When the unit of intelligence is portable, context travels across tools, teams, and time.
Any model that can call tools or skills can be a runtime.
Latest · 2026-05 · v0.6 prototype
Spec v0.6 + two reference examples + the medical-journal Kaggle submission landed this week.
Spec · SDKs · Reference example · Kaggle
Start here
Pick the page that matches the job: install the runtime layer, teach an agent the capsule format, read the research, or test a domain workflow.
A capsule is not simply a document. It is a portable unit of collaborative work,
combining content, state, and a verifiable event ledger in a single .capsule file. Any actor can
open it, validate it, continue the work, and hand it off.
The human-readable work product
Machine-readable snapshot of current context
Append-only event history proving how it evolved
Capable of continuing work wherever it is opened
Multiple actors contribute with full attribution
History matters as much as current state
Resumable across tools, teams, and time
Cryptographic proof of who did what and when
Carries logic that actors follow, not just data to read
In any modern LLM with coding capabilities.
Install the Capsule skill. Your LLM handles the full lifecycle: reading, appending, verifying, and handing off.
Get the skill →Connect via the Capsule MCP server. Works in any MCP-compatible client with full protocol access.
Install →Drop a .capsule into any LLM with a short bootstrap prompt. No tooling required. Works
anywhere.
Capsules ship as templates: named formats aligned to real workflows. Each one specifies fields, steps, rules, and permitted AI interactions. The protocol stays underneath; teams interact with the shape of the work.
Real domains. Real workflows. Nine experiments testing the protocol under conditions that matter: compliance, healthcare, journalism, security, and more.
A capsule moves through a lifecycle. Any actor, human or AI, enters at the appropriate stage and continues the work from there.
Capsules represent a shift from documents to work artifacts.
Traditional systems scatter context across documents, databases, software, email, and human memory. Capsules consolidate this into portable artifacts that carry their own history, structure, and execution context.
surface.md
Inside every capsule, surface.md is the object the human and the AI are working toward:
the readable result of the chain. Today it’s markdown, because markdown is the lowest common denominator
that every model already speaks. That choice is intentional, and it is also temporary.
The surface could be anything in the future: code, libraries, live artifacts, HTML, programs, journals, images, any media type that requires provenance. Capsules could even become hash-less for some use cases, or remain hash-rooted for v1.0 while we work through the security and gaps in today’s LLM architectures. We are starting with the primitives on purpose.
Once the bones are in, the morphic nature of open source will evolve .capsule into a new file
type for LLM-to-LLM, human-in-the-loop, intelligence-to-intelligence collaboration: verifiable, encryptable,
event-rich, context-aware documents that travel between minds, models, runtimes, and other capsules.
Capsules are verifiable, encryptable, event-rich intelligence documents. They begin as markdown-producing skill artifacts, but are designed to become media-agnostic containers for collaborative work between humans, models, runtimes, and other capsules.
The artifact becomes the unit of collaboration.