Filepad is a file-first workspace for AI-assisted work. Documents live in a real, editable file tree — not a chat log, not a hidden prompt database. Every change to a document is a new version, and higher-risk changes wait for a human decision before they land.
Workspaces
A workspace is a file tree plus AI settings. Personal workspaces get a single private root folder; shared workspaces get a shared root (visible to every member) and a private root per contributor. There is no fixed folder scaffold — you organize the tree however your work needs.
Shared workspaces also have a candidate root: a staging area where contributors draft edits that a moderator reviews before they land in the shared tree (see Submissions, below).
Artifacts
An artifact is a document living at a location in the file tree — a note, a sheet, a diagram, a PDF, code, or an image. Every artifact has a version chain: each edit creates a new ArtifactVersion rather than overwriting the last one, so you can diff any two versions or revert to an earlier one.
Each version records who made it — a human, FilepadAI, an external agent, or a workflow — so you always know the provenance of a line, not just that it changed.
FilepadAI
FilepadAI is the native in-app assistant. You talk to it in a thread; its turn is composed of blocks — a working trace of the steps it took (search, read, plan, edit) and a final transcript you actually read. When FilepadAI edits an artifact, that edit goes through the same versioning and governance path as any other write.
Ask it to explain a run, summarize a file, or suggest a fix to your instructions — it has access to workspace context and activity history.
Governed writes
Not every write needs a human in the loop, and not every write skips one. Each action FilepadAI or a connected agent can take is classified:
- Immediate — low-risk actions (most reads, some direct writes) apply right away.
- Reviewable — higher-risk actions (editing a document, moving or deleting files) can pause and return
approval_required, asking a human to approve before the change lands. - Always requires approval — destructive actions (batch deletes) always pause for a decision, regardless of who's asking.
This is the real mechanism behind "review before it lands" — it's part of how every write is authorized, not a separate feature you turn on.
Submissions (shared workspaces)
Submissions are the human co-authoring review flow: a contributor drafts into the candidate root, and a moderator with Editor/Admin access approves or rejects it before it merges into the shared tree. This is distinct from the agent governance above — it's for people reviewing people's drafts, scoped to shared workspaces.
Memory
Workspace memory holds durable facts, decisions, and context that shouldn't need restating in every conversation — shared or private, active or retired, each with a confidence level and provenance. FilepadAI and connected agents can read and propose memory (subject to the memory:read/memory:write scopes), and you can browse the memory timeline to see what's been learned and when.
Connecting an external agent
External agents connect to a workspace with a scoped Agent Access key over OAuth. Two things are available, and they serve different purposes:
- Agent Access API (
/agent-api/v1) — a lightweight channel for pushing activity events, raising signals, and reading a mailbox. This is essentially async reporting: an agent tells Filepad what it did. - MCP tools (
/mcp) — a governed, scope-limited set of tools for searching, reading, and editing workspace content: artifacts, files, and memory. Every call is authorized against the connection's granted scopes and audited; higher-risk writes go through the same approval flow described above.
Both are workspace-scoped: a key issued for one workspace cannot touch another, and access is revocable instantly.
See Agent Access API v1 for the REST reference and Agent Access overview for how a connection is set up.