<p align="right"><a href="README.zh-CN.md">简体中文</a></p>
<p align="center">
<a href="https://hiroute.ai/">
<img src="apps/website/public/brand/github-social-preview.png" alt="HiRoute — intelligent routing for long-horizon agents" width="100%">
</a>
</p>
<p align="center">
<strong>Make Codex and Claude Code subscriptions go further.</strong><br>
A local model router that gives routine agent work to economical models and keeps your strongest model for the hard parts.
</p>
<p align="center">
<a href="https://hiroute.ai/">Website</a> ·
<a href="https://hiroute.ai/download/">Download</a> ·
<a href="https://hiroute.ai/en/docs/">Documentation</a> ·
<a href="https://hiroute.ai/en/docs/decision-extensions/">Decision models</a> ·
<a href="CONTRIBUTING.md">Contributing</a>
</p>
<p align="center">
<a href="https://github.com/higress-group/HiRoute/releases/latest"><img alt="Latest release" src="https://img.shields.io/github/v/release/higress-group/HiRoute?display_name=tag&sort=semver&style=flat-square"></a>
<a href="LICENSE"><img alt="Apache-2.0 license" src="https://img.shields.io/badge/license-Apache--2.0-5A6FE8?style=flat-square"></a>
<a href="https://hiroute.ai/download/"><img alt="macOS and Linux" src="https://img.shields.io/badge/platform-macOS%20%7C%20Linux-171A21?style=flat-square"></a>
</p>
<p align="center">
<a href="https://hiroute.ai/en/#tour">
<img src="apps/website/public/media/hiroute-promo-en-readme.jpg" alt="Watch the 1-minute HiRoute product tour" width="720">
</a>
</p>
Keep using Codex, Claude Code, Qoder, Pi, OpenCode, Hermes or DeepSeek Harness. With a difficulty-routing plan, HiRoute
judges verified native user input: routine work such as reading code or gathering facts can go to an
economical model, while design decisions and difficult fixes stay with your subscription or
primary model. Subscriptions, API keys and local models are managed in one place, and each
session shows which model actually ran, its usage and its cost. Independent tasks can also be
delegated to other agents.
Tool calls within a turn usually keep that turn's decision, so the growing context prefix stays
cacheable when continuity and the captured decision remain valid. HiRoute decides again when a
usable new task input arrives, and may keep the same model; cache reuse depends on the provider.
HiRoute runs locally as a Desktop application or a headless CLI and daemon.
Classification, history capture and cross-turn evaluation remain available when a native client
cannot prove input origin. Such input carries explicit uncertainty; missing captured execution
stays partial. Codex compressed sessions can recover on new questions after a Gateway restart
without creating a new conversation. Claude, Hermes and Qoder retain smart routing while task
association and scoring may be inaccurate. Qoder's shared main/child identity is an accepted
limitation. Known auxiliary summaries stay outside main history. Successful output alone does
not establish routing or scoring accuracy. See the [routing and recovery limits](apps/website/content/guides/model-routing.en.md).
## News
- **Latest stable release** — macOS Desktop and Linux headless packages for x86_64 and ARM64. [Release notes](https://github.com/higress-group/HiRoute/releases/latest) · [Download](https://hiroute.ai/en/download/)
- **2026-10-14** — [Switching Models Loses Cache, Not the Savings](news/2026-10-14-routing-cache-cost.en.md). [Website article](https://hiroute.ai/en/news/routing-cache-cost/)
- **2026-10-04** — [GPT-6 Astra × Qwen-3.8 Flash: same quality at 90% lower cost](news/2026-10-04-astra-qwen.en.md). [Website article](https://hiroute.ai/en/news/astra-qwen-smart-routing/) · [Experiments](experiments/README.md)
- [HiRoute Is Open Source: Help Your Agents Do More and Your Subscriptions Go Further](news/hiroute-open-source.en.md). [Website article](https://hiroute.ai/en/news/hiroute-open-source/)
- [All news](news/README.md)
## Why HiRoute
| | |
| --- | --- |
| **Route long tasks by stage** | Use the model that fits the work ahead instead of committing an entire task to one model. |
| **Protect quality while saving** | Use observed competence to block risky downgrades, while task complexity creates opportunities to use economical models. |
| **Keep handoffs bounded** | Enforce capability requirements, ordered candidates, and explicit exhaustion instead of silently changing behavior. |
| **Learn from execution** | Retain the task category, actual model group and candidate, tool outcomes, accepted output, and optional stage competence for later decisions. |
HiRoute currently provides four routing modes: fixed model, difficulty routing, custom branches, and free first.
Each mode uses the plan's ordered candidates for bounded failover. It also provides agent work plans
and Worker delegation; local session, usage, cost, and competence views; and a custom extension API
for connecting your own decision service.
## Get started
### macOS Desktop
Download the current DMG from [hiroute.ai](https://hiroute.ai/download/), drag HiRoute into
Applications, and follow the [self-signed package instructions](docs/macos-installation.md) if
macOS asks for manual approval. HiRoute supports macOS 15 or later on Apple silicon
(`arm64`) and Intel (`x86_64`).
### Linux headless
Install the current-user CLI and daemon on Linux `x86_64` or ARM64 (`aarch64`) without `sudo`:
```sh
curl -fsSL https://hiroute.ai/install.sh | sh
hiroute service start --output json
hiroute system status --output json
```
The installer verifies the published package manifest and checksums and does not start a
service automatically. See the [Linux installation guide](https://hiroute.ai/en/docs/install-linux/)
and [CLI reference](docs/standalone-cli.md) for service lifecycle and machine-readable commands.
### Configure your first route
1. Connect a supported subscription, registered API, or compatible custom API.
2. For model-based judgment, open **Models → Decision models** and connect Bailian,
OpenRouter Jev, TypeSafe, or a compatible endpoint. Built-in connections need no self-hosted
Jev service.
3. Create and publish a routing plan, using difficulty routing or custom task branches when needed.
4. Connect an Agent client, or expose a Worker plan for delegated tasks.
5. Inspect sessions and model performance to understand the task category, actual model group,
and execution result.
Continue with [model routing](https://hiroute.ai/en/docs/model-routing/),
[task routing](https://hiroute.ai/en/docs/task-routing/), or the
[headless CLI guide](https://hiroute.ai/en/docs/cli/).
## See routing performance
<p align="center">
<img src="decision-extensions/assets/quality-native-en.png" alt="HiRoute desktop session: stage competence and model routing" width="100%">
</p>
Review model performance within a task, alongside execution records and user feedback, to inform
the next model choice and task delegation. Each score stays linked to its execution stage.
## How routing works
Built-in decision models judge the task category when needed and return simple/complex
probabilities for routes with two model groups. They can also assess the previous execution stage.
HiRoute supplies the published conditions and scoring criteria, applies your thresholds,
and chooses the model group and ordered candidates.
Configure these connections in **Models → Decision models**; you do not need to deploy an extension.
**Difficulty routing** has one task scope with regular and primary model groups. Regular is selected
when this turn's simple-task probability meets your threshold and no applicable, complete
assessment falls below the competence floor. Otherwise HiRoute selects primary. Missing or
partial scores stay unrated. An old low score does not lock future turns to primary, and a new
turn does not automatically return to regular: HiRoute judges the current work again.
**Custom branches** separate task categories such as drafting and review. HiRoute first uses
the category choice, then the simple/complex degree within that category to choose its regular
or optional primary group. A category with only a regular group needs no degree judgment.
An assessment belongs to the stage that actually ran; a low drafting score does not upgrade
the review category. Only a complete, applicable assessment for the same category, published
revision, and frozen criteria can affect this turn's group choice.
Tool continuations keep the decision when HiRoute can recognize the same turn and reuse its
frozen decision. Candidate failures use bounded failover: regular candidates may relay to the
same category's primary group, while a primary selection stays within primary. Exhaustion fails
explicitly. Failover does not create a competence score or change the task category.
The principle is:
> **Competence blocks risky cost-cutting; complexity creates opportunities to save.**
Difficulty routing also supports heuristic rules. A custom extension is optional: use it when you
want to own the model integration, judgment, scoring, and context trimming. The official
[Jev extension](decision-extensions/extensions/jev-decider/README.md) is one implementation
of that boundary. Start with [decision models and the routing mechanism](decision-extensions/README.md),
the [custom extension API guide](decision-extensions/api/README.md), or the canonical
[OpenAPI document](decision-extensions/api/decision.openapi.json).
## For technical contributors
HiRoute is a Rust workspace with a Tauri Desktop client and an Astro website. The Rust toolchain
is pinned in `rust-toolchain.toml`; Desktop uses Node.js 24 and the website uses Node.js 22.
```sh
git clone https://github.com/higress-group/HiRoute.git
cd HiRoute
cargo fmt --check
cargo test --locked --workspace --exclude hiroute-desktop --all-features
```
For frontend work, run the checks in the affected application:
```sh
cd apps/desktop # or apps/website
npm ci --ignore-scripts
npm test
npm run build
```
See [CONTRIBUTING.md](CONTRIBUTING.md) for native dependencies, Clippy, validation scope, and
language conventions. Native Desktop behavior and real-provider integration require their
respective environments; a source build alone is not product acceptance.
### Repository map
The [code map](docs/code-map/README.md) connects user capabilities to architecture,
production owners and representative tests.
| Path | Responsibility |
| --- | --- |
| `apps/desktop` | Desktop UI and Tauri host |
| `crates/cli`, `crates/daemon` | Headless client and local role-all service |
| `crates/application`, `crates/client-core` | Application workflows and shared client behavior |
| `crates/gateway`, `crates/gateway-core` | Agent-facing protocols, routing, execution, and fallback |
| `crates/observation`, `crates/local-storage` | Local execution evidence, queries, and persistence |
| `decision-extensions` | Decision model guides, routing mechanism, extension API, screenshots, and optional Jev extension |
| `contracts`, `assets` | Current runtime contracts, model metadata, and product resources |
| `apps/website` | Bilingual `hiroute.ai` site, downloads, and release publication |
| `e2e`, `tools`, `scripts` | Product scenarios, validation tools, packaging, and automation |
## Project status
The latest stable release provides macOS Desktop and Linux headless packages for both
x86_64 and ARM64. Download the latest version and read the release notes on the
[website](https://hiroute.ai/en/download/).
HiRoute is licensed under the [Apache License 2.0](LICENSE). Use
[GitHub Issues](https://github.com/higress-group/HiRoute/issues) for bugs and feature requests.
Report security issues and crash reports through the process in [SECURITY.md](SECURITY.md).
Links: [LINUX DO](https://linux.do/) — a Chinese-language tech community.