에이전트 트랜잭션을 위한 레일 중립적 증거 레이어를 제공하며, 검증 가능한 컴퓨팅, 서명된 영수증 및 서비스 게시 기능을 지원합니다.
다른 이름 armanas/froglet, froglet-mcp
- 스타
- 8
- Tools
- —
- 다운로드, 30일
- —
- 최신 버전
- 0.4.0
- 마지막 커밋
- 2026년 11월 9일
- 라이선스
- Apache-2.0
Froglet MCP란?
에이전트 트랜잭션을 위한 레일 중립적 증거 레이어를 제공하며, 검증 가능한 컴퓨팅, 서명된 영수증 및 서비스 게시 기능을 지원합니다.
프로젝트 README에서 생성된 요약
Froglet MCP server 정보
Froglet은 에이전트 트랜잭션에 대한 검증 가능한 기록을 제공하도록 설계된 레일 중립적 증거 레이어입니다. 개발자와 AI 에이전트는 승인된 내용, 적용된 제한 사항, 실행 당사자가 수행한 증명(attestation)을 기록하는 서명 및 해시 연결 영수증을 생성할 수 있습니다. 이 시스템은 자금 이동에 사용되는 특정 금융 레일과 독립적으로 작동하므로, 노드나 특정 계정 없이도 오프라인 검증이 필요한 모든 사용자에게 유용합니다.
사용자는 npx froglet-mcp를 통한 네이티브 MCP server 또는 의존성을 최소화한 네이티브 브릿지를 포함한 여러 인터페이스를 통해 시스템과 상호작용할 수 있습니다. 소프트웨어는 주로 Rust로 구축되었으며 Apple Silicon macOS와 Linux x86_64 및 arm64를 지원합니다. froglet-node 바이너리는 프로바이더이자 런타임 노드 역할을 수행하여, 에이전트가 프로바이더 ID와 URL을 통해 이름이 지정된 서비스, 데이터 기반 서비스 또는 로우 컴퓨팅(raw compute)을 게시할 수 있도록 합니다. 이러한 설정을 통해 봇은 스크립트 가능한 서비스를 생성하고 로컬에서 검증하며, OCI 이미지를 관리할 필요 없이 서비스를 게시할 수 있습니다.
활용 사례
- 1에이전트 서비스를 마켓플레이스에 게시AI 에이전트가 데이터나 Wasm 함수를 준비한 후 marketplace_publish 도구를 사용하여 다른 프로바이더와 공유할 수 있습니다.
프로젝트의 README를 바탕으로 작성되었습니다. 모든 활용 사례는 해당 내용의 인용문과 대조하여 확인합니다.
Froglet MCP 설치 방법
형식 확인: 2026년 11월 10일Cursor에 추가Goose에 추가Claude Code에 Froglet MCP 설치
$ claude mcp add froglet -- npx -y froglet-mcpCursor에 Froglet MCP 설치
~/.cursor/mcp.json
{
"mcpServers": {
"froglet": {
"command": "npx",
"args": [
"-y",
"froglet-mcp"
]
}
}
}VS Code에 Froglet MCP 설치
.vscode/mcp.json
{
"servers": {
"froglet": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"froglet-mcp"
]
}
}
}VS Code CLI에 Froglet MCP 설치
$ code --add-mcp '{"name":"froglet","type":"stdio","command":"npx","args":["-y","froglet-mcp"]}'Claude Desktop에 Froglet MCP 설치
claude_desktop_config.json
{
"mcpServers": {
"froglet": {
"command": "npx",
"args": [
"-y",
"froglet-mcp"
]
}
}
}Codex에 Froglet MCP 설치
$ codex mcp add froglet -- npx -y froglet-mcpCodex config.toml에 Froglet MCP 설치
~/.codex/config.toml
[mcp_servers.froglet] command = "npx" args = ["-y", "froglet-mcp"]
Gemini CLI에 Froglet MCP 설치
$ gemini mcp add -s user froglet npx -- -y froglet-mcpGemini settings.json에 Froglet MCP 설치
~/.gemini/settings.json
{
"mcpServers": {
"froglet": {
"command": "npx",
"args": [
"-y",
"froglet-mcp"
]
}
}
}구성
FROGLET_LIGHTNING_MODEFROGLET_NODE_ROLEFROGLET_PAYMENT_BACKENDFROGLET_PRICE_EXEC_WASMFROGLET_PROFILEFROGLET_PROVIDER_AUTH_TOKEN_PATHFROGLET_PROVIDER_URLFROGLET_RUNTIME_AUTH_TOKEN_PATHFROGLET_RUNTIME_URLFROGLET_RUN_COMPOSE_SMOKE
| 이름 | 설명 | 유형 |
|---|---|---|
| FROGLET_LIGHTNING_MODE | — | |
| FROGLET_NODE_ROLE | — | |
| FROGLET_PAYMENT_BACKEND | — | |
| FROGLET_PRICE_EXEC_WASM | — | |
| FROGLET_PROFILE | — | |
| FROGLET_PROVIDER_AUTH_TOKEN_PATH | — | |
| FROGLET_PROVIDER_URL | — | |
| FROGLET_RUNTIME_AUTH_TOKEN_PATH | — | |
| FROGLET_RUNTIME_URL | — | |
| FROGLET_RUN_COMPOSE_SMOKE | — |
자동 검사
2026년 11월 13일에 실행됨전체 보고서 →자동 검사를 통과했습니다. 이는 보안 감사가 아닙니다. 코드를 직접 읽고 서버에 필요한 최소한의 권한만 부여하십시오. 검사 방식
- 소스 저장소 연결 및 접근 가능통과
- 오픈 소스 라이선스 (OSI)통과
- 최근 1년 내 커밋 또는 릴리스통과
- 패키지 이름이 인기 패키지와 유사하지 않음통과
- 원격 엔드포인트가 HTTPS 사용검사되지 않음
- 파괴적인 도구에 라벨 표시됨검사되지 않음
- 모델에 대한 숨겨진 지침 없음통과
- 패키지 및 해당 종속성에 중요한 보안 권고 없음통과
README
GitHub에서 보기Froglet
The rail-neutral evidence layer for agent transactions.
Whatever rail moves the money, Froglet produces a signed, hash-linked record of what was authorised, under which limits, and what the executing party attested to — verifiable offline by anyone, with no node and no account.
Verify a chain yourself in one command:
cargo run -p froglet-verify -- conformance/kernel_v1.json
Created and maintained by Armanas Povilionis. Copyright Alithea Biotechnology GmbH, which operates the hosted services.
Table of Contents
Overview
Launch preparation, 2 October 2026: the current scope is free selected-data sharing and bounded Wasm through native MCP, with optional configured A2A. Use the production handoff and explicit remaining gates. Local software checks, local packages and the public documentation website do not qualify a new production compute deployment or an immutable release.
Development checkpoint, 1 October 2026: work has resumed on the free native MCP workflow, optional configured A2A transport, and selected-data sharing with bounded Wasm computation. The final candidate's actual Codex CLI workflow and extracted demo bundle are qualified on the current Apple Silicon Mac. The same production inputs also passed 625 library and nine native MCP tests, plus the terminology workflow, in an isolated Linux arm64 development container. Thirteen library tests were ignored; Python OS isolation and a Linux agent host remain unqualified. Claude Code host qualification is blocked on sign-in. Selected-data/profile examples and bounded local capacity have recorded checks; clean machines, public compute HTTPS/capacity, and first-user value remain separate gates. See current work and the qualification evidence.
The public documentation website was restored on 1 October; seven checked routes returned HTTP 200 and matched the local build bytes. See the deployment record and remaining gates. The last observed hosted compute trial was unavailable; this pass has not requalified that deployment. Production and available hosted compute qualification remain open. The immutable v0.4.7-beta.1 prerelease is now published; its downloaded Apple Silicon binary passed a separate deterministic localhost replay. Distribution and exact released-Mac evidence also record a separate native-MCP Mac-to-Fly proof: six bounded free cases, restart persistence and offline task recovery on one isolated test host, now removed. That proof does not qualify production, load, a fresh agent host, a clean operating system or paid operation. The September pause and rollback record is preserved. File downloads remain experimental; Node/UI/relay qualification is open. General CPU jobs, GPU rental, and new payment work are outside the resumed scope.
The optional research profile makes namespace, version, field types, units, provenance and mapping assumptions explicit. The final local example retrieves a pinned public GO catalog in 100- and 40-row pages with verified receipts; this establishes data exchange and declared metadata compatibility, not scientific correctness or independent user demand.
On one shared machine, eight concurrent tiny workflows using independent requester identities completed without errors in the measured sweep. At 16 and 32, six selected-data jobs hit the four-process limit and returned independently verified failure chains. Shared requester identities still face about one identical requester/workload/terms Quote per second; collisions now fail before Deal admission and leave no pending requester submissions. These measurements do not establish production capacity, performance superiority or developer-time savings. See the exact local qualification.
The current product candidate starts with “Make this catalog usable by another agent.” Codex or Claude Code can prepare selected JSON, typed CSV, or SQLite data, or a small Wasm function; run a local example; ask for exact publication approval; and return a provider/service share link. Native setup preserves existing agent settings. Recipients use the same native integration without npm.
See the publishing workflow and the qualification gates. This is a public prerelease: clean-machine agent tests, the real public relay journey through this profile, and first-time-user acceptance remain qualification gates. The first-user trial kit separates setup usability from evidence of a useful job. Supported targets are Apple Silicon macOS and Linux x86_64/arm64. The host must stay online.
The local captured-program replay now has complete evidence: Alice's retained
81,222-byte Wasm program matched four independently frozen answers, and all
seven admitted signed chains were verified offline. The seven include an owner
control, selected-data retrieval, four program executions and a separate looping
program's budget failure. Exact completed data and compute retries recovered
cached results after Bob stopped and Alice restarted. In that historical replay,
get_task returned HTTP 502 with Bob stopped. The 3 October correction now
revalidates saved terminal task evidence locally; unsigned or unresolved tasks
still require Bob for refresh. Current-source local regressions pass; a fresh
optimized native candidate also passed independent verification of all seven
chains and all six saved Alice terminal task reads after Bob stopped and Alice
restarted. Package/release gates and website publication remain separate checks.
This is deterministic replay of a program created in an earlier actual Codex session. That earlier fresh-session qualification remains failed because its complete chains were not captured and persistent Codex configuration changed. The replay starts no LLM, compiler or agent settings operation. See the reusable runbook and local finalization evidence.
Froglet gives one signed economic primitive for three product shapes:
| Shape | Description |
|---|---|
| Named Services | Discoverable, published service endpoints |
| Data-Backed Services | Services backed by bot-authored data or projects |
| Open-Ended Compute | Raw compute targeted via provider_id or provider_url |
The primary bot-facing integration surfaces are intentionally simple:
- One OpenClaw/NemoClaw plugin id:
froglet(atintegrations/openclaw/froglet/) - One MCP server under
integrations/mcp/froglet/, published asfroglet-mcpfornpx froglet-mcp, plus a dependency-minimal native bridge infroglet-node mcp - Both surfaces register a single agent-facing tool named
frogletto the host (Claude Code, Codex, Cursor, Windsurf, etc.). The headline action is two-stepmarketplace_publish: the first call is a non-mutating plan, and the second must carry the exact approvedconsent_hash. Both native and JavaScript adapters delegate to the same Rust publication module and provider-control contract. - The
froglet-nodebinary is both the daemon (running as a provider) and the author CLI (froglet-node init/build/publish/whoami,prepare-service,invoke,doctor, andcheck-updates).status --open --jsonreturns a short-lived read-only local status URL.
Bots should be able to create small scriptable services directly, validate them locally, and publish them without starting from OCI images. OCI containers remain a supported packaging and deployment path.
Product Model
- Any Froglet node can publish resources and invoke remote resources
- Published resources are execution bindings backed by bot-authored projects, explicit source, or prebuilt artifacts
- Easy bot authoring and local checking of scriptable services is a core product requirement
- Identity is first-class in signed artifacts
- Clearnet HTTPS, outbound relay HTTPS, and Tor v3 onion transport are supported registration paths for self-hosted providers. Bootstrap plans the official relay URL and suffix by default but keeps them dormant; publication fails closed until an exact durable grant makes the endpoint ready. Planned configuration is not a claim that its public operator service is deployed.
[!NOTE] Marketplace, ranking, incentive, and broker policy live above the protocol. Payment rails are adapter-level surfaces for local or self-hosted operators, not normal buyer onboarding. Lightning, Stripe, and x402 are the launch adapters in this repo. Lightning and x402 have standardized kernel settlement methods with conformance vectors; Stripe's
stripe_mpp.v1receipt shape is standardized but attested rather than cryptographic. Only Lightning escrow uses the signed invoice-bundle flow. x402's wire format is settled (x402.eip3009.v1,conformance/x402_v1.json) but its rail is not: no publish-path exposure and no live transcript on any network. The first-party hostedtry.froglet.devtrial was retired on 8 October 2026; the free public beta at https://froglet.dev/services/ replaces it. Hosted paid rails must not be claimed live until Lightning and Stripe have public payment transcripts, and users should not be asked to manage LND channels or payment secrets just to try Froglet.
- Named services and data services are discovered through discovery
- Open-ended compute uses the provider's direct compute offer via
run_compute, targeted withprovider_idorprovider_url - Publication and bootstrap adapters may include Nostr-style publication without making any single relay or network the kernel source of truth
Components
Product-wise, Froglet is one node that can both provide and consume. The reference implementation exposes these binaries:
| Binary | Purpose | Default Port |
|---|---|---|
froglet-node |
Provider and/or runtime node (role configured via env) | 8080 / 8081 |
[!TIP] Marketplace integration is part of the public Froglet surface. Runtimes can point at the default public marketplace with
FROGLET_MARKETPLACE_URL; providers can self-register there after exposing a public HTTPS URL, identity-assigned relay URL, Tor v3 onion URL, or claimed*.providers.froglet.devhostname. See docs/MARKETPLACE.md.
Prerequisites
Binary install (quickest): curl, tar, sha256sum (Linux) or shasum (macOS). Supported: Linux x86_64/arm64, macOS arm64.
Build from source: Rust 1.91+, Python 3.12+ (for tests), Node 18+ with npm (for Claude Code/Codex MCP setup and integration tests).
Docker (optional fallback/contributor stack): Docker with Compose v2.
Quick Start
Canonical onboarding lives in
docs-site/src/content/docs/docs.mdx
(the learn/ index remains as a legacy route).
Use the repo README for the product and codebase overview, the docs-site
manual for the public launch path, and docs/ for specs, operator notes, and
integration reference.
The public launch story still has exactly two entry points:
1. Try the public beta
- Open https://froglet.dev/services/ in a browser: no installation, signup or invitation. It runs your bounded Wasm program or a synthetic catalog query on a real remote node and returns signed evidence the browser verifies.
- Scope and limits: docs-site/src/content/docs/learn/cloud-trial.mdx
- The older hosted trial at
try.froglet.dev(and itsai.froglet.devorigin) was retired on 8 October 2026; see docs/HOSTED_TRIAL.md. - The public beta does not prove paid rails, persistent identity, service publication, marketplace depth, or general runtime access.
2. Run Locally
- Start with docs-site/src/content/docs/learn/quickstart.mdx
- Then use docs-site/src/content/docs/learn/agents.mdx and docs-site/src/content/docs/learn/payment-rails.mdx
- Self-host and operator follow-ons live in docs/DOCKER.md, docs/GCP_SINGLE_VM.md, and docs/MARKETPLACE.md
Minimal full local stack from zero:
set -eu
repo=AlitheaBio/froglet
metadata="$(mktemp "${TMPDIR:-/tmp}/froglet-release.XXXXXX")"
bootstrap="$(mktemp "${TMPDIR:-/tmp}/froglet-agent-bootstrap.XXXXXX")"
release_url="$(curl -fsSL --proto '=https' --proto-redir '=https' --tlsv1.2 -o /dev/null -w '%{url_effective}' "https://github.com/$repo/releases/latest")"
tag="${release_url%/}"; tag="${tag##*/}"
printf '%s' "$tag" | grep -Eq '^v[0-9A-Za-z][0-9A-Za-z.+-]*$'
curl -fsSL --proto '=https' --proto-redir '=https' --tlsv1.2 -H 'Accept: application/vnd.github+json' \
-H 'X-GitHub-Api-Version: 2026-03-10' \
"https://api.github.com/repos/$repo/releases/tags/$tag" -o "$metadata"
[ "$(sed -n 's/^ "immutable": \([a-z]*\),*$/\1/p' "$metadata")" = true ]
[ "$(sed -n 's/^ "tag_name": "\([^"]*\)",*$/\1/p' "$metadata")" = "$tag" ]
asset_record="$(awk '
/^ \{/ { in_asset=1; name=digest=state=""; next }
in_asset && /^ "name":/ { v=$0; sub(/^ "name": "/,"",v); sub(/",*$/,"",v); name=v }
in_asset && /^ "digest":/ { v=$0; sub(/^ "digest": "/,"",v); sub(/",*$/,"",v); digest=v }
in_asset && /^ "state":/ { v=$0; sub(/^ "state": "/,"",v); sub(/",*$/,"",v); state=v }
in_asset && /^ \},*$/ { if (name=="agent-bootstrap.sh") print digest "|" state; in_asset=0 }
' "$metadata")"
[ "$(printf '%s\n' "$asset_record" | sed '/^$/d' | wc -l | tr -d ' ')" = 1 ]
bootstrap_digest="${asset_record%%|*}"; asset_state="${asset_record#*|}"
[ "$asset_state" = uploaded ]
printf '%s' "$bootstrap_digest" | grep -Eq '^sha256:[0-9a-f]{64}$'
bootstrap_digest="${bootstrap_digest#sha256:}"
curl -fsSL --proto '=https' --proto-redir '=https' --tlsv1.2 \
"https://github.com/$repo/releases/download/$tag/agent-bootstrap.sh" -o "$bootstrap"
if command -v sha256sum >/dev/null 2>&1; then actual="$(sha256sum "$bootstrap" | awk '{print $1}')";
elif command -v shasum >/dev/null 2>&1; then actual="$(shasum -a 256 "$bootstrap" | awk '{print $1}')";
else actual="$(openssl dgst -sha256 "$bootstrap" | sed 's/^.*= //')"; fi
[ "$actual" = "$bootstrap_digest" ]
chmod 0700 "$bootstrap"
VERSION="$tag" sh "$bootstrap" plan
# Show the complete plan and wait for approval. Then copy its exact hash:
VERSION="$tag" sh "$bootstrap" execute '<install_approval_hash>'
rm -f "$bootstrap" "$metadata"
plan is the safe default and writes nothing outside its temporary workspace.
It resolves the immutable release, verified manifest and target-platform binary
asset digests, exact bootstrap/install/configuration script digests, persistent
paths, and service-manager impact into one canonical approval hash. execute
recomputes that contract and stops before any persistent write if the release,
script bytes, profile, paths, or supplied hash changed.
After approval, the agent bootstrap verifies a Release Bundle, installs the checksum-verified
froglet-node, starts the native dual-role service through launchd/systemd,
writes native MCP config, and returns only after health plus a non-seeding MCP
status proof and a transient read-only data proof that is confirmed-unpublished.
The proof first requires an empty active-offer feed and no lifecycle for its
exact service ID, records a private cleanup intent before publishing, and
removes both its offer and authoring directory only after an empty-feed check.
A failed proof attempts exact unpublish while the service is still live and
preserves a recovery marker if cleanup cannot be proved.
A digest-pinned dual-role GHCR image is the fallback
when native service management is unavailable.
The approval plan includes the official relay endpoint and DNS suffix by
default. This is dormant configuration: it derives an exact planned HTTPS URL
but opens no WSS connection and exposes no service before an exact durable
publication grant. Set both FROGLET_RELAY_URL='' and
FROGLET_RELAY_PUBLIC_SUFFIX='' before plan to opt out. The first-party
DNS/TLS endpoint is not yet live-proven, so relay publication remains an
explicit external gate rather than an install success claim. See
docs/RELAY.md. The implementation/evidence matrix in
docs/AGENT_FIRST_PUBLICATION_PLAN.md
tracks which clean-host and external gates remain open.
From a trusted checkout, run the disposable-host proof with
bash scripts/fresh_host_quickstart_smoke.sh.
The lower-level scripts/install.sh is an internal release-bundle installer;
agents should use the approval-gated bootstrap above.
Source-checkout Compose and generated host-side agent configs depend on
FROGLET_HOST_READABLE_CONTROL_TOKEN=true; the default user path is the
no-clone /agent bootstrap. The quickstart page carries the step-by-step
MCP-first explanation, payment-rail decisions, Tor registration, managed
subdomains, and contributor/source-mode fallbacks.
# Provider node
FROGLET_NODE_ROLE=provider \
FROGLET_PRICE_EXEC_WASM=10 \
FROGLET_PAYMENT_BACKEND=lightning \
FROGLET_LIGHTNING_MODE=mock \
cargo run -p froglet --bin froglet-node
# Runtime node
FROGLET_NODE_ROLE=runtime \
FROGLET_PAYMENT_BACKEND=lightning \
FROGLET_LIGHTNING_MODE=mock \
cargo run -p froglet --bin froglet-node
The normal model is one node running both provider and runtime roles
(FROGLET_NODE_ROLE=dual), so it can publish local resources and invoke
remote ones.
Set FROGLET_MARKETPLACE_URL on runtime nodes to search through an external
marketplace. Providers can self-register with the default public marketplace
after they advertise a matching public HTTPS origin, assigned relay HTTPS URL,
Tor v3 onion URL, or claimed *.providers.froglet.dev hostname.
Bot Surfaces
OpenClaw, NemoClaw, and MCP-compatible hosts are the primary bot-facing surfaces today. Distribution status and marketplace/plugin ordering live in PLUGIN_DISTRIBUTION.md.
OpenClaw & NemoClaw
Use the shared plugin package in
integrations/openclaw/froglet.
Current OpenClaw plugin install/inspect and gateway invocation require
Node.js 22.14.0 or newer. Gateway-mediated local actions should launch with
FROGLET_PROVIDER_AUTH_TOKEN_PATH and FROGLET_RUNTIME_AUTH_TOKEN_PATH
pointing at the local data/runtime/ token files.
| Key | Purpose |
|---|---|
hostProduct |
Target host product |
providerUrl |
Provider/public API base URL |
runtimeUrl |
Runtime API base URL |
providerAuthTokenPath |
Path to the provider control token |
runtimeAuthTokenPath |
Path to the runtime auth token |
baseUrl |
Legacy single-surface fallback URL |
authTokenPath |
Legacy single-token fallback path |
requestTimeoutMs |
HTTP request timeout |
defaultSearchLimit |
Default discovery result limit |
maxSearchLimit |
Maximum discovery result limit |
The generated local OpenClaw config uses the split provider/runtime keys above.
Legacy baseUrl and authTokenPath remain supported for single-surface
configs such as the checked-in NemoClaw examples.
The one froglet tool covers:
- Service discovery and invocation
- Agent-grade service publication via
marketplace_publish - Local artifact publication via
publish_artifact - Settlement visibility and current marketplace wrappers
- Status and task polling
- Exact install approval via non-mutating
plan_install, then approved command generation viaget_install_guide - Post-install workflow planning via
plan_use_case - Raw compute
summaryis metadata only; it does not generate codepublish_artifactis the current local publication pathrun_computeis the low-level path for open-ended compute and should includeprovider_idorprovider_url- Project authoring, log tailing, and node restart are not part of the current public tool API
Froglet MCP 질문
Froglet MCP를 Claude Code에 추가하려면 어떻게 하나요?
터미널에서 claude mcp add froglet -- npx -y froglet-mcp를 실행합니다. 모든 프로젝트에서 사용할 수 있게 하려면 --scope user를 추가합니다.
Froglet MCP를 Cursor에 추가하려면 어떻게 하나요?
~/.cursor/mcp.json를 열고 mcpServers 아래에 서버를 추가합니다. 위 설치 섹션에 표시된 Cursor 구성을 사용합니다.
Froglet이 AI 에이전트에게 제공하는 기능은 무엇인가요?
에이전트가 작은 스크립트 기반 서비스를 생성하고, 이를 로컬에서 검증하며, OCI 이미지를 관리할 필요 없이 게시할 수 있도록 지원합니다.