07of 12 · Features

Pick a model by task, not by logo.

Claude, GPT and Codex, Gemini, Kimi, DeepSeek, Grok, Qwen, GLM, and local models on your own machine. One window, one interface, one set of directives. Choose per session, and switch mid-thread.

One queue. The right engine for each request.

Routed by task

overwatch · queue · 4 requests

01Build the retry path on checkoutcoding

02What do competitors charge? Cite sourcesresearch

03Fix the typo on the pricing pagesmall edit

04Review the diff before it shipscode review

/OVERWATCH routes each request to the model that fits it

Coding

Claude · Opus

Deep reasoning for the build that touches forty files.

→ 01 · retry path

Research

Perplexity · Sonar

Live web answers with citations, straight into the session.

→ 02 · competitor pricing

Small edits

Gemini · Flash

Fast and cheap for the one-line fix that needs no ceremony.

→ 03 · pricing typo

Code review

Codex · GPT-5.5

A second set of eyes from a different engine before it ships.

→ 04 · review the diff

Same window, same keys. The task picks the model — you never re-explain the project.

Every engine Overwatch speaks to

Claude

GPT · Codex

Gemini

Perplexity

Kimi

Meta · Muse

DeepSeek

Grok · xAI

Qwen

GLM

Ollama

LM Studio

Frontier, open-weight, and local — one integration, your key for each.

What people actually say

“Significantly dumber. It ignored its own plan, messed up the code, and started to lie about the changes it made.”

The "AI shrinkflation" backlash. Anthropic's own postmortem confirmed two bugs degraded models Aug 5 to Sep 5 2025. A six-week track of identical prompts logged correctness falling 4.2 to 3.1 and hallucinations rising 8% to 22%.
View the source ›

Demand rank #2 Demand score 93/100 Overwatch mitigates it Full research ledger ›

What Overwatch does about it

When one model has a bad week, being locked to it is the whole problem. Overwatch lets you move the same session onto a different engine and keep going. The per-session intensity dial also pins high effort where a silent "medium" default quietly took it away.

What that gets you

07 / Multi-model

Every major model, one window

One integration, 75+ providers. Frontier models, open-weight models, and local models through Ollama or LM Studio.

Per session, not per app

Reasoning model on the hard thread, a fast cheap one on the scratch thread. Both open at the same time.

Your key, your cost

Bring your own account. The key stays on your device and you control what gets spent.

How it works

In the app

Switch mid-thread without losing the thread

Not happy with the answer? Change the engine and ask again in the same session, with the same context. No copy-paste into another app, no re-explaining the project.

Same transcript across engines
Same directives, same folder, same files
Compare in a split pane, side by side

switch mid-thread

YouThat answer drifted. Try it on a different engine.

GeminiSame session, same files, same directives. Here is the version that keeps the schema untouched.

OpusGeminiCodex

No copy-paste into another app. No re-explaining the project.

Directives route work to the right engine

Type //perplexity for live research, //image to generate, //gemini for a second opinion. The session stays where it is and the right model does the piece it is best at.

//perplexity, //image, //gemini and more
Set a default model per folder
Intensity dial: how hard it should think

directives route the work

//perplexityLive web research, cited

//imageGenerate an image in this session

//geminiSecond opinion on the last answer

//localRun it on the model on this machine

The session stays where it is. The right model does the piece it is best at.