Blekline — Plano de Controle de Ingresso para Agentes IA
Blekline/blekline-oss
Govern, mask, and audit every agent call before it touches your systems
Ver no GitHubMCP server que funciona como um plano de controle de ingresso para agentes de IA, implementando mascaramento de dados sensíveis, aplicação de políticas de acesso e auditoria completa de chamadas. Coloca governança entre seus agentes e as ferramentas que eles acessam, garantindo conformidade com regulações como a Lei de IA da UE.
Instalação
npx convez-ai Blekline/blekline-oss1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Proteger dados sensíveis (credenciais, PII) de serem expostos em contextos de modelo
- Implementar políticas de acesso granular para agentes autônomos
- Auditoria e rastreamento completo de chamadas de ferramenta para compliance regulatório
- Prevenir execução não autorizada de comandos críticos em bases de dados e APIs
Veredito

Solução madura e essencial para times que deployam agentes IA em produção. Endereça o gap crítico de governança entre agentes e recursos, com suporte a EU AI Act e compliance. Arquitetura sólida (proxy + SDK) e open core tornam viável desde POCs até enterprise.
README
A tradução para o português deste README ainda está na fila.
The problem nobody wants to talk about
Cursor writes your code, Claude answers your support tickets, autonomous pipelines touch your databases, your APIs, your customers' data. The ecosystem is accelerating, MCP servers let agents pick up tools like apps pick up plugins, and that is genuinely exciting.
But here's the thing: your agents have no idea what they're not allowed to do. They'll happily pass an AWS key to a model context window. They'll call a tool with a customer's email as an argument. They'll execute a shell command that wasn't in the plan. Not out of malice, out of the fundamental nature of language models: they optimize for task completion, not for the organizational policies you haven't written yet.
This is the AI governance gap. And right now, there's nothing sitting between your agents and everything they can touch.
Why this is becoming urgent
The EU AI Act isn't theoretical anymore. GPAI obligations have been enforceable since August 2025. Transparency and human oversight requirements land in August 2026. High-risk system conformity assessments follow. Fines reach up to €35 million or 7% of global turnover for the worst violations. And these rules aren't just about the model, they're about the systems you build with them: how you govern tool access, how you audit decisions, how you prove a human was in the loop.
Meanwhile, enterprises running AI at scale, sandboxed, parallelized, thousands of agent calls, have no native answer for: what happened in that session? Who authorized that tool call? Did any PII leave the context window?
The compliance question is catching up to the capability question. And most teams aren't ready.
What Blekline is
Blekline is an open-core MCP ingress control plane, infrastructure that sits between your agents and everything they can touch. It does three things, in real time, before any LLM sees a prompt or any tool executes:
Mask - strip PII, secrets, and sensitive context from prompts before they hit model APIs (MCP Server docs)
Enforce — evaluate tool calls against policy; allow, flag, or block before execution
Audit — emit a structured, tamper-evident event trail for every agent interaction
You can run it locally in two minutes. You can deploy it as a sidecar alongside any L1 sandbox (Daytona, Modal, E2B, Cloudflare, Vercel Sandbox). You can plug it into Cursor, Claude Desktop, or Codex today, without changing your agent code. This is the infrastructure that makes governed AI deployment real: not a checkbox, not a policy document, but a running system that enforces your intentions at the call level.
Start here
pnpm install && pnpm build:packages
export BLEKLINE_WORKSPACE_TOKEN="blw_..."
export BLEKLINE_API_URL="https://app.blekline.com"
export BLEKLINE_CLIENT_SURFACE="sdk"
pnpm demo:mcp-smoke
Headless guide: cli/README.md · CI template: ci/
Connect Blekline
| Surface | Path | BLEKLINE_CLIENT_SURFACE |
|---|---|---|
| CLI / SDK | cli/ | sdk |
| CI / CD | ci/ | sdk |
| Claude Code | .claude/settings.json.example | claude-code |
| Cursor | .cursor/mcp.json.example | cursor |
| GitHub Copilot | .vscode/mcp.json.example | github-copilot |
| Continue | .vscode/continue.config.json.example | continue |
| Claude Desktop | config/claude_desktop_config.json.example | claude-desktop |
| Codex | .codex/config.toml.example | codex |
Full index: integrations/README.md · Docs: app.blekline.com/docs
pnpm generate:mcp-configs # *.example configs
pnpm verify:integrations # manifest + schema checks
In any client: "Use blekline_mask_prompt on: Contact Jane at jane@acme.com — API key AKIAIOSFODNN7EXAMPLE"
Architecture
Blekline sits at Layer 4 — between L5 agents (Cursor, Claude, Codex) and L1 sandboxes (Daytona, Modal, Vercel Sandbox, Cloudflare, E2B) and model APIs.
L5 Agents → L4 Blekline (mask · enforce · audit) → L1 sandbox MCP / model APIs
AI Enablement Stack · Architecture · Trust boundaries · Latency SLO
Open core vs cloud
| Capability | OSS (this repo) | Cloud (app.blekline.com) |
|---|---|---|
| MCP server / proxy | Yes | Yes |
| Local tool + secret enforce | Yes (@blekline/contracts) | Yes |
| Azure authoritative PII mask | — | Yes |
| Workspace fleet policy (SSE) | — | Yes |
| Investigations / billing | — | Yes |
License: AGPL for proxy/server (self-host or buy cloud). Apache for contracts/SDK (embed in your agent stack).
Who this is for
Developers building with Cursor, Claude Desktop, or Codex who want their agents to stop leaking secrets and start respecting tool boundaries.
Platform teams deploying AI workloads in L1 sandboxes who need a governance layer that travels with the execution environment.
Enterprise architects preparing for EU AI Act compliance — specifically human oversight, audit trails, and tool call transparency requirements that become enforceable in August 2026.
Open source contributors who believe that the infrastructure for safe AI should be auditable, forkable, and owned by the community — not locked inside a vendor's cloud.
Packages
| Package | Install | License |
|---|---|---|
@blekline/mcp-server | npm i @blekline/mcp-server | AGPL-3.0 |
@blekline/mcp-proxy | npm i @blekline/mcp-proxy | AGPL-3.0 |
@blekline/client | npm i @blekline/client | Apache-2.0 |
@blekline/contracts | workspace / embed | Apache-2.0 |
ingress-proxy | Docker / Helm | AGPL-3.0 |
OpenAPI: packages/contracts/openapi.yaml
MCP tools
| Tool | Purpose |
|---|---|
blekline_mask_prompt | Redact PII / secrets before model context |
blekline_classify_risk | Risk tier → allow / review / block |
blekline_evaluate_tool_call | Policy on tool name + arguments |
blekline_emit_event | Metadata audit trail |
Proxy path: agent → Blekline → allow/mask/block → downstream MCP (Daytona, E2B, Modal, Cloudflare, Vercel Sandbox, custom).
Client libraries
TypeScript
npm install @blekline/client
import { BleklineClient } from "@blekline/client";
const blekline = new BleklineClient({
workspaceToken: process.env.BLEKLINE_WORKSPACE_TOKEN!,
metadata: { clientSurface: "sdk" },
});
await blekline.mask({ text: "alice@corp.com", platform: "MyAgent" });
await blekline.enforceToolCall({
toolName: "run_shell",
arguments: { cmd: "curl https://api.internal/deploy" },
});
Python
pip install blekline-client
Local-only (no API token)
import { enforceToolCallLocally, scanTextForSecrets } from "@blekline/contracts";
scanTextForSecrets("export AWS_KEY=AKIAIOSFODNN7EXAMPLE");
enforceToolCallLocally({
toolName: "run_shell",
arguments: { cmd: "export AWS_KEY=AKIAIOSFODNN7EXAMPLE" },
requestId: "local-1",
});
Works with
Full integration guides on app.blekline.com/docs — not mirrored in this repo.
L5 agent clients
| Client | Guide |
|---|---|
| Hub | Agent clients |
| Continue | Continue MCP |
| GitHub Copilot | Copilot MCP |
| OpenHands | OpenHands |
| Sourcegraph Cody | Cody |
L2 model providers
| Provider | Guide |
|---|---|
| Hub | Model providers |
| Azure OpenAI | Azure stack |
| AWS Bedrock | Bedrock stack |
| OpenRouter | OpenRouter stack |
L2 frameworks & RAG
| Integration | Guide |
|---|---|
| Hub | Frameworks & RAG |
| LangChain | LangChain stack |
| Pinecone | Pinecone stack |
L3 eval & safety
| Partner | Guide |
|---|---|
| Hub | Eval & safety |
| LangSmith | LangSmith stack |
| Guardrails | Guardrails stack |
L1 sandboxes
| Provider | Integration guide |
|---|---|
| All five | Sandbox providers hub |
| Daytona | Daytona stack |
| Modal | Modal stack |
| Vercel Sandbox | Vercel Sandbox stack |
| Cloudflare | Cloudflare stack |
| E2B | E2B stack |
Deploy
| Mode | Command / link |
|---|---|
| MCP (global) | npx -y @blekline/mcp-server |
| Edge sidecar | pnpm docker:ingress — Helm |
| L1 sandboxes | Sandbox providers |
Development
Client demos and smoke tests: demo/README.md.
git clone https://github.com/Blekline/blekline-oss.git && cd blekline-oss
pnpm install && pnpm build:packages && pnpm demo:mcp-smoke
Documentation
All docs: app.blekline.com/docs
| Doc | Link |
|---|---|
| Quick start | introduction/quick-start |
| EU AI Act & compliance | introduction/eu-ai-act |
| AI Enablement Stack | introduction/ai-enablement-stack |
| MCP proxy | mcp/proxy |
| Cursor setup | mcp/cursor |
Community & design partners
Questions or integration feedback — GitHub Discussions (Q&A, integration feedback).
Design partners — shipping agents in production and want help wiring Blekline into your stack (proxy or MCP server)?
- Open a design partner issue
- Or start a discussion — choose Design partner interest
Contributing
CONTRIBUTING.md · SECURITY.md · CHANGELOG.md
Private Blekline team: develop in the blekline monorepo, run pnpm sync:oss from the root repo.
License
| Component | License |
|---|---|
mcp-server, mcp-proxy, ingress-proxy | AGPL-3.0 |
contracts, client, client-python | Apache-2.0 |
Managed SaaS at app.blekline.com is not licensed under this repository.