Edgegate Mcp

io.github.frozo-ai由 GitHub 帳號 frozo-ai 發布v0.19.0更新於 Oct 24, 2026

Regression-test AI models on real Snapdragon and Jetson hardware, with signed evidence for CI.

概覽

AI 產生的概覽

EdgeGate 的 MCP 伺服器,讓助理在真實 Snapdragon 與 Jetson 硬體上建立並執行 AI 模型回歸閘門,並為 CI 產生簽章證據。

可以做什麼
  • 建立與管理工作區、成員、角色和 API 金鑰
  • 定義含裝置矩陣與指標閘門的回歸流水線,觸發執行、輪詢狀態並與基準比較
  • 匯入 Hugging Face ONNX 模型、預測 NPU 覆蓋率,並透過 Qualcomm AI Hub 編譯 LLM
  • 管理行為評測集、promptpack、參考基準擷取與行為閘門執行
  • 取得簽章稽核與合規報告、匯出執行報告,並產生 GitHub Actions CI 設定
  • 連接 Qualcomm AI Hub、Hugging Face 與自有 S3 儲存整合,並向自行連接的裝置派送基準測試
適用情境
當你希望助理端到端建置邊緣 AI 回歸測試時使用:挑選裝置、設定延遲或記憶體閘門、在真實硬體上執行,並把簽章證據接上 CI 或合規紀錄。
執行需求
從 npm 套件 edgegate-mcp 安裝的本機 stdio 程序(需 Node.js 與 npx)。需要 EdgeGate 帳號與 EDGEGATE_API_KEY 環境變數;自架部署可用 EDGEGATE_API_URL 覆寫預設 API 位址。執行還需為工作區連接 Qualcomm AI Hub 權杖,並需要網路存取。
安裝前請注意
API 金鑰是寫入用戶端設定的機密。多個工具具破壞性或高權限:撤銷 API 金鑰、移除成員、中斷整合、刪除自有儲存授權,部分需要 owner 或 admin 角色。建立 API 金鑰時明文權杖只回傳一次。執行會消耗 Qualcomm AI Hub 額度,LLM 閘門成本約為一般視覺執行的三倍。伺服器會為工作區保存 Hugging Face 與 AI Hub 權杖,並可在你的 AWS 帳號中扮演 IAM 角色。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 Edgegate Mcp,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。

Desktop only,透過 STDIO。 STDIO 服務會啟動本機處理程序,因此需要 SourceWeft 桌面主機。

其他 MCP 客戶端

參照 儲存庫 中的啟動說明。

README

edgegate-mcp

MCP server for EdgeGate — set up edge-AI regression gates on Snapdragon devices directly from Claude Code, Cursor, or Claude Desktop.

What does it do?

EdgeGate runs AI model regression tests on real Snapdragon hardware via Qualcomm AI Hub, then produces signed evidence bundles you can attach to CI gates. This npm package exposes EdgeGate's REST API as 61 MCP tools, plus bundled skills, so you can drive the whole flow from a prompt:

> Use the edgegate MCP to set up a CI gate for my MobileNet ONNX model.> Gates: inference_time_ms ≤ 10, peak_memory_mb ≤ 150.> Devices: Galaxy S24, Galaxy S23.

Install

bash
# 1. Generate an API key in the EdgeGate dashboard# https://edgegate.frozo.ai/workspace/<id>/settings#api-keys
# 2. Run the installer from the edgegate-mcp package (writes config for Claude Code / Cursor / Desktop)npx [email protected] edgegate-mcp-install

Restart your MCP client. Done.

Use it on the web (claude.ai, ChatGPT)

No install and no API key. Add the hosted endpoint as a custom connector and approve it in your browser:

https://edgegate.frozo.ai/mcp
  • Claude — Settings → Connectors → Add custom connector.
  • ChatGPT — Settings → Apps & Connectors → enable Developer mode, then create a connector with OAuth authentication.

You sign in to EdgeGate and approve once; disconnect any time under Settings → Connected AI apps in the dashboard. The hosted endpoint has no access to your filesystem, so edgegate_export_run_report returns the report inline instead of writing a file.

Any client that speaks Streamable HTTP can also use it with an API key as a bearer token:

bash
claude mcp add --transport http edgegate https://edgegate.frozo.ai/mcp --header "Authorization: Bearer egk_live_..."

Self-hosting: handleMcpRequest from edgegate-mcp/http is a Web-standard (Request) => Response handler you can mount on any fetch runtime.

Manual config

If you'd rather edit config files yourself, the server is a standard stdio MCP. Add this to your client's config:

Claude Code (~/.claude.json)

json
{  "mcpServers": {    "edgegate": {      "type": "stdio",      "command": "npx",      "args": ["-y", "edgegate-mcp"],      "env": {        "EDGEGATE_API_KEY": "egk_live_...",        "EDGEGATE_API_URL": "https://edgegateapi.frozo.ai"      }    }  }}

Cursor (~/.cursor/mcp.json)

Same shape as Claude Code without the type field.

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json on macOS)

Same shape as Cursor.

Tools

See docs/tools.md for input schemas and examples.

EdgeGate exposes 61 MCP tools. This table is generated from src/server.ts — run npm run docs:tools after adding a tool.

ToolPurpose
edgegate_setup_workspaceConfirm or list EdgeGate workspaces visible to the API key. Run this first in a fresh conversation to lock in which workspace_id the other tools should use.
edgegate_create_pipelineCreate a new EdgeGate regression pipeline. Define which model(s), which device(s), and which gates (e.g. inference_time_ms ≤ 10) the pipeline will enforce. For LLMs: set llm_compile_source on a model instead of artifact_id. EdgeGate will compile + link via AI Hub on first run; subsequent runs reuse the cached composite. ttft_ms + tps gates work; both are derived from per-component profile (prompt-role inference_time → TTFT; 1000/token-role inference_time → TPS). Each LLM gate run = 3 AI Hub profile jobs (one per component) ≈ 3× CV cost.
edgegate_run_gateTrigger an EdgeGate run against a pipeline. Returns a run_id you can poll with edgegate_check_status.
edgegate_check_statusGet the current status of an EdgeGate run, including per-device metrics and which gates passed or failed.
edgegate_get_reportList recent EdgeGate runs in a workspace with status, duration, and trigger.
edgegate_get_audit_reportGet the signed audit report PDF URL for a completed EdgeGate run. Used for compliance records.
edgegate_setup_github_actionGenerate the GitHub Actions workflow YAML + gh secret commands so every PR runs EdgeGate as a CI gate.
edgegate_compare_runsDiff two EdgeGate runs — standard pipeline metrics/gates or Behavioral-Gate summary signals. Reports REGRESSION / IMPROVEMENT / NEUTRAL, or NO BASELINE / NOT COMPARABLE / INSUFFICIENT EVIDENCE when a reliable comparison cannot be made. Auto-selects a recent PASSED run from the same pipeline when baseline_run_id is omitted. Runs without a pipeline (including Behavioral-Gate runs) require an explicit baseline_run_id with matching eval-set and gate configuration.
edgegate_export_run_reportDownload a human-readable markdown report for an EdgeGate run and save it to disk. Returns the absolute file path plus a preview of the first 30 lines. Optionally includes a run-vs-baseline diff section (include_diff=true).
edgegate_import_huggingface_modelImport a public Hugging Face model that contains a pre-built ONNX file. EdgeGate downloads the file and registers it as an Artifact. Returns the artifact_id you can pass directly to edgegate_create_pipeline. Polls until the import completes by default (poll_for_completion=true); set to false to return immediately with the job id.
edgegate_predict_npu_coveragePredict which ONNX ops will run on the Qualcomm Hexagon NPU vs fall back to CPU, BEFORE spending any AI Hub credits. Returns a compute-weighted NPU coverage % (the latency-honest number), op-count coverage, risk band, per-op CPU fallbacks, and fix recommendations. Heuristic — the real device run remains authoritative.
edgegate_list_promptpacksList all promptpacks in an EdgeGate workspace. Returns a markdown table with promptpack_id, version, case count, published status, and creation date. Use include_unpublished=false to hide draft packs.
edgegate_create_promptpackCreate a new promptpack in an EdgeGate workspace. A promptpack defines the test cases (prompts, expected outputs, per-case overrides) that regression pipelines evaluate. Requires admin role on the workspace. Packs are immutable after creation — bump the version to update.
edgegate_publish_promptpackPublish a promptpack version in an EdgeGate workspace so it can be referenced in pipelines. Newly created packs start as unpublished — call this after edgegate_create_promptpack to complete the create → publish → use lifecycle. Requires admin role on the workspace. The operation is idempotent.
edgegate_connect_huggingfaceStore a personal HuggingFace access token for this workspace so the import flow can read private / gated / Qualcomm-org repos (most qualcomm/, Intel/, and many Xenova/* repos 401 the anonymous endpoint). The token is validated against HF whoami before encryption and is never echoed in plaintext. If an integration already exists this tool rotates the token. Requires admin role.
edgegate_get_huggingface_integrationShow whether a personal HuggingFace token is connected to this workspace (and whether it is currently active or disabled). Does not return the token itself.
edgegate_disconnect_huggingfacePermanently delete the workspace's HuggingFace integration. Future HF imports fall back to anonymous access. Requires owner role.
edgegate_connect_qaihubStore a Qualcomm AI Hub API token for this workspace so EdgeGate can submit compile + profile jobs on real Snapdragon devices. The token is encrypted at rest and is never returned in plaintext after the initial connect. If an integration already exists this tool transparently rotates the token. Requires admin role.
edgegate_get_qaihub_integrationShow whether a Qualcomm AI Hub token is connected to this workspace and whether it is currently active. Does not return the token itself.
edgegate_disconnect_qaihubPermanently delete the workspace's Qualcomm AI Hub integration. Any new EdgeGate runs in this workspace will then fail with NO_AIHUB_TOKEN until a fresh token is connected. Requires owner role.
edgegate_create_workspaceCreate a new EdgeGate workspace. The caller automatically becomes the owner. Subject to plan-tier workspace limits. After creation, connect Qualcomm AI Hub and define pipelines as usual.
edgegate_list_api_keysList all API keys in this workspace (id, name, prefix...suffix, status, last_used). Plaintext is never returned. Requires owner role.
edgegate_create_api_keyCreate a new API key for this workspace. The plaintext token is returned EXACTLY ONCE in the response — copy it to your CI secrets or local env immediately. Requires Pro tier or above. Requires owner role.
edgegate_revoke_api_keyRevoke an API key by id. The key is immediately rejected for all subsequent requests; the row is preserved (with revoked_at set) so the audit trail survives. Destructive. Requires owner role.
edgegate_list_membersList all members of this workspace with their email + role. Requires at least viewer role.
edgegate_list_devicesList every Qualcomm AI Hub device EdgeGate can target (Snapdragon phones, QRD/CRD reference platforms, IoT Dragonwing, automotive, XR). Returns a markdown table grouped by category. Use the id column verbatim when building a create_pipeline device matrix. No workspace_id needed — the catalog is global.
edgegate_list_device_targetsList the customer's OWN connected devices (Jetson, Snapdragon hosts, gateways) with live/offline status computed from each device's heartbeat (30s beat, 90s window). This is the workspace fleet connected via edgegate-runner agent — NOT the global AI Hub catalog (use edgegate_list_devices for that). If empty, the response includes copy-paste connect instructions for the customer.
edgegate_run_device_benchmarkDispatch an ONNX benchmark to one of the customer's connected devices (Jetson, Snapdragon host, gateway) by name/id. The on-device agent picks it up within ~30s and reports latency/memory results. Multi-silicon: works for any vendor the customer has connected. Requires admin role.
edgegate_invite_memberAdd an existing EdgeGate user to this workspace by email at the given role (owner / admin / viewer). v1 only attaches existing users — does not send invitation emails to external addresses. Requires admin role; only owners can add other owners.
edgegate_change_member_roleChange a member's role in this workspace. Cannot downgrade the last owner — promote another member to owner first. Requires owner role.
edgegate_remove_memberRemove a member from this workspace. The user loses access immediately; their pipelines and runs are preserved. Cannot remove the last owner. Destructive. Requires owner role.
edgegate_setup_byo_storageEnterprise only. Zero-friction BYO storage setup — creates a pending grant with EdgeGate (returns the External ID) and returns the exact AWS CLI commands the agent should run to create the IAM role in the customer's AWS account. Pair with edgegate_attach_byo_role to finalize. Prefer this over edgegate_register_byo_bucket for new setups — the agent doesn't have to figure out the trust policy or guess at role names. Requires owner role. 402 from non-Enterprise workspaces; if a non-pending grant exists, instructs the agent to disconnect first.
edgegate_attach_byo_rolePhase 2 of edgegate_setup_byo_storage. Hands the freshly-created Role ARN back to EdgeGate, which runs sts:AssumeRole + a deny-by-default HEAD probe to verify the trust + permission policies are correct. Flips the grant to 'active' on success, or returns a typed BYO_* error with a checklist of common misconfigurations on failure. Re-callable with the same role_arn after fixing IAM.
edgegate_register_byo_bucketEnterprise only. Use edgegate_setup_byo_storage instead for new setups — this tool only works if you already have an IAM role and just want to register its ARN. Registers the workspace's customer-owned S3 bucket + IAM role as a BYO storage grant. EdgeGate's workers will AssumeRole into your AWS account to read model bytes — they never leave your account. Returns the External ID you must add to your role's trust policy. Requires owner role. 402 from non-Enterprise workspaces; 409 if a grant already exists.
edgegate_check_byo_bucketRe-run the AssumeRole + HeadObject readiness probe against the workspace's BYO grant. Returns the updated grant status with the typed BYO_* error code if it failed, plus a checklist of common IAM / bucket / KMS misconfigurations to inspect. Requires admin role.
edgegate_register_byo_artifactRegister an existing S3 URI in your registered bucket as an EdgeGate Artifact. EdgeGate HeadObjects the URI to confirm the key exists + capture size/etag — bytes are NOT uploaded through EdgeGate. Returns an artifact_id you can pass directly to edgegate_create_pipeline / edgegate_run_gate. Requires admin role. Pre-conditions: Enterprise plan + active BYO grant + bucket matches the grant.
edgegate_disconnect_byo_bucketDelete the workspace's BYO storage grant. EdgeGate stops attempting to read from your bucket. Refuses (409) if artifacts still reference it — surface lists the safe paths forward (rotate External ID via dashboard, or drop the artifacts first). Destructive. Requires owner role.
edgegate_llm_compileSubmit a multi-component LLM compile + link job via Qualcomm AI Hub. Returns a compile_job_id; poll with edgegate_check_llm_compile_status. Spend is gated by the workspace's monthly LLM compile cap (default 100/mo Pro tier). Each compile produces a composite QNN_DLC linked model + 3 component artifacts (prompt / token / kv_cache). Compile target_runtime is QNN_DLC under the hood despite the genie label — the label is for downstream profile dispatch hints only.
edgegate_check_llm_compile_statusPoll an LLM compile job. Returns status (queued|running|completed|failed), composite_artifact_id when complete, error_detail when failed, and progress.compile_jobs_done / total.
edgegate_get_byo_auditFetch the workspace's append-only BYO storage audit log (every AssumeRole, HeadObject, GetObject, KMS Decrypt). Supports filters by artifact_id, run_id, since timestamp, plus cursor pagination. Each row's aws_request_id is the join key for cross-referencing against your own CloudTrail. Requires admin role.
edgegate_list_eval_packsList the bundled behavioral eval-set starter packs a customer can clone from. Returns each pack's id, name, case count, and balance (must_refuse / task counts). Clone one into a new eval set via edgegate_create_eval_set(clone_from=). No workspace_id needed — the pack library is global.
edgegate_create_eval_setCreate a new behavioral eval set with its first draft version. Seed it from a bundled pack (clone_from), with explicit cases, both, or neither (empty draft). Each case has six fields: case_id, prompt, category (jailbreak|forbidden_action|task|format), forbidden_actions, must_refuse, expected_task_answer. Drafts are NOT validated here — only edgegate_publish_eval_set freezes + gates. Requires workspace write access.
edgegate_list_eval_setsList the workspace's behavioral eval sets — eval_set_id, name, latest version, and creation date. Use the eval_set_id with the update / publish / new-version tools.
edgegate_update_eval_setReplace ALL cases of a DRAFT eval-set version (full replacement list, not a patch). Published versions are immutable — a 409 here means you must fork a fresh draft with edgegate_new_eval_set_version first. Still leaves the version a draft; publish separately. Requires workspace write access.
edgegate_publish_eval_setValidate and freeze a draft eval-set version into an immutable, signed, hash-anchored version. On success returns eval_set_sha256 + artifact_id (feed both into a 3b Behavioral-Gate run). On validation failure returns the balance/structural violations and the version stays a draft (floor: ≥5 must_refuse-with-forbidden cases + ≥1 task case). Requires workspace write access.
edgegate_new_eval_set_versionFork a fresh draft (version+1) from a PUBLISHED version, seeded with its cases — the edit-after-publish path. Edit the new draft with edgegate_update_eval_set, then re-publish. The original published version (and any references / runs bound to its sha) is untouched. Requires workspace write access.
edgegate_list_workflow_endpointsList the workspace's saved API/workflow endpoints — the targets of an API gate (n8n, Zapier, Make, or any OpenAI-compatible API). Returns each endpoint_id to pass to edgegate_capture_reference and edgegate_create_bg_run. Credentials are never returned; only whether one is stored and its last 4 characters.
edgegate_create_workflow_endpointSave an API/workflow endpoint so a behavioral gate can target it. This is the ONLY way to gate an authenticated endpoint: the credential is stored envelope-encrypted here and is never returned, whereas an inline http descriptor is rejected (422) if it carries one. Using a saved endpoint also lets EdgeGate refuse a run whose endpoint differs from the baseline's. Requires workspace admin access.
edgegate_probe_workflow_endpointSend one test request to a saved endpoint and show the raw response beside the text EdgeGate extracted from it. Run this BEFORE capturing a baseline: it is the only thing that catches a wrong response_text_path, which yields empty text — and empty text scores as a refusal, so an all-empty baseline makes every later gate pass trivially.
edgegate_capture_referenceTrigger a reference-oracle capture — the known-good baseline the Behavioral Gate diffs against. Specify EXACTLY one flavor: hf_repo (auto-FP16, EdgeGate runs the un-quantized same model), reference_upload_artifact_id (golden, customer-supplied), or endpoint_id/http (an API or workflow endpoint — n8n, Zapier, Make, or any OpenAI-compatible API; EdgeGate calls it once per case, nothing to install). Returns a job_id to poll. Requires workspace admin access.
edgegate_check_reference_capture_statusPoll a reference-capture job. When status is done, returns reference_artifact_id — an ArtifactKind.REFERENCE artifact — to feed into edgegate_create_bg_run as the gate's trustworthy baseline. Requires workspace admin access.
edgegate_compile_genieSubmit a 3-lane genie compile for the Behavioral Gate's self-hosted runner. Specify EXACTLY one lane selector: hf_repo (Lane A — EdgeGate compiles a HuggingFace repo), onnx_artifact_ids (Lane B — genie-ready multi-part ONNX, compile-and-link), or bundle_artifact_id (Lane C — an already-precompiled bundle). Returns a job_id to poll. Requires workspace admin access.
edgegate_check_genie_compile_statusPoll a genie-compile job. When status is done, returns bundle_artifact_id — the compiled genie bundle to feed into edgegate_create_bg_run. Requires workspace admin access.
edgegate_create_bg_runWire what you are gating + a published eval set + a reference oracle into a behavioral-gate Run. Three targets: a compiled genie bundle (bundle_artifact_id, runs on-device via the self-hosted runner), an on-device LLM (geniex_model), or an API or workflow endpoint (endpoint_id/http — n8n, Zapier, Make, OpenAI-compatible; EdgeGate executes it by default, so nothing is installed, and results are tagged API-verified rather than hardware-certified). Errors with a mismatch hint when the reference was captured against a different eval-set version (eval_set_sha256 differs) or a different endpoint. Requires workspace admin access.
edgegate_cancel_runCancel a non-terminal run, freeing the workspace's single active-run slot. Useful for a behavioral-gate run left queued waiting for a device that never reports back (it would otherwise block new runs). 409 if the run is already terminal.
edgegate_rerun_bgRe-run an existing behavioral-gate run: clones its already-validated config (same bundle + eval set + reference + system prompt + device) into a fresh queued run — no need to re-supply artifact ids. 409 if the workspace already has an active run (cancel it first).
edgegate_setup_bg_github_actionGenerate the Behavioral-Gate GitHub Actions setup — the self-hosted-runner prerequisites, the workflow YAML (using the edgegate-bg composite action), and the gh secret commands. Unlike the standard run gate, BG runs the model on a real device, so it needs a self-hosted runner with the device attached.
edgegate_export_compliance_reportExport the compliance-preset report for a run (e.g. ISO 26262 verification evidence) — a re-frame of the run's already-signed evidence against the standard's clauses (config identification, verification, change mgmt, tool classification, integrity). Verification evidence, NOT a compliance certification. The formatted assessor PDF is on the dashboard run page.
edgegate_export_field_recorder_reportExport the Field Recorder EU AI Act Article 12 record-keeping report for a date range — the signed, hash-chained evidence of what deployed models did on-device (clause 12(1)/12(2)(a)/12(2)(b)/integrity mapping, event summary, Ed25519 signature). Verification evidence, NOT a compliance certification. The formatted assessor PDF is on the Field Recorder dashboard page.
edgegate_recorder_statusField Recorder status for a workspace: recorded-event counts, replay/divergence breakdown (passed/diverged/no_reference), hash-chain integrity (verified, gaps, signing keys), and the devices reporting in. Read-only.
edgegate_trigger_replayReplay pending recorded events against their certified reference (a PASSED gate run for the same model) plus the input-matched baseline. Verdicts land asynchronously — poll edgegate_recorder_status. Events with no certified reference resolve to no_reference.

Skills

Slash commands you can invoke directly:

  • /edgegate-init — full onboarding flow (zero → CI gate)
  • /edgegate-gate — trigger a run on an existing pipeline
  • /edgegate-status — check a run's status + metrics
  • /edgegate-audit — fetch the evidence bundle for a run
  • /edgegate-compare — diff two runs (auto-baseline) with REGRESSION/IMPROVEMENT/NEUTRAL verdict
  • /edgegate-export — save a run report as a markdown file (for PR comments, Slack, compliance)
  • /edgegate-import — import a Hugging Face ONNX model (anonymous, or workspace-token for private / gated / qualcomm-org repos)
  • /edgegate-connect-huggingface — attach a personal HuggingFace token to the workspace so the import flow can read private / gated / qualcomm-org repos
  • /edgegate-connect-qaihub — connect a Qualcomm AI Hub token so runs can actually compile and profile on Snapdragon devices
  • /edgegate-workspace-setup — bootstrap a brand-new workspace end-to-end (create → connect AI Hub → API key → invite teammates)
  • /edgegate-members — list / invite / change role / remove workspace members
  • /edgegate-promptpacks — list existing promptpacks or create a new one with generated test cases
  • /edgegate-byo-storage — Enterprise BYO storage onboarding (register grant → paste External ID → verify probe → first artifact → first run → audit log)

License

MIT — see LICENSE.

來源:README.md,提交 e47608a

工具

0
SourceWeft 尚未在沙箱中啟動過該軟體包,因此其工具未知。

版本歷史

1
  1. v0.19.0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