Say hello, or type a message. Hold the mic to dictate. Type
/ for commands, @path to attach a file.Available rules
Create new rule
Describe the directive in plain language. The currently-selected
LLM (Vantage or AskSage, per Settings) drafts the rule's
markdown (YAML frontmatter + body) and validates the
frontmatter. On a validation miss, the drafter retries once
with the errors fed back. You then review the draft and click
Create.
Review the draft. Edit either field if needed, then click
Create rule. Frontmatter is re-validated on save.
Available skills
Create new skill
Describe what the skill should do. The currently-selected LLM
(Vantage or AskSage, per Settings) drafts SKILL.md and the Python
module, then runs the same three-stage validator (frontmatter
parse → py_compile → load probe) the manual path uses. On a
validation miss, the drafter retries once with the errors fed
back. You then review the draft and click Create.
Review the draft. Edit either field if needed, then click
Create. The Create button re-runs validation — anything
unsafe still blocks the write.
Import skill
Upload a .zip containing SKILL.md and one or more .py files. Other
file types are rejected. Path-traversal entries and symlinks are
blocked. Total uncompressed size must be under 5 MB.
Workspace
Your 50 GB cloud workspace. The agent saves here by default;
drag files anywhere on this panel to upload. To touch files on
your actual machine instead, ask the agent — local-device
operations route through your helper exe.
Loading…
/
Upload
Or drag & drop files anywhere on this panel. Uploads land in
the folder shown above.
Scheduled tasks
Scheduled tasks fire a synthetic user prompt through the orchestrator —
every loaded skill and tool is available. Schedules only run while
Bastard AI is running; if you close the app, the next trigger fires
whenever you reopen it (within 5 minutes of the missed time).
Loading…
Pick a schedule on the left to view its details and run history,
or click + New schedule to create one.
Settings
LLM provider
Pick the inference enclave for both the chat LLM and the knowledge-graph extraction LLM. Vantage / Envision use an OpenAI-compatible Foundry proxy; AskSage uses api.genai.army.mil; Local talks to any OpenAI-compatible server you run yourself (llama.cpp, vLLM, Ollama). Switching is hot — no restart, no re-sign-in. The bearer JWT / API key expires periodically — paste a fresh value here when the provider returns 401.
Pick from the curated list. Click "List models" to verify which are actually available to your AskSage account.
Must end in
/v1 for OpenAI-compatible servers. Default points at the homelab llama-server on cluster-bebo.Only needed if you've put an authenticating reverse proxy in front of your local server. Stored encrypted alongside the Vantage/AskSage keys.
llama-server returns whatever GGUF is loaded regardless of this value, but vLLM / Ollama use it to pick between served models.
Microsoft 365 (Graph)
Sign-in always requests the full read+write scope set, so flipping the write toggle below is instant — no re-sign-in needed. The runtime guard remains the source of truth for whether mutations are permitted.
Your default browser should have opened. If it didn't, copy the URL below and the code into a browser on this machine, then approve the sign-in.
Diagnostics
Debug shows event flow; verbose adds full payload dumps to the per-subsystem log files. Triggers off bypasses the regex pre-execution layer.
File Size Limits
Caps on what the agent can read from disk, what you can upload via
the chat, and what the rag image tools accept. Changes apply
immediately — no restart. Out-of-range values are clamped to the
per-knob safe range. Click "Reset to factory" to revert all seven
to their shipped defaults.
Allowed Drives & Paths
Paths the agent's file_system tools may read and write under. Anything
outside this list is refused. Changes apply immediately — no restart.
Paths that don't exist on disk are dropped silently and listed in the
save status. The base directory must be inside one of the allowed paths.
Comms Streamlining
Settings for the comms-streamlining skill. The OneDrive root is where
tracker syncs (action items, projects, info logs) are mirrored.
Default "/" is your actual OneDrive root.
Chain patterns (planner seeds)
Markdown source of the workflow patterns the planner uses to match
user requests to multi-agent chains. Lives at
chain_patterns.md
in the project root — edit here or directly on disk. Each pattern is a
## Pattern Name block with **ID:** `pattern_id`,
a description, **Keywords:**, and a yaml code block of
steps. Save validates parses; if zero patterns parse, the save is
refused so you don't accidentally wipe every chain pattern.
After save the planner DB is reseeded immediately — no restart.
Helper devices
Each device that runs your helper exe authenticates via a
long-lived bearer token bound to your account. Generate a
one-time pair code here, paste it into the helper's first-run
prompt, and the helper exchanges it for a token it stores
DPAPI-sealed on the device. Revoke a device any time — its
token stops validating immediately on the next connect.
Loading…