Pre-MVP · local-first · open source

Capture context.
Compound knowledge.

CarpeOS keeps the decisions, evidence, and failure paths from AI-agent work in one place you control — with provenance intact and status explicit.

Install CarpeOS Explore the source

Public code. Private knowledge. Not hosted SaaS.

A crystalline knowledge core connected to a network of smaller nodes
accepted Decision with provenance
retrievable Context for the next agent
Append-only events Explicit acceptance Local by default Rebuildable projections
Why CarpeOS

Your most valuable context should not disappear with the session.

A real decision, a rejected approach, or a bug path often ends up split across chat history, terminal scrollback, and notes. CarpeOS keeps the trail without flattening every statement into “memory.”

01

History without trust

Agent transcripts mix facts, guesses, drafts, and abandoned paths.

Keep evidence, claims, acceptance, and supersession as separate records.
02

Knowledge in silos

Each coding tool starts with a different memory of the work.

Capture into a provider-neutral event trail and retrieve through shared interfaces.
03

Notes become authority

Generated documents drift until no one knows what was canonical.

Treat notes, indexes, and context packs as rebuildable projections.
04

Continuity breaks

The next machine or session has to rediscover the same decisions.

Use a local-first store with optional private, operator-run sync.
How it fits

Capture once. Reuse with the trail attached.

Provider hooks feed a private canonical event store. Accepted facts are derived at query time; MCP, CLI, and Obsidian remain interfaces and projections.

CarpeOS flow from agent hooks through capture, canonicalization, retrieval, and MCP, CLI, and Obsidian interfaces
  1. 1

    Capture

    Selected lifecycle events enter an encrypted outbox and local store.

  2. 2

    Canonicalize

    Evidence, observations, claims, decisions, and supersession keep distinct roles.

  3. 3

    Retrieve

    Search and context packs return bounded results with provenance rechecked.

  4. 4

    Use

    People and agents work through CLI, local MCP, and rebuildable projections.

Local interfaces

The same knowledge trail, available where work happens.

Current interfaces are local and pre-MVP. They expose the same underlying model without turning a projection into the source of truth.

CLI

Search, retrieve, and build context packs.

Use local commands for memory search, individual records, and bounded task context.

carpeos memory context-pack
MCP · stdio

Eight local tools for agents.

Trace evidence, capture observations, propose claims, and retrieve context through a local stdio server.

memory_context_pack
Obsidian projection

Readable notes without changing authority.

Generate Markdown from the local store. Delete and rebuild it whenever the projection drifts.

projection ≠ source of truth
Trust model

A model suggestion is not automatically a fact.

CarpeOS preserves the difference between what was observed, what was proposed, what was accepted, and what was later replaced.

  • Accepted status is derived at query time.
  • After acceptance, the event log is append-only.
  • Sensitive plaintext stays outside the event body.
  • Trust zones are isolation boundaries, not decorative labels.
Evidencewhat can be inspected
Observationwhat was seen
Claimwhat is proposed
Acceptancewhat was decided
Supersessionwhat replaced it
Accepted facts, derived with provenance
Get started

Install locally. See the plan before applying it.

CarpeOS requires Node.js 22.22 or newer. The setup flow can show resolved paths and actions before it changes anything.

Open the docs hub Full install guide
terminal 
# install the published package
npm install -g @innocarpe/carpeos

# inspect actions first
carpeos setup plan

# apply the resolved setup
carpeos setup run --apply

Pre-MVP. Local interfaces are implemented; no hosted Worker, hosted MCP, or SaaS deployment is claimed.

Documentation

Start with the model. Go deeper when you need to.

The website stays intentionally small. Detailed contracts, guides, ADRs, and implementation status live with the public source.

Website guide

Documentation hub

A focused map of architecture, installation, capture, retrieval, and MCP references.

Open locally System design

Architecture overview

Follow the event model, trust boundaries, storage layers, and projection rules.

Read on GitHub Agent interface

MCP tool contract

Review the local stdio tools and the contracts agents can rely on.

Read on GitHub
Keep the trail

Give the next session more than a transcript.

Install CarpeOS View source