Flow ships a command line tool, an MCP server and a skill file. Your agent gets taught how to use it, and you decide what it's allowed to do.
$ nyxe session.search --query "migration" --limit 5 { "ok": true, "value": { "matches": 11 } } $ nyxe insight.generate --kind summary --id 4f2a { "ok": true, "value": { "kind": "summary" } } $ nyxe insight.emit --id 4f2a --to reminder ⚠ requires grant: emitExternal exit 4 · consent denied
Nyxe installs a nyxe command onto your PATH on launch, and it describes its own commands, so an agent can list them itself instead of hunting for a doc.
Every session, meeting and note is one command away. Search across them, generate a summary, or list what you shipped. Reading is free and needs nothing but the binary.
nyxe schema emits the whole command tree, so the guide an agent reads is generated from the tool rather than written by hand.
Every result is a JSON envelope, so an agent can act on the answer instead of scraping a pretty table.
No API keys and no server to babysit. It reads Flow's own local database directly, so Flow doesn't even need to be open.
Flow registers its MCP server with the agent apps it finds and drops a skill file beside each. A connected agent gets a read-only subset: it can search your sessions and generate summaries on-device, and change nothing until you grant it.
Registering isn't enough on its own. Flow also drops a skill file next to each agent, so it knows what the tool is for and where the line is before it touches anything.
The guide says the same thing every time. Reading is free, acting needs consent, and a denied grant is surfaced rather than retried. An agent that follows it stays inside the lines by default.
# nyxe-flow — how to use it Reading is free. Query sessions and notes with nyxe session.search and session.get. Summaries run on-device. Sending and deleting need a grant, and return exit 4 without one. # Don't retry a denied grant. Surface it.
A connected agent can read your sessions and generate summaries on-device. Sending and deleting are blocked until you grant them, and every request is logged, allowed or not.
A grant is narrow and named. Let an agent draft a reply on-device without letting it send anything, and take the grant back the moment you're done.
What was asked for and whether it was allowed, in one log. Anything blocked shows up plainly, and never a word of what was said.
$ nyxe insight.emit --id 4f2a --to reminder ⚠ requires grant: emitExternal exit 4 · consent denied $ NYXE_GRANTS=emitExternal \ nyxe insight.emit --id 4f2a --to reminder { "ok": true, "value": { "emitted": "reminder" } }
The CLI, MCP server and skill file ship with Flow.
macOS · included with every paid Nyxe plan