OPEN SOURCE · APACHE 2.0

A framework forproductionAI agents.

Zil is an open-source CLI and Python SDK for validating, packaging, and deploying AI agents. Open source. Composable. Deploys anywhere.

Get Started →DocumentationJoin Slack →
EtymologyAg → silver → zilver → Zil
Composes withADK · MCP · DeepEval · OpenTelemetry
Statusv0.1 · Early preview
01 — Why Zil

There is no standard way to ship an AI agent.

Most agent projects jump from notebook to production with no manifest, no eval gate, no signed artifact, and no security audit. The agent works in a demo — then breaks in ways that compound once real users, real data, and real tools are involved.

Zil fills this gap. A declarative manifest (manifest.yaml) defines your agent's identity, adapters, tools, evals, and environment. A CLI validates, audits, evaluates, packages, and deploys — so every agent ships as a signed, portable .zil archive through a repeatable pipeline.

02 — The CLI

One CLI. Seven commands. Init to deployed.

Each command does one thing well. Together they form a complete pipeline from project scaffolding to production deployment.

  1. 01

    zil init

    Scaffold a complete agent project — manifest, identity, adapters, evals, guardrails, Dockerfile, CI pipeline. Choose your LLM provider and MCP preset.

  2. 02

    zil validate

    Check manifest schema, file references, env var declarations, MCP server config, and guardrail structure. Catch misconfigurations before they reach production.

  3. 03

    zil audit

    Security scan: prompt injection resilience, PII leakage, indirect injection surface, guardrail coverage scoring. Actionable findings, not compliance theater.

  4. 04

    zil eval

    Run evaluation suites with LLM-as-judge metrics — answer relevancy, faithfulness, contextual recall. Gate promotions on quality thresholds.

  5. 05

    zil pack

    Build a signed .zil archive: manifest, agent code, MCP tools, SBOM, eval results — one portable artifact. Cosign signature and SLSA provenance included.

  6. 06

    zil push

    Push the archive to any OCI-compatible registry — Google Artifact Registry, GHCR, ECR, Docker Hub. Same tooling you already use for containers.

  7. 07

    zil deploy

    Deploy to Cloud Run from source or from a registry artifact. MCP server host dependencies, env vars, and tracing — handled automatically.

03 — The artifact

.zil — a signed, portable agent bundle.

Every agent ships as a single signed archive containing its manifest, identity, adapters, MCP tools, evaluation results, and SBOM — packaged together, versioned together, deployed anywhere.

Signed bycosign / Sigstore
ProvenanceSLSA Level 3
DistributionOCI registry
Deployable toCloud Run + more
# my-agent-1.0.0.zil
apiVersion: zil/v1
kind: Agent
metadata:
  name: my-agent
  version: 1.0.0

spec:
  runtime:
    framework: adk
    language: python
    llm:
      adapter: ./adapters/llm.yaml
  identity: ./identity
  evals: ./evals
  tools: ./tools
  observability: ./observability
  env:
    # declared, validated, injected at deploy

# bundled attestations
SBOM.cyclonedx.json
EVAL_RESULTS.json
BUILD_META.json
04 — Provenance

Built by FluentData.

Zil is built and maintained by FluentData. Born from real production agent work, open-sourced under Apache 2.0. We build production AI agents for clients — Zil is the toolchain we use.

About FluentData ↗Get in touch →