filepadai
sign inget started
features

Built for operators,
not prompt engineers.

Every feature in Filepad exists to make agent work observable, auditable, and safe to hand off.

01/The workspace

Real files. Standard structure. Always readable.

Every Filepad workspace is a structured folder system, not a hidden prompt database or chat log. Every important file is something you can open, edit, version, and review.

.filepad/
Workspace manifest and managed Filepad metadata.
agents/
Optional external-agent operating notes and access instructions.
skills/
Reusable procedures exposed as prompts.
memory/
Durable workspace facts and decisions.
sources/
Grounding material — documents, data, references.
uploads/
Raw inputs: PDFs, spreadsheets, exports.
artifacts/
Outputs from users, FilepadAI, or external agents.
automations/
Reviewed automation definitions.
agents / research-agent.md
1# Research Agent Notes
3name: external-agent
4model: gpt-4o
6## Scope
7read: agents/, skills/, memory/, sources/
8write: artifacts/
9propose: memory/
Live file treeSyntax-highlighted editorInline editingRevision history
02/Visible context

The context stays visible to you.

No hidden prompts. No black-box context. Skills, memory, sources, and operating notes are files you can open, read, and edit directly.

03/Scoped agents

Permissions per folder. Not per account.

Each Agent Access key is granted explicit scopes — read, write, propose, or locked — per environment folder. A key that can write to artifacts/ cannot touch .filepad/. Revoke any key instantly. Every scope change is logged.

external-agent● connected
read
agents/
memory/
skills/
sources/
write
artifacts/
propose
memory/
locked
.filepad/
automations/
Per-folder scopesHMAC-SHA256 authRevoke instantlyFull audit log
memory/facts.mdpending
···## Operational Facts
−primary_language: Python
+primary_language: TypeScript
···module_pattern: canonical
+updated: 2025-04-28
✓ Accept✕ Rejectexternal-agent · 2h ago
Line-by-line diffAudit trailConfigurable per folderStatus bar indicator
04/Review & approval

No write lands without your review.

Agents that need to update memory/ or write artifacts/ submit proposals. You see a line-by-line diff: what they want to add, what they want to remove. Accept it, reject it, or send it back. Every decision is timestamped and attributed.

05/FilepadAI

Ask questions and produce reviewable work.

FilepadAI is the native in-app assistant. Ask it to explain access, summarize files, debug unexpected behavior, or propose better organization with the context of your actual file tree.

FilepadAI — workspace context
YOU
What does external-agent have access to?
FILEPADAI
external-agent can: • read → agents/, skills/, memory/, sources/ • write → artifacts/ • propose → memory/ 1 pending review in memory/facts.md.
Workspace contextExplains permissionsDebug agent behaviorReviewable outputs
Agent Access packages
MCPnpx -y @filepad/mcp-server@latesttools/list
SDK@filepad/agent-access-sdktyped client
GET/agent-api/v1/workspaces/:id/file-treeenv:read
POST/agent-api/v1/workspaces/:id/artifactsartifacts:write
POST/agent-api/v1/workspaces/:id/files/:id/proposalsfiles:propose
POST/agent-api/v1/workspaces/:id/eventsevents.write
06/Agent Access

Point any MCP-capable agent at any workspace.

Use the published MCP server for OpenClaw, Claude Desktop, Cursor, Windsurf, and custom agents. The server turns scoped Agent Access keys into safe tools for reading context, searching files, creating artifacts, proposing edits, and reporting activity.

Ready to operate agents properly?

Start free. No credit card. One workspace, one agent, full visibility from day one.