Skip to content

ArcavenAE/flyloft

Flyloft

A curated, graph-aware retrieval substrate for human-AI teams. Rig the materials. Fly them in on cue. Strike what's stale.

Flyloft is a private-deployable reference system designed for small teams of humans and AI agents working together on sustained, evolving problems — security operations, platform engineering, research programs, anywhere a shared working memory needs to grow, be groomed, and be trusted.

It is a RAG substrate with a reranker, a graph overlay, and a curation surface — but more importantly, it is a role: a home for the emerging researcher-librarian, the flyperson, who tends the collection that the rest of the troupe draws from.

Where it sits

Flyloft is part of the Arcaven theatrical stack. Each component is independently useful; the stack composes them.

Project Role
forestage The agent console — persona-themed CLI wrapping Claude Code (rename to apron planned)
marvel Agent control plane — Kubernetes-like orchestration (teams, roles, sessions, shifts)
tmux-cmc tmux control mode client library — the wiring beneath forestage
switchboard Remote tmux relay — serial console for distributed agents
director Inter-agent communication (planned) — supervisor/crew coordination protocol
curtain Minimal agent sandbox — kernel-enforced containment
spectacle Standards-based spec templates + Claude Code commands (ISO 29148, IEEE 1016, 42010, arc42, MADR)
sideshow Content pack manager — distributes rules, skills, themes across repos and AI CLI tools
Flyloft Retrieval substrate — the casebook
kos Knowledge graph and epistemological process — the promptbook
beads Task tracker — dolt-backed issues across the fleet

Flyloft is the raw, discoverable corpus. kos is the distilled, authoritative spec layer. Material is rigged into Flyloft, flown to agents on demand, and — once proven and stable — promoted into kos.

Flyloft also federates. Alongside the local stacks (material the team has rigged), Flyloft queries catalogs — registered external sources exposed through adapters (GitHub, Confluence, peer Flylofts, research databases). Held and cataloged battens are curated with the same verbs. See docs/CATALOGS.md.

Deployment modes

Flyloft is designed to run two ways:

  1. Managed by marvel — teams running marvel-managed agent workloads get a provisioned Flyloft as part of their workspace, reachable over MCP by the agent sessions marvel schedules.
  2. Standalone (BYOA) — Bring Your Own Agents: run Flyloft as a private service against any agent runtime (Claude Code, forestage, or a custom stack). Expose via MCP and a REST fallback.

Same binary, same data model, different front doors.

Quickstart (placeholder)

# Initialize a new flyloft at the current directory
flyloft init

# Rig some source material
flyloft rig ./docs/runbooks/
flyloft rig https://example.com/article

# Start the MCP server so agents can fly material in
flyloft serve --mcp

# Open the flyperson's grooming surface
flyloft groom

The Flyperson

A Flyloft without a flyperson is a pile. The flyperson is the team member — sometimes a dedicated role, more often a rotating responsibility — who:

  • reviews what agents are asking for and what they're finding
  • rigs new sources
  • spikes battens that prove authoritative
  • strikes what's stale
  • promotes mature material into KOS
  • watches the cue sheet for retrieval gaps

This role is the dramaturg of the human-AI team. Flyloft is the workshop they work in.

Status

Design-stage. Schema, architecture, and interface specs are in docs/. A minimal Rust workspace skeleton lives in crates/. First implementation milestone is in docs/ROADMAP.md.

Further reading

About

Curated, graph-aware retrieval substrate for human-AI teams — RAG + reranker + curation surface. The casebook to kos's promptbook.

Topics

Resources

License

Code of conduct

Contributing

Security policy

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors