Visible, never silent
Every pick shows a 13X RECOMMENDS or DEFAULT badge, three editable dropdowns, an always-visible price-vs-quality line, a confidence label, and an honest "why." Nothing is auto-applied — you click APPLY ALL or just send.
Stop paying opus for typo fixes
Config Advisor recommends an agent kind, model, and effort for each task you dispatch — visible, justified, and overridable in a click. It never auto-applies anything silently.
overview
Config Advisor picks an agent configuration — agent kind (claude/codex), model, and effort — for each task you send, so you stop overpaying on trivial work and under-powering hard refactors. Every recommendation is shown with a badge, three editable dropdowns, a price-vs-quality line, a confidence label, and an expandable "why." It draws on a cheap local complexity classification of the task text plus your own past choices for that exact task, then applies by typing /model and /effort into the target claude terminal before the context packet.
- Recommendations are visible and overridable — never auto-applied
- No API keys stored — reuses your own authenticated claude CLI
- No 13x telemetry — every signal is local
- History lives in <root>/.13x/decisions.jsonl, append-only
- Hard 600ms dispatch budget with an honest static fallback
- Config commands inject into claude slots only; codex gets text only
why it matters
Local-first, no API keys, no telemetry
Classification reuses your own installed claude CLI (which brings its own auth), and past-choice history lives in a local .13x/decisions.jsonl. 13x stores no keys and sends nothing to a 13x server.
Deterministic, explainable combiner
No black box: a fixed priority order — agreed history, then blast radius, then text, then static default — produces the config, and the reasons list never claims a source that did not fire.
Learns from your corrections
Each send logs recommended vs. chosen and whether you overrode it. After 3+ runs of the exact same task with a 2/3 majority, the Advisor reuses your proven config at high confidence — plain frequency counting, no trained model.
Hard 600ms dispatch budget
At dispatch the recommendation only does cache and file lookups under a 600ms race. If anything is slow it falls back to the honest static default — nothing keeps running in the background.
Never mutates a foreign agent
Config commands are injected only into claude slots. Codex and other CLIs receive the task text only, and the row says so honestly — a config can never convert a running agent to another CLI.
how it works
01
Type your task
In the AUTO CONTEXT task box or the Broadcast input, type what you want done. Broadcast only shows the advisor for task-like text (no leading slash or bang, at least 16 chars); short snippets stay silent.
02
Get a justified pick
After a short pause the config row fills in. While you type, 13x classifies complexity off-thread via your claude CLI; at recommend time it races your local history and the task-text signal under a 600ms timer and shows the deciding source.
03
Read the why and tradeoff
The badge, confidence, and an always-visible price-vs-quality sentence sit under the row. Expand "why" for the honest reasons — e.g. "prompt classified as complex" or "chosen in 4 of 5 past runs of this task."
04
Tweak or apply
Override the agent-kind, model, or effort dropdowns, or click APPLY ALL to snap rows back to the recommendation. Editing any dropdown marks the row overridden.
05
Send and log
On send, the chosen claude config is typed as /model and /effort before the context packet, the packet goes to your selected slots, and one decision is logged to .13x/decisions.jsonl per task.
faq
Does Config Advisor apply anything automatically?
No. Every recommendation is shown with a badge, editable dropdowns, a tradeoff line, and a "why." The config is only applied when you click APPLY ALL or send — it is never auto-applied silently.
Does it send my task text or code anywhere?
Not to 13x. Complexity classification runs through your own installed claude CLI via stdin (capped at 4000 chars). All other signals are local — a classify cache in memory and decisions.jsonl on disk. 13x stores no API keys and sends nothing to a 13x backend.
How does it learn from me?
Each send logs the recommended and chosen config and whether you overrode it. After 3 or more runs of the exact same task with a 2/3 majority, it reuses your proven config at high confidence. It is plain frequency counting over an exact text fingerprint — no trained model, no cross-task inference.
What's still a stub?
Blast-radius (Schicht B) is deliberately stubbed and always returns null until the Map rework ships, so today's picks rest on task text plus history only. Codex injection is unverified (codex rows recommend but inject nothing), and there is no in-app UI for the classifier model — it's hand-edited in .13x/settings.json.
get started
one project. many shells. zero noise.
13x is local-first and free of telemetry — your code, commands and shells never leave the machine. Join the waitlist for the next build.