FarkiBot Reference

Technical reference for FarkiBot, Farkitect’s AI assistant. For guidance, see How to set up FarkiBot and How to choose an LLM.

Opening FarkiBot

Action	How
Open the panel	View > FarkiBot, or Ctrl+Shift+K
Open settings	The gear icon in the FarkiBot toolbar

FarkiBot is a dockable panel, tabbed by default alongside Properties, Context, Notes, and Log. It can be dragged, split, and (on desktop) popped out like any other panel.

Tasks

Task	Writes to project	Minimum role	Output
Ask	No	Viewer	A streamed answer in the panel
Report	No	Viewer	A Markdown document, saved via Save .md
Generate M2	On import only	Contributor	A staged metamodel package
Generate M1	On import only	Contributor	A staged model package
Modify M1	On replace only	Contributor	A diff, then a replaced package

Tasks a role can’t run are shown disabled, labelled (needs contributor access), rather than hidden.

Your LLM choice is remembered per task, so Ask and the generation tasks can each keep their own.

Report kinds

Report	Purpose
Model Overview	Narrative document for stakeholders who need to understand the estate without opening the model
Gap & Risk Assessment	Prioritised findings, each with evidencing model content, impact, and a recommendation
Reference Documentation	Systematic walkthrough of every package in scope
Executive Summary	One page, business language, conclusion first

A brief is optional for all four. Saved files are named from the document’s own title.

Providers

Provider	Reached	Notes
OpenRouter	Directly from the browser	Recommended. One key, 300+ models, free tier. Reports per-run cost.
Anthropic	Directly from the browser	Under Advanced. Stronger privacy posture — no gateway in the path. No cost reporting.
Google Gemini	Directly from the browser	Under Advanced. No cost reporting.
Custom endpoint	Directly from the browser	Any OpenAI-compatible base URL — Ollama, LM Studio, an internal gateway. No web research.

OpenAI has no direct option — its API does not permit browser connections. Add an OpenAI key inside OpenRouter’s BYOK settings instead.

Hosts the app is permitted to contact

Enforced by Content-Security-Policy on both the web and desktop apps:

openrouter.ai
api.anthropic.com
generativelanguage.googleapis.com
localhost:*          (custom/local endpoints)

…plus Farkitect’s own Convex backend. Any other destination is refused by the browser.

Attachments

Task	Accepted
Ask, Report	`.md`, `.txt`
Generate M2, Generate M1, Modify M1	`.md`, `.txt`, `.pdf`, `.docx`, `.farki`

PDFs are sent to the model natively — no text extraction step
.docx is read from the file directly, mapping heading styles to Markdown headings and list paragraphs to bullets
.doc (the pre-2007 format) is not supported — save as .docx or PDF

Key storage

Setting	Behaviour
Session only (default)	In memory; cleared when the tab closes
Remember on this browser	AES-256 ciphertext in browser storage, wrapped by a non-extractable browser-held key

Keys are masked on entry, display only their last 4 characters after saving, and have no reveal affordance. Settings are per browser, per device, and are never synced to Farkitect’s servers.

Limits

Limit	Value
Repair attempts per generation	3
Staged packages retained	10 per project (oldest evicted)
Staged package persistence	Survives page reload
Run lifetime	The browser tab — a reload or close ends a run
Free-model requests (OpenRouter)	50/day, or 1,000/day once you’ve purchased $10+ of credits
Output token limit	Model-dependent; raised automatically when a generation runs out of room

Cost reporting

Provider	Token counts	Cost
OpenRouter	Yes	Yes
Anthropic	Yes	No — shows “cost not reported”
Google Gemini	Yes	No — shows “cost not reported”
Custom endpoint	Varies	Usually not

A researched answer costs roughly 3–4× a plain one, because search results are added to the prompt.

Web research

Property	Value
Default	Off, chosen per run
Where the search runs	On the provider’s servers, inside the completion call
Supported on	OpenRouter (all models), Anthropic and Google (current models)
Not supported on	Custom / local endpoints
Citation	Every web claim carries an inline link
Marking	Answers that used research carry a Web badge

Farkitect’s servers are not in the path, and no new network destination is contacted — the provider hosts already on the allow-list perform the search.

Provenance

Every generated package carries a provenance note recording:

The model that produced it
The provider
A prompt hash
The date

Runs are also recorded in the project’s governance metadata, including whether web research was used.

Modify M1 — the replace sequence

Step	Stops the replace if
1. Freshness check	The target model changed since the rewrite was generated
2. Confirmation dialog	You decline
3. Snapshot save	The snapshot can’t be confirmed written — you’re offered retry, replace-without-snapshot, or abandon
4. Replace	—

Snapshot filename format: <Model Name> (before FarkiBot modify YYYY-MM-DD HHMM).farki

Snapshot confirmation by browser: Chrome and Edge report whether the file was written. Firefox and Safari do not — a started download is all the app can observe.

M2 metamodels are excluded from this task. See How FarkiBot Works.

Error messages

Message	Cause and fix
”Your API key was rejected”	Wrong or revoked key. Re-paste it in FarkiBot Settings and use Test connection.
”…out of credits”	Top up at your provider (OpenRouter: Credits page).
“…hidden reasoning” errors	A reasoning model burned its output budget thinking. FarkiBot already tried raising the limit and disabling reasoning, so this means those recoveries were exhausted. Pick a different model.
”Generation failed validation”	The model couldn’t produce a valid model file after 3 attempts. Retry with a starred model; the raw output is downloadable for salvage.
”Generation interrupted”	Runs live in the browser tab — a reload or tab close ends them. Try again resubmits the same brief.
”…cut off at the output token limit”	FarkiBot raises the limit automatically up to the model’s maximum, then asks for a smaller complete file. If it persists, narrow the brief or use a model with a larger output cap.
”…remaining credit can’t cover the full output budget”	Your key’s spend cap can’t pre-authorise the full output budget. Raise the cap or top up.
”Rate-limited”	Free models allow 50 requests/day (1,000 once you’ve bought $10+ of credits). Wait, or use a paid model.
Empty model list on a custom endpoint	The gateway has no `/models` route. Type the model id directly.
”The model returned this package unchanged”	A Modify M1 run produced no changes. Try a more specific instruction.