Fergana Labs Stash Stash

Stash — Files, Skills, and Memory System

Concept: Your Stash and Skills

Stash is your personal home for agent work. Everything you create is scoped to your account — it is yours alone, with nothing to pick or set up first. Your Stash has three primary surfaces:

• Sessions — agent transcripts uploaded under /api/v1/me/sessions.
• Files — folders, markdown pages, HTML pages, uploads, and tables.
• Skills — modules of agent-usable knowledge: local SKILL.md folders and shareable bundles of sessions and Files.

To give your agents a skill, create a Files folder whose immediate children include a file named SKILL.md. The body of SKILL.md starts with YAML frontmatter:

---
name: dd-respond
description: Draft response packets to investor diligence asks
when_to_use: When an investor sends a DD checklist
version: 2.1
mcp_exposed: true
---

The folder may contain any number of supporting .md files (examples.md, checklist.md, etc.) — they all become part of the skill payload. Stash exposes skills via:

• GET /api/v1/me/skills — list your skills
• GET /api/v1/me/skills/{name} — full skill (SKILL.md + siblings)
• MCP: stash_list_skills, stash_read_skill

This is the same skills convention Claude Code uses, so a skill authored in Stash works directly when dropped into any agent's ~/.claude/skills/ folder.

Overview

Stash is the product surface for you and your agents.

It provides:

• pages organized in nestable folders
• tables (typed columns, rows, CSV import/export, semantic row search)
• session events (with file attachments)
• file uploads (S3-backed; PDF/image text extraction when available)
• Skills for publishing sets of pages, sessions, and files

Design boundary:

• Stash owns persistent state and plugin-based memory access
• external orchestration layers own multi-agent delegation
• Claude-session memory access should go through the Stash plugin, not side-channel polling

Base URL

{{PUBLIC_URL}}

Authentication

All endpoints (except registration and a few public lookups) require an API key:

Authorization: Bearer mc_xxxxxxxxxxxxx

Your account is the scope. GET /api/v1/users/me returns the authenticated user; everything else hangs off the /api/v1/me prefix.

Quick Start

1. Register

curl -X POST {{BASE_URL}}/api/v1/users/register \
  -H "Content-Type: application/json" \
  -d '{"name": "my-agent", "description": "A helpful assistant"}'

Response includes api_key — save it, it's shown only once.

2. Push a Session Event

curl -X POST {{BASE_URL}}/api/v1/me/sessions/events \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"agent_name":"cli","event_type":"note","content":"Hello"}'

3. Create a Page

curl -X POST {{BASE_URL}}/api/v1/me/pages/new \
  -H "Authorization: Bearer $API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"name":"Notes","content":"# Hello"}'

Pass "folder_id": "<uuid>" to drop the page into a specific folder; omit it to create the page at your root.

4. Upload a File

curl -X POST {{BASE_URL}}/api/v1/me/files \
  -H "Authorization: Bearer $API_KEY" \
  -F "file=@./report.pdf"

Response includes the file id, a signed url, and basic metadata. For PDFs with embedded text and text-based documents, extracted text is available at GET /api/v1/me/files/{id}/text once the background extractor has processed the file (typically a few seconds).

Route Surfaces

Everything is scoped to your account under /api/v1/me. Single shared objects also have canonical, scope-free URLs (/api/v1/{pages,files,tables}/{id}) used when linking to a specific object by id.

CRUD verbs are standard: POST to create, GET list/detail, PATCH update, DELETE remove. Semantic search hangs off the page surface (GET /api/v1/me/pages/semantic-search?q=...).

Page Content (content_markdown)

Pages store markdown. The editor implements a deliberately small subset — agents writing content should stick to what's listed below, since anything else silently passes through as plain text on render.

Links

Use ordinary markdown links for everything:

The viewer renders all three with the same style; an ↗ glyph marks off-origin URLs. Internal /p/<uuid> and stash absolute URLs are SPA-routed (same tab, no reload); externals open in a new tab.

There is no [[...]] syntax. Use ordinary markdown links with the page's id URL.

Block elements

Anything starting with #### (H4+), > (blockquote), triple-backtick code fences, or --- (hr) is not parsed — the markdown renders literally. Treat it as unsupported.

Inline

Strikethrough (~~), underline markup, LaTeX, footnotes, and raw HTML tags are not supported.

Authoring from agents

When generating page content programmatically, follow these rules and you'll round-trip cleanly through edit mode:

1. Never emit [[…]] — use the id-URL link form.
2. Never emit relative paths like [text](foo.md) or ![](foo.jpg) — the renderer treats them as dead and the editor strips them if a user saves the page.
3. Don't rely on H4 or deeper headings. Restructure with H3 + bold.
4. Images need an absolute URL (external or /api/v1/me/files/<id>/download).

Session Events

Events are structured append-only records keyed by (agent_name, event_type).

POST /api/v1/me/sessions/events
{
  "agent_name": "cli",
  "event_type": "note",
  "content": "text body",
  "session_id": "optional",
  "tool_name": "optional",
  "metadata": {},
  "attachments": [
    {"file_id": "<uuid>", "name": "report.pdf", "content_type": "application/pdf"}
  ]
}

attachments entries must reference a previously-uploaded file. The CLI wrapper (stash sessions push --attach ./path) uploads and attaches in one step.

Query/search:

• GET /events?agent_name=&event_type=&limit=&after=
• GET /events/search?q=&limit=
• GET /events/{event_id}

Files

• POST /files — multipart upload (field file), 50 MB cap.
• GET /files — list.
• GET /files/{id} — metadata (with signed URL).
• GET /files/{id}/text — extracted text. Response shape: {"text": ..., "status": "pending|processing|done|failed", "error": ...}. Works for PDFs with embedded text and for plain-text / JSON / XML uploads. Extraction runs asynchronously after upload — poll this endpoint until status is done or failed.
• DELETE /files/{id} — best-effort S3 cleanup plus DB row delete.

Rate Limits

• Registration: 5/min
• Login: 10/min
• CLI auth session polling: 60/min

Tips for Agents

• Everything you create is scoped to your account — your API key is the scope.
• For extracted text on an uploaded file, poll GET /files/{id}/text — it returns status alongside the text so you can distinguish "still extracting" (pending/processing) from "done, no text available" (done with text: null).
• Attach files to session events rather than embedding base64 — keeps event payloads small and allows reuse across events.
• When authoring page content that links to another page, use the page id URL form: [text](/p/<uuid>).

Source: Fergana-Labs/stash — distributed by TomeVault.