filepadai
sign inget started

Core concepts

The mental model behind Filepad — workspaces, scoping, proposals, and FilepadAI.

Filepad is a file-first workspace for AI-assisted work. Not a generic chat interface, not a hidden prompt database, not a workflow builder. The core idea: keep the files, sources, memory, artifacts, and reviewed automations visible, scoped, and reviewable.

The environment

Every Filepad workspace is a set of 8 standard folders:

FolderPurpose
.filepad/Workspace manifest and managed Filepad metadata
agents/Optional external-agent operating notes and access instructions
skills/Reusable prompt modules stored as Markdown files
memory/Persistent facts and operating context
sources/Grounding material — documents, data, references
uploads/Raw inputs from outside the workspace
artifacts/Outputs from users, FilepadAI, or external agents
automations/Reviewed automation definitions

Humans, FilepadAI, and scoped external agents work from the same visible folder structure. There is no separate hidden "agent context" database you cannot inspect.

Agents folder

The agents/ folder is optional file space for external-agent operating notes. It is not a deployment target and it does not mean agents live inside Filepad.

An operating note can describe what an external agent should do and what access key scope it expects:

## Scope

read:   agents/, skills/, memory/, sources/
write:  artifacts/
propose: memory/

The file is editable and versioned. An external agent sees it only when its Agent Access key allows it to read the workspace environment.

Skills

Skills are reusable prompt modules stored as Markdown files in skills/. An external agent discovers them via GET /agent-api/v1/workspaces/:id/prompts or /mcp/v1/workspaces/:id/prompts, then reads the skill content through the returned contentUrl.

A skill might be:

  • skills/summarize.md — instructions for summarizing a source document
  • skills/extract-kpis.md — a structured KPI extraction prompt
  • skills/format-report.md — formatting rules for a specific output format

Skills are files. You can read them, edit them, and version them.

Memory

memory/ holds durable workspace facts, decisions, and operating context. Unlike a hidden vector store, memory files are visible and versioned.

Proposals

When FilepadAI or an external agent needs to update a file that requires review, it creates a proposal instead of writing directly.

A proposal is a staged diff: what the agent wants to add, what it wants to remove. You review it line-by-line in Filepad and accept or reject. Nothing lands until you approve it.

External Agent Access V1 uses narrow scopes: read-only discovery and search, direct artifact creation under artifacts/, reviewable proposals, and activity events.

FilepadAI

FilepadAI is the native assistant inside Filepad. It uses the app's session auth, workspace context, /runs, runtime tools, approvals, and context preview. It can:

  • Explain what any connected agent can and cannot access
  • Summarize the contents of any file
  • Identify why a run produced unexpected output
  • Suggest improvements to instructions or skills

Risky or durable changes should become reviewable outputs, proposals, or approvals.

Agent Access API

External agents connect via Agent Access API v1 at /agent-api/v1. Each Agent Access key is:

  • Workspace-scoped (a key for workspace A cannot access workspace B)
  • Scope-limited (env:read, artifacts:write, files:propose, memory:read, events.write)
  • Revocable instantly

Files hidden from agents stay visible to humans but are omitted from Agent Access file-tree, search, prompt discovery, MCP resources, file reads, and proposals.

See Agent Access API v1 for the full reference.