Skip to content

arikyp/agentcapsule

Repository files navigation

Agent Capsule

Tests PyPI License

Agent Capsule is the verifiable handoff format for agents.

It is a small open protocol + developer toolkit for safe agent handoffs. It wraps exact machine-readable payloads in a text-native envelope so receivers can detect, verify, policy-check, and unpack safely.

2-Minute Proof

python3 -m pip install agentcapsule
agentcapsule pack handoff.json --out capsule.txt
agentcapsule ingest thread.txt --out ./sandbox --strict --json

If ingest exits 0, the handoff passed verification/policy and unpacked safely. If ingest exits non-zero in --strict, treat it as a CI/governance failure.

One Command, One Function

CLI:

agentcapsule ingest thread.txt --out ./sandbox --policy ./policy.json --json --strict

Python:

from agentcapsule import ingest_messages

result = ingest_messages(
    messages=thread_messages,
    out_dir="./sandbox",
    policy="./policy.json",
)

print(result.inline_capsules)
print(result.references)
print(result.unpacked_files)

What It Solves

Normal agent channels are lossy for machine payloads (truncation, formatting drift, silent edits). Agent Capsule turns handoffs into verifiable artifacts.

Protocol Layers

  • Capsule: exact payload bytes + metadata + hash/signature context.
  • Envelope: text wire format with boundary markers, headers, and encoded payload.
  • Manifest: handoff intent (creator, task, files, capabilities, policy hints).
  • Delivery modes: inline, attachment, reference (URI + capsule hash + payload hash).

Delivery Modes

  • inline: full capsule in message body.
  • attachment: full capsule as file/blob.
  • reference: descriptor in message, full capsule fetched by URI.

Reference descriptors are not authoritative by themselves. Receivers must fetch the full capsule and verify capsule_sha256, payload_sha256, signature trust policy, and receiver policy.

Not A Transport Platform

Agent Capsule does not replace transport. It travels through existing systems: chat, tickets, email, GitHub, A2A/MCP workflows, and object storage.

Security And Trust Model

Baseline:

  • SHA256 payload integrity checks.
  • Local policy checks.
  • Safe unpacking into a chosen output directory.

Optional hardening:

  • HMAC-SHA256 signatures.
  • Ed25519 signatures and trust registry checks.
  • AES-256-GCM payload encryption.
  • Zstandard compression.
  • Resumable reference fetching.

Current Limitations

  • No hosted trust service: signature trust resolution is local-file policy/registry driven.
  • No remote/global key-discovery protocol yet: receivers must supply local trust inputs.
  • No first-party JS/TS reference implementation yet.
  • Governance output is JSON-first; there is no built-in long-running dashboard service.
  • Reference fetching requires optional install extras (agentcapsule[fetch] or agentcapsule[all]).

Typical Flow

  1. Sender packs payload into a capsule.
  2. Sender transports inline/attachment/reference.
  3. Receiver scans and ingests.
  4. Receiver verifies metadata, hashes, signature trust, and policy.
  5. Receiver unpacks verified payload into sandbox.
  6. Receiver runs downstream logic on unpacked files.

Install

PyPI:

python3 -m pip install agentcapsule

Full optional capabilities:

python3 -m pip install "agentcapsule[all]"

Reference fetching support only:

python3 -m pip install "agentcapsule[fetch]"

Docs

About

Exact, verifiable payload capsules for agent-to-agent handoffs over chat, email, tickets, and A2A.

Topics

Resources

License

Contributing

Stars

Watchers

Forks

Packages

 
 
 

Contributors