an agent
fork_blueprintFork a blueprint into an executable workflow
verify_receiptLook up a settlement receipt by hash
check_trustCheck if an agent meets a minimum trust threshold
settle_batchSettle multiple proposals in one call
register_agentRegister a new agent identity with trust tracking
get_agentLook up an agent's profile and reputation
list_agentsBrowse agents by trust level and role
whoamiAgent self-introspection (trust, tools, next promotion)
execute_stepExecute a single workflow step in a molecule
list_moleculesList active workflow molecules with progress
get_moleculeGet full molecule state with all beads
register_webhookSubscribe to settlement events
list_webhooksList registered webhook subscriptions
unregister_webhookRemove a webhook subscription
claim_authorshipCreate a verifiable authorship record
verify_authorshipCheck authorship claims by content hash
start_writing_sessionStart tracking authorship provenance
record_writing_eventRecord a creative event in a session
approve_writingRecord final approval of content
get_provenance_reportGenerate a verifiable provenance report
list_writing_sessionsList active authorship sessions

Framework Adapters

Seven adapters settle agent outputs with zero hard dependencies on the framework:

FrameworkAdapterEntry Point
LangGraphSwarmNodeWrapperWraps node functions
CrewAISwarmTaskCallbackTask completion callback
AutoGenSwarmReplyCallbackAgent reply observer
OpenAI AssistantsSwarmRunHandlerRun and step settlement
OpenAI Agents SDKSwarmAgentHookRunner result and tool call settlement
Strands (AWS)SwarmStrandsCallbackTool and agent completion callbacks
HaystackSwarmSettlementComponentPipeline component returning receipts

Install adapters via optional extras:

bash
pip install swarm-at-sdk[langgraph]pip install swarm-at-sdk[openai-agents]pip install swarm-at-sdk[strands]pip install swarm-at-sdk[haystack]

Blueprint Catalog

58 pre-validated blueprints across 9 categories:

CategoryCountExamples
Procurement & Supply Chain5vendor-negotiation, purchase-approval, delivery-confirmation
Software Development5code-review-pipeline, pr-merge-audit, incident-escalation
Finance & Compliance5invoice-matching, kyc-verification, financial-close
Content & Knowledge5research-workflow, fact-checking, translation-verification
Customer Operations5escalation-routing, refund-approval, sla-compliance
Specialty6audit-chain, healthcare-referral, insurance-adjudication
AI & ML Ops9model-validation, training-pipeline, prompt-regression
Security & Compliance9vulnerability-triage, access-review, incident-response
Data Engineering9etl-pipeline, data-quality, schema-migration

Browse: api.swarm.at/public/blueprints

Each blueprint defines 2-4 steps with role assignments (worker, auditor, specialist, orchestrator, validator), dependency chains, and credit costs (2.0-8.0 per settlement).

Settlement Types

42 types covering knowledge verification, agent behaviors, and protocol operations

Knowledge Verification

TypeDescription
text-fingerprintPublic domain text ingestion + hashing
qa-verificationFactual Q&A with multi-agent consensus
fact-extractionStructured entity extraction from text
classificationGenre/topic/tone multi-label tagging
summarizationText condensation with cross-verification
translation-auditCross-language verification
data-validationMath constants + periodic table verification
code-reviewAlgorithm correctness + complexity analysis
sentiment-analysisDimensional sentiment scoring
logical-reasoningSyllogism and formal logic verification
unit-conversionMetric/imperial/scientific unit verification
geo-validationGeographic fact verification
timeline-orderingChronological event ordering
regex-verificationPattern matching correctness
schema-validationData schema conformance checking

Agent Behaviors

TypeDescription
code-generationNew code creation with language + framework metadata
code-editModifications to existing code with diff tracking
code-refactorStructural improvements preserving behavior
bug-fixDefect resolution with root cause analysis
test-authoringTest creation with coverage and assertion tracking
codebase-searchCode search operations with match scoring
web-researchWeb research with source verification
planningTask decomposition and execution planning
debuggingDiagnostic investigation with hypothesis tracking
shell-executionShell command execution with safety classification
file-operationFile system operations with change tracking
git-operationGit operations with ref and diff metadata
dependency-managementPackage and dependency management
agent-handoffTask delegation between agents
consensus-voteMulti-agent consensus participation
task-delegationWork distribution to sub-agents
documentationDocumentation creation and updates
api-integrationExternal API calls with endpoint tracking
deploymentBuild, deploy, and release operations
conversation-turnConversational exchange settlement

Protocol Operations

TypeDescription
blueprint-forkBlueprint forked into executable workflow
guard-actionPre-action settlement check (settle before you act)
trust-checkAgent trust level verification against threshold
credit-topupCredits added to agent balance
receipt-verifySettlement receipt lookup by hash
badge-requestTrust badge SVG generation
adapter-settlementFramework adapter settling agent output

Lexicon

TermDefinition
SettlementA verified, hash-chained record of an agent action. One credit = one settlement.
LedgerAppend-only JSONL file where settlements are recorded. Each entry's hash chains to the previous.
ProposalAn agent's request to settle an action. Contains a header (task ID, parent hash) and payload (data, confidence score).
ReceiptProof that a settlement occurred. Contains hash, task ID, timestamp, and parent hash. Publicly verifiable.
BlueprintA pre-validated workflow template with ordered steps, role assignments, and credit cost. Forkable by any agent.
WorkflowAn executable instance of a blueprint. Created by forking. Each step maps to one settlement.
Trust LevelAgent reputation tier: untrusted, provisional, trusted, senior. Computed via Bayesian credible intervals.
CreditUnit of settlement currency. 1 credit = 1 settlement. New agents get 100 free.
Guard ActionPattern: settle before you act. Propose, settle, get receipt in one call. Raises error on rejection.
Shadow AuditCross-model verification where a second model independently checks the primary's output.
DivergenceWhen a shadow audit disagrees with the primary. Penalizes the agent's trust score.
AdapterFramework integration that settles agent outputs without hard-importing the framework. Uses duck typing.
Agent CardA2A discovery document at /.well-known/agent-card.json describing the protocol's capabilities.
TierSettlement strictness level: sandbox (log-only), staging (writes, no chain enforcement), production (full).

Links

© 2026 Mediaeater. All rights reserved.

來源：README.md，提交 097b9d3

工具

0
SourceWeft 尚未讀取此伺服器的工具。

版本歷史

1
  1. v0.7.1最新Sep 16, 2026

評分

已滿足適用檢查項的 50%

該評分衡量此收錄條目的完整程度，依據註冊中心條目、端點與公開來源計算得出。它不是對伺服器本身的評價，也不表示伺服器安全或目前可用：可連通性是帶日期的檢查結果，評分也不包含安裝量或使用者評價。

  • 近 30 天內可連通必需25 分

    請保持端點能夠回應 MCP 握手；SourceWeft 每 30 天探測一次。

  • 已宣告認證方式必需10 分

    請在註冊中心條目中宣告憑證請求標頭，或保持端點能夠回應探測，以便判定其認證方式。

  • 已列出工具15 分

    請讓伺服器在無憑證時回應 tools/list，或發布可在沙箱中無憑證啟動的軟體包。

  • 描述必需15 分
  • 圖示10 分
  • 文件15 分
  • 授權條款10 分