eMachina

The domain-neutral knowledge loop — harvest → curate → retrieve → feedback — as a machine any actor can use: agents, people, and services stop re-learning what's known, and what's known keeps getting better from use. Charter: CHARTER.md.

Two layers, three install channels:

| Layer | What it is | Ships as | | ----- | ---------- | -------- | | Skills (the thinking) | skills/ — agent-skills packages (spec); flagship: harvest | plugin marketplace · npx skills add | | Substrate (the executable) | src/mcp/ — the machina MCP server: yield management, curate (memorialize, charter-gated), retrieve (offline corpus recall), over an owned minimal atom kernel + markdown corpus | npm: @ema.co/machina |

Install

Everything (skills + MCP, recommended) — Claude Code:

/plugin marketplace add Ema-Unlimited/eMachina
/plugin install emachina@ema

The plugin's .mcp.json auto-launches the MCP server from npm (npx -y @ema.co/machina@latest) — nothing to build. Cursor: open a clone (skills load from .agents/skills/), or install via Cursor's marketplace when published.

Skills only — any supported agent, via skills.sh:

npx skills add Ema-Unlimited/eMachina

MCP server only — any MCP client:

{ "mcpServers": { "machina": { "command": "npx", "args": ["-y", "@ema.co/machina@latest"] } } }

Tools: config (readiness) · yield (validate/list/get/scaffold harvest yields) · curate (publish/advise — content-addressed dedup, blocking validation, charter gate) · retrieve (search the local corpus + yields, offline, audience-filtered). Server docs: src/mcp/README.md.

Skills

| Skill | What it does | Docs | | ----- | ------------ | ---- | | harvest | Universal context harvester — gather/synthesize context from files, URLs, transcripts, CRM, web, or tools into one contract-valid harvest-yield.md (findings only; derivations are projections). Invoke with /harvest. | Architecture · Charter |

The knowledge model (one paragraph)

Knowledge lives as atoms — markdown docs with a minimal frontmatter manifest (schema): content-addressed identity (identical content ⇒ identical id ⇒ dedup by construction), a single fail-closed audience ACL token, active|retracted lifecycle. The local corpus (.knowledge/atoms/, gitignored) is the staging tier; the org tier (GCS + Discovery Engine) is consumed at runtime via MCP interfaces — never a build dependency. Versioning has no lineage pointers: re-curating a changed artifact under the same (kind, name) tombstones the prior version. Charters (kind: charter) are gated at the write path — no mission/owner/status, no entry.

Develop on this repo

1. Clone and open — packages under skills/ are live via the symlink hub; the MCP builds with npm run check (tsc + bundle + ESM verify + full test suite; 15 tests incl. golden-vector wire-conformance and the charter gate).
2. Read the guidelines — AGENTS.md is canonical (CLAUDE.md imports it); skills authorship: skills/AGENTS.md; MCP changes: src/mcp/AGENTS.md.
3. Validate before you push:

claude plugin validate .claude-plugin/marketplace.json --strict   # plugin manifest
npm run check                                                     # MCP: build + tests
python3 skills/harvest/evals/scripts/validate_yield.py <yield.md> # yield contract (needs pyyaml)

4. Ship via PR — main is protected. Releases are cut as GitHub Releases (vYYYY.M.D); CI publishes @ema.co/machina via npm OIDC trusted publishing (no tokens).

Layout

CHARTER.md                        # the project charter (mission · goals · bar)
.agents/plugin/marketplace.json   # canonical manifest — marketplace "ema" → plugin "emachina"
.claude-plugin/ · .cursor-plugin/ # symlinked views of the manifest — never edit directly
.agents/skills → ../skills        # agent-agnostic hub (.claude/skills routes through it)
skills/<name>/                    # canonical skill packages (each may carry its CHARTER.md)
src/mcp/                          # the machina MCP server (+ src/mcp/atom/ — the owned kernel)
schemas/atom.schema.json          # the atom contract, as a served artifact
docs/charters/                    # entity-less standing charters (e.g. curator-memorialization)
.harvest/                         # harvest runtime yields (gitignored; AGENTS.md tracked)
.knowledge/                       # local atom corpus — the staging tier (gitignored)