agentpk

The open source CLI and Python SDK for packaging AI agents.

pip install agentpk

Quickstart

pip install agentpk

Both agentpk and agent are installed. agentpk is the canonical form.

agentpk init my-agent
# edit my-agent/manifest.yaml
agentpk pack my-agent/

That's it. You now have a portable my-agent-0.1.0.agent file you can share, deploy, or register.

# Run it
agentpk run my-agent-0.1.0.agent

# Sign it
agentpk keygen --out my-key.pem
agentpk sign my-agent-0.1.0.agent --key my-key.pem

What is the .agent format?

An .agent file is a portable archive containing your agent source code, a manifest.yaml that describes what your agent does and what it needs, and a checksums.sha256 file that verifies nothing was tampered with.

The manifest is the important part. It tells runtimes how to start your agent and tells registries how to list it. One file, two audiences.

Python SDK

All CLI operations are available as typed Python functions:

from agentpk import pack, analyze, validate, inspect_package, init

# Pack an agent
result = pack("./my-agent", analyze=True)
print(result.trust_score)    # 87
print(result.trust_label)    # "High"
print(result.package_path)   # PosixPath('./dist/my-agent-1.0.0.agent')

# Analyze without packing
analysis = analyze("./my-agent", levels=[1, 2, 3])
print(analysis.discrepancy_count)   # 0

# Validate
val = validate("./my-agent")
print(val.valid)    # True

# Scaffold a new project
r = init("my-node-agent", runtime="nodejs")
print(r.project_dir)   # PosixPath('./my-node-agent')

All functions return typed dataclasses. Errors raise typed exceptions (AgentpkError, ManifestError, PackagingError, AnalysisError, PackageNotFoundError) — no sys.exit(), no string parsing.

Multi-language support

agentpk packages agents written in any language. The manifest declares the runtime; analysis depth depends on the language:

Language	Analysis	Extractor
Python	Full AST	stdlib ast module
Node.js	Full AST	acorn (via bundled helper)
TypeScript	Full AST	@typescript-eslint/parser
Go	Pattern-based	Regex on source text
Java	Pattern-based	Regex on source text
Other	Structural only	Level 2 skipped, reason logged

Scaffold for any runtime:

agentpk init my-node-agent --runtime nodejs
agentpk init my-go-agent --runtime go
agentpk init my-java-agent --runtime java
agentpk init my-ts-agent --runtime typescript

Naming convention

Agent names must be lowercase with hyphens and digits only. They must start with a letter.

Valid	Invalid
fraud-detection	Fraud_Detection
my-agent-2	my agent
data-pipeline	data.pipeline

CLI commands

Command	Description
agentpk init <name>	Scaffold a new agent project
agentpk pack <dir>	Pack a directory into a .agent file
agentpk validate <target>	Validate a .agent file or project directory
agentpk inspect <file>	Display metadata and AIR bundle from a .agent file
agentpk unpack <file>	Extract a .agent file to a directory
agentpk diff <old> <new>	Show differences between two .agent files
agentpk test	Run built-in self-tests (22 cases)
agentpk generate [dir]	Generate a manifest.yaml from code analysis
agentpk list [dir]	List all .agent files in a directory
agentpk run <file>	Execute a packed .agent file as a subprocess
agentpk sign <file>	Sign a .agent file with a private key
agentpk verify <file>	Verify the signature on a .agent file
agentpk keygen	Generate an Ed25519 key pair for signing
agentpk serve	Start the REST API and packaging UI

REST API and packaging UI

Package and certify agents from a browser or remote system without the CLI:

pip install agentpk[api]
agentpk serve
# API on http://localhost:8080
# Packaging UI on http://localhost:8080

The UI lets you select an agent folder directly from your browser, runs analysis, and returns a trust score with a download link — no terminal required. The UI automatically detects whether an LLM API key is configured and enables or disables Level 3 accordingly.

Via any HTTP client:

# Submit a packaging job
curl -X POST http://localhost:8080/v1/packages \
     -F "source=@my-agent.agent" \
     -F "analyze=true" \
     -F "levels=1,2,3"

# Poll for completion
curl http://localhost:8080/v1/packages/{job_id}

# Download the .agent file
curl http://localhost:8080/v1/packages/{job_id}/download -o my-agent.agent

Options:

agentpk serve --port 9000
agentpk serve --host 127.0.0.1
agentpk serve --reload          # dev mode

Listing agents

agentpk list
agentpk list ./agents/
agentpk list ./agents/ --recursive
agentpk list ./agents/ --json

Running agents

agentpk run my-agent-1.0.0.agent
agentpk run my-agent-1.0.0.agent --dry-run
agentpk run my-agent-1.0.0.agent --keep
agentpk run my-agent-1.0.0.agent --env API_KEY=abc123
agentpk run my-agent-1.0.0.agent -- --flag value

The runner extracts the package to a temp directory, validates it, and launches the entry point using the runtime declared in the manifest. Extra arguments after -- are forwarded to the agent process.

Flag	Effect
--dry-run	Validate and extract without executing
--keep	Keep the temp directory after execution
--env KEY=VALUE	Set environment variables (repeatable)

Warning: agent run executes code from the package. Only run agents from sources you trust.

Code analysis and trust scores

agentpk can analyze agent source code and assign a trust score indicating how well the manifest matches what the code actually does.

See TRUST.md for the full trust score reference and docs/agent_analyzer.md for the analysis architecture.

Generating a manifest from code

agentpk generate ./my-agent
agentpk generate ./my-agent --level 3

The generated manifest includes # REVIEW markers on fields that could not be determined from code analysis alone.

Packing with analysis

agentpk pack my-agent/ --analyze
agentpk pack my-agent/ --analyze --level 3
agentpk pack my-agent/ --analyze --level 3 --strict

Flag	Effect
--analyze	Run code analysis before packing
--level N	Analysis depth 1-4 (default: auto)
--strict	Fail if requested level cannot be reached
--on-discrepancy warn|fail|auto	Discrepancy handling (default: warn)
--memory	Bundle an AIR memory snapshot with the package
--memory-components	Comma-separated component list (default: all)

Analysis levels

Level	Source	Needs	Weight
1	Structural validation	Nothing	+20 pts
2	Static analysis (AST or pattern-based)	Nothing	+30 pts
3	LLM semantic analysis	API key	+25 pts
4	Runtime sandbox	Container runtime	+25 pts

Skipped levels subtract points (Level 3 skip: -15, Level 4 skip: -25). The maximum score is 100 when all four levels pass with no discrepancies.

Trust score labels

Score	Label
90-100	Verified
75-89	High
60-74	Moderate
40-59	Low
0-39	Unverified

Portable agent memory (AIR)

Pack an agent with its accumulated intelligence:

agentpk pack my-agent/ --memory
agentpk pack my-agent/ --analyze --memory
agentpk pack my-agent/ --memory --memory-components fingerprint,trust,org_context

The --memory flag bundles an AIR (Agent Intelligence Record) snapshot alongside the package. AIR is an open standard for portable agent memory — behavioral history, trust trajectory, organizational context, and distilled insights in platform-agnostic JSON schemas.

The snapshot lives in _package.air in the packed manifest and in intelligence/ inside the archive. A receiving platform can rehydrate the agent's behavioral state without rebuilding it from scratch.

Full intelligence export requires pip install agentpk[memory]. Without it, a spec-compliant stub is embedded instead.

See AIR.md for the full specification.

Interactive pack mode

Run agentpk pack in a terminal with no flags and the CLI walks you through environment detection, analysis level selection, and memory bundling interactively:

agentpk pack ./my-agent

The interactive flow detects available API keys and container runtimes, presents guided options, shows live progress, and prints a summary with next steps. It activates automatically in a terminal and is disabled when piped, when explicit flags are passed (--analyze, --memory), or with --no-interactive.

Signing and verification

Generate a keypair

agentpk keygen --out my-key.pem

Creates two files:

• my-key.pem — Ed25519 private key (keep secret, do not commit)
• my-key.pub.pem — Ed25519 public key (share with recipients)

Sign an agent

agentpk sign fraud-detection-1.0.0.agent --key my-key.pem
agentpk sign fraud-detection-1.0.0.agent --key my-key.pem --signer "Acme AI"

Produces fraud-detection-1.0.0.agent.sig — a JSON file containing the manifest hash, Ed25519 signature, algorithm identifier, and optional signer metadata.

Verify a signature

agentpk verify fraud-detection-1.0.0.agent --key my-key.pub.pem

Manifest structure

The manifest has two zones:

Zone 1 (open core) — authored by the developer: identity, runtime, capabilities, permissions, execution settings, and resource requirements.

Zone 2 (_package) — generated automatically at pack time: hashes, timestamps, file counts, and package size. Never edit by hand.

Validation

agentpk validate ./my-agent/
agentpk validate my-agent-1.0.0.agent
agentpk validate my-agent-1.0.0.agent --verbose

The --verbose flag displays a per-stage breakdown. Directories skip stages 5-6 (checksums and package integrity) since those only apply to packed files.

Verifying your installation

agentpk test
agentpk test --verbose

Examples

Seven valid examples and fourteen intentionally broken examples in examples/.

agentpk pack examples/valid/fraud-detection
agentpk pack examples/valid/fraud-detector-with-memory --memory
agentpk inspect fraud-detector-with-memory-1.0.0.agent
agentpk pack examples/invalid/04-invalid-name

The memory examples demonstrate AIR bundling (fraud-detector-with-memory, healthcare-agent-strict-redaction) and AIR validation failure modes (memory-hash-mismatch, memory-missing-component, memory-malformed-air-json). See examples/README.md for the full index.

Specification

See SPEC.md for the full agent package format specification.

Development

pip install -e ".[dev]"
pytest

# With API extras
pip install -e ".[dev,api]"
pytest tests/test_api.py

Core dependencies: click, pyyaml, pydantic, rich, cryptography. API extras: fastapi, uvicorn, python-multipart.

About

Built by Nomotic AI.

License

Code: MIT Specification (SPEC.md): CC BY 4.0