Security

The boundary stays
where you draw it.

Local-first files, outbound-only connections, and a human gate on anything that writes.

Where your data lives

Your .md filesYour disk — plain files, Git if you wantOnly when you connect them
Provider API keysOS keychain / encrypted storeNever written into files
Agent & chat stateLocal — works offlineOnly for docs you share
Bridge trafficOutbound socket to the relayNothing listens inbound

AI-specific leak paths

What could leak — and what stands in the way

AI agents read real data and can take real actions. These are the honest leak paths, and the control on each one.

Prompt contextAnything an agent reads is sent to the model you picked — cloud providers see what the agent sees.Run a local model (Ollama / LM Studio) and nothing leaves the machine, or scope each agent to only the tools it needs.
Secrets in contextA key pasted into a file could ride along inside a prompt.Tokens resolve from env or the OS keychain at run time — referenced, never written into .workflow.md.
Over-broad tool callsAn agent with write or network tools could act beyond what you intended.Per-agent tool scoping, process-isolated MCP servers, and approval prompts before writes or commands.
Prompt injectionHostile text inside a file or fetched page can try to steer the agent into leaking or acting.Tool output is untrusted input — outbound and write actions stall at the human gate until you approve.
Provider retentionCloud inference is covered by each provider's data terms.Bring your own API keys to keep prompts under your agreement — or use managed AI for convenience, per agent.

Process-isolated tools

Every MCP server runs in its own process over stdio — no shared memory, with timeouts and OS-level limits.

Human approval gates

checkpoint nodes pause workflows for a person, and agent permission prompts gate every write action.

Least-privilege sharing

Owner, editor, or viewer per doc — read-only links and time-limited access when you share.

Your models, your call

Point agents at Ollama or LM Studio and nothing leaves the machine — or bring your own API keys.

No secrets in files

Workflow auth pulls tokens from env or the OS keychain — .workflow.md stays safe to commit.

Inspect everything

Every run, step, and tool call is logged locally — and every save is a version you can restore.

Found something? Report it via support.