Only changes how times are shown to you. Job schedules and the server clock stay as they are.
Falcon Orchestrator
Pick a model and start the conversation.
AIStrategist
Project Falcon
Pick a model and start the conversation.
This clears the conversation history on the server (a backup is kept if supported).
The model will read the conversation and write a summary. The full history is backed up first, then replaced with the summary — this frees up context. This can take a minute or two.
Only changes how times are shown to you. Job schedules and the server clock stay as they are.
Large windows need a lot of memory (KV cache).
A catalog is a set of model patterns (qwen/*, ext:anthropic, ext:*). Attach catalogs to workgroups with a daily token cap each; users get the union of their workgroups' catalogs. Until the first catalog exists everyone sees every model. Administrators are exempt.
For users who belong to no catalog-bearing workgroup.
A team is a tool-capable orchestrator (a local model, an external agent whose provider supports tools, or auto: the best orchestrator by assessment) plus named roles, each pinned to a model (local or an external agent) or given a profile (orchestrator, researcher, writer) and left on auto, so the best eligible model by assessment answers at call time. Users send a message to it with @team:<name>: the orchestrator plans, delegates briefs to the roles within the budgets and writes the final answer. A user sees a team when their catalogs entitle them to its orchestrator model; each role call is charged like any other use of its model.
Every model Falcon knows, with the scope scores routing uses (🧪 measured by the assessment battery, 📖 curated by hand in AvailableModels.md), its hosts and your tags. Tags are searchable here and usable as catalog patterns (tag:fast). Each assessed model also gets a suggested role (signal-engine, fast-worker, reviewer, composer, vision; helper, the platform how-to role, is shown but never counted as the best) with a 0–100 fit; ≈ marks an indicative fit. Assessments run only on loaded models; administrators only.
Non-secret tunables of this server. Precedence: environment variable > saved value > built-in default. Values set by the environment are read-only here; changes marked restart needed apply after a server restart, every other change applies at once.
Keys live in the encrypted settings store; create them on the Bybit sub-account tagged bot with trade+read only, withdrawals off.
The S3-compatible bucket where you keep model artifacts for the worker Macs (Wasabi, AWS S3, Cloudflare R2, Backblaze B2, MinIO). Keys live in the encrypted settings store; workers only ever receive time-limited download links, never the keys.
Files dropped into the chat are encrypted with the master key and belong to the session or workgroup they were uploaded in; deleting that session or workgroup deletes them.
How long each log is kept and how it rotates. Housekeeping runs every hour; Purge now applies a retention immediately, Rotate now starts a fresh file. The audit log rotates by UTC day on its own (its hash chain must stay intact).
Shown once. The install command carries a one-time enrollment token: valid 15 minutes, usable on one machine only. No key is minted until that machine redeems it.
Enrollment tokenInstalls into ~/falcon-worker (worker script + LaunchAgent, no venv), detects or installs LM Studio headless, enrols this machine and keeps the key in the macOS Keychain. The worker keeps itself up to date.
When some members already have the model loaded, only those are considered: loading a large model just to balance work is never worth it.
Approvers can approve actions proposed for this workgroup. Only current members can be approvers.
Project Falcon