hi

Zero‑intrusion multi‑backend proxy for Claude Code. No Claude Code modifications, no forks, no plugins. hi sits between Claude Code and the network — intercepting only API calls — and restores everything on exit. Your Claude Code stays vanilla.

Connect to any Anthropic‑compatible API. Officially supports Claude, DeepSeek, MiniMax, GLM, Kimi, OpenRouter, internal gateways, and custom endpoints. Hot‑switch between all of them without restarting your session.

The name comes from the first word you type into any Claude Code or AI agent session — hi. The agent replies "Hello, how can I help you?" and the work begins. Simple, memorable, and fitting for the AI era. May evolve if a better name comes along.

Why hi?

hi currently prioritizes Claude Code — arguably the best harness agent tool available today. However, CC is also the most expensive. For everyday users who rely on harness agents heavily but don't have a bottomless budget, the cost can be prohibitive. DeepSeek, on the other hand, offers some of the most affordable large models in the world. Pairing Claude Code with DeepSeek makes for a fantastic combination: top-tier agent capabilities at a fraction of the price.

Looking ahead, hi plans to support more agent tools and the OpenAI API protocol, making it a universal multi-backend proxy.

Quick start

Prerequisites

Claude Code installed and logged in
hi binary — download from Releases (recommended) or build from source
API key — hi works with any Anthropic-compatible endpoint. You need an API key from one of:
- Anthropic Console — Claude API key
- DeepSeek Platform — DeepSeek API key
- Your organization's internal API gateway

Feature	Status
Claude Code version	Tested up to 2.1.168 (latest)
Hot-switch backends	✅ Claude, DeepSeek, and all custom backends
Web Search	✅ Works on all supported models

Install (Linux / macOS)

curl -fsSL https://raw.githubusercontent.com/mars-base/hi/main/install.sh | sh

This downloads the latest release binary and installs it to /usr/local/bin/hi (or ~/.local/bin/hi if /usr/local/bin requires sudo).

Install (Windows)

PowerShell:

irm https://raw.githubusercontent.com/mars-base/hi/main/install-windows.ps1 | iex

CMD:

curl -fsSL https://raw.githubusercontent.com/mars-base/hi/main/install-windows.cmd -o install.cmd && install.cmd && del install.cmd

Installs to %USERPROFILE%\.local\bin\hi.exe and adds the directory to your user PATH. Restart your terminal after installation.

Run

# 1. Initialize configuration (auto-detects settings.json)
hi init-config

# 2. Edit config (Linux/macOS: ~/.hi/config.yaml, Windows: %USERPROFILE%/.hi/config.yaml)
#    Skip if the auto-detected defaults are fine.

# 3. Launch proxy + Claude Code
hi

# Or with inline API keys (no config file needed)
ANTHROPIC_API_KEY=sk-xxx DEEPSEEK_API_KEY=sk-xxx hi

# Windows PowerShell
$env:ANTHROPIC_API_KEY="sk-xxx"; $env:DEEPSEEK_API_KEY="sk-xxx"; hi

# Windows CMD
set ANTHROPIC_API_KEY=sk-xxx && set DEEPSEEK_API_KEY=sk-xxx && hi

# Attach additional agents to an already-running proxy
hi cc

# Same as above, explicit
hi launch --backend deepseek

# Proxy only (no Claude Code, no settings.json patch)
hi proxy --log-file /tmp/hi.log

# Show config and status
hi status

Command reference

Command	Proxy	Claude Code	settings.json patch	Slash commands	Use case
`hi`	✅	✅	✅	✅	Full workflow (recommended)
`hi launch`	✅	✅	✅	✅	Same as bare command
`hi proxy`	✅	❌	❌	✅	Proxy only
`hi agent` / `hi cc`	❌	✅	✅	✅	Attach to existing proxy
`hi status`	❌	❌	❌	❌	Show config
`hi statusline`	❌	❌	❌	❌	Claude Code status bar

CLI options

Option	Default	Description
`-b, --backend <name>`	`claude`	Backend: `claude` / `deepseek`
`--log-file <path>`	`~/.hi/logs/hi.log`	Log file path (auto-rotated by date)
`--log-level <level>`	`info`	Log level: `debug` / `info` / `warn` / `error`
`--preserve-statusline`	—	Keep existing statusLine command (don't replace with hi)

Log files are automatically rotated by date. The filename is decorated with today's date before the extension:

~/.hi/logs/hi.log  →  ~/.hi/logs/hi-2026-06-05.log
/tmp/hi.log      →  /tmp/hi-2026-06-05.log

At midnight the logger switches to a new file automatically — no signal or restart needed. Old files are never deleted; manage retention manually or with a cron job.

Watch logs

tail -f /tmp/hi-$(date +%F).log | grep -E "#[0-9]|Control:|backend |env:"

All paths shown as ~/.hi/ resolve to your OS home directory automatically:

Platform	Path
Linux	`/home/user/.hi/`
macOS	`/Users/user/.hi/`
Windows	`C:\Users\user\.hi\`

Supported Models

hi works with any Anthropic‑compatible API. Here are the officially supported providers and their endpoints:

Provider	Type	Base URL	Official Guide
Anthropic (Claude)	`anthropic`	`https://api.anthropic.com`	—
DeepSeek	`deepseek`	`https://api.deepseek.com/anthropic`	—
MiniMax	`anthropic`	`https://api.minimax.io/anthropic`	Docs
GLM (z.ai)	`anthropic`	`https://api.z.ai/api/anthropic`	Docs
Kimi	`anthropic`	`https://api.kimi.com/coding/`	Docs

Any gateway or proxy that speaks the Anthropic API protocol (e.g. Leihuo, OpenRouter, OneAPI, internal gateways) also works — just set type: anthropic and the appropriate base_url.

Configuration

First run of hi status auto-generates ~/.hi/config.yaml:

active_backend: deepseek
proxy_port: 18799

backends:
  claude:
    type: anthropic
    base_url: https://api.anthropic.com
    api_key: "${ANTHROPIC_API_KEY}"
    strip_thinking: true
    pricing:
      input: 3.00
      output: 15.00
    models:
      opus: claude-opus-4-8
      sonnet: claude-sonnet-4-6
      haiku: claude-haiku-4-5-20251001

  deepseek:
    type: deepseek
    base_url: https://api.deepseek.com/anthropic
    api_key: "${DEEPSEEK_API_KEY}"
    strip_thinking: true     # remove top-level thinking config
    reasoning_effort: max    # optional: output_config.effort (deepseek only)
    pricing:
      input: 0.42
      output: 0.83
    models:
      opus: deepseek-v4-pro[1m]
      sonnet: deepseek-v4-pro[1m]
      haiku: deepseek-v4-flash[1m]

The type: anthropic backend works with any Anthropic‑compatible API — OpenRouter, OneAPI, internal gateways, Azure, etc. Just update base_url and api_key. See Adding custom backends for details.

API key resolution priority

When hi expands ${VAR} references in ~/.hi/config.yaml, it checks:

OS environment variable (os.Getenv)
~/.claude/settings.json env block (fallback)
Non-${...} format — passed through as-is

This is hi's own resolution order. Claude Code's startup priority is the reverse — settings.json env > OS process env — which is why hi must patch settings.json to redirect API traffic through the proxy.

Startup logs show the resolution source:

[hi] DEBUG env: ANTHROPIC_API_KEY=...KoKG (from OS env)
[hi] DEBUG env: ANTHROPIC_API_KEY=...KoKG (from ~/.claude/settings.json)
[hi] DEBUG env: DEEPSEEK_API_KEY=<not set>

Model name mapping

Claude Code uses Anthropic native model names. hi auto-remaps them:

Claude model	deepseek backend	claude backend
`claude-opus-4-8`	`deepseek-v4-pro`	`claude-opus-4-8`
`claude-sonnet-4-6`	`deepseek-v4-pro`	`claude-sonnet-4-6`
`claude-haiku-4-5-20251001`	`deepseek-v4-flash`	`claude-haiku-4-5-20251001`

Adding custom backends

You can add any Anthropic‑compatible API endpoint as a backend. Define it in ~/.hi/config.yaml under backends, then restart hi:

backends:
  claude:
    type: anthropic
    base_url: https://api.anthropic.com
    api_key: "${ANTHROPIC_API_KEY}"
    pricing: { input: 3.00, output: 15.00 }
    models:
      opus: claude-opus-4-8
      sonnet: claude-sonnet-4-6
      haiku: claude-haiku-4-5-20251001

  deepseek:
    type: deepseek
    base_url: https://api.deepseek.com/anthropic
    api_key: "${DEEPSEEK_API_KEY}"
    pricing: { input: 0.42, output: 0.83 }
    models:
      opus: deepseek-v4-pro[1m]
      sonnet: deepseek-v4-pro[1m]
      haiku: deepseek-v4-flash[1m]

  # Example: custom backend via an internal gateway
  internal:
    type: anthropic
    base_url: https://llm.internal.example.com
    api_key: "${INTERNAL_API_KEY}"
    pricing: { input: 0.50, output: 1.00 }
    models:
      opus: claude-opus-4-8
      sonnet: claude-sonnet-4-6
      haiku: claude-haiku-4-5-20251001

Key points:

type: anthropic — for Anthropic‑compatible API endpoints
type: deepseek — strips thinking blocks before forwarding
api_key — supports ${ENV_VAR} expansion or literal keys
models.opus/sonnet/haiku — maps Claude model names to backend‑specific IDs
pricing — USD per 1M tokens, used by cost tracking
context_window — max context window size in tokens for statusline display. Default: 1000000 for deepseek, 200000 for other types
reasoning_effort — sets output_config.effort for deepseek backends: max / high. Empty to disable. Default: empty

DeepSeek 1M context window: for the official DeepSeek API, append [1m] to the model name to unlock the 1M‑token context window. Write deepseek-v4-pro[1m] and deepseek-v4-flash[1m] in the models block. If your API gateway rejects the [1m] suffix, just use the plain names (deepseek-v4-pro, deepseek-v4-flash) instead.

Hot-switching backends

Switch backends mid-session without restarting Claude Code. The switch only affects the next API call.

From within Claude Code, use the auto-generated slash commands:

/deepseek   →  Switch to DeepSeek
/claude     →  Switch to Claude

Or manually via the control endpoint:

curl -sX POST http://127.0.0.1:18799/_proxy/mode -d 'backend=deepseek'

Or create a slash command ~/.claude/commands/deepseek.md:

Switch the proxy to DeepSeek. Run silently and report the result:
curl -sX POST http://127.0.0.1:18799/_proxy/mode -d 'backend=deepseek'
If HTTP 200, say "Switched to DeepSeek." Otherwise report the error.

Then type /deepseek in any Claude Code session to switch instantly.

Switch log output

[hi] INFO  Control: backend switched deepseek → claude
[hi] INFO  New backend env:
[hi] INFO    ANTHROPIC_MODEL                = claude-sonnet-4-6
[hi] INFO    ANTHROPIC_DEFAULT_OPUS_MODEL   = claude-opus-4-8
[hi] INFO    ANTHROPIC_DEFAULT_SONNET_MODEL = claude-sonnet-4-6
[hi] INFO    ANTHROPIC_DEFAULT_HAIKU_MODEL  = claude-haiku-4-5-20251001

Multi‑agent workflows

A single hi proxy instance can serve multiple Claude Code agents simultaneously. Each agent's API calls go through the same proxy, sharing cost tracking and backend switching.

There are two ways to run multiple agents against a single proxy:

Method 1: `hi proxy` + `hi cc` (recommended)

Start the proxy once, then attach agents with zero manual config:

# Terminal 1: start proxy
hi proxy

# Terminals 2+: attach agents
hi cc
hi cc --backend claude
hi cc

Method 2: `hi launch` + bare `claude`

hi launch patches settings.json on startup. Once patched, any bare claude invocation picks up the proxy address automatically:

# Terminal 1: full launch (proxy + agent + settings patch)
hi launch --backend deepseek

# Terminals 2+: just claude — settings.json already points to proxy
claude
claude

Caveat: hi launch shuts down the proxy when its Claude Code exits, killing the shared proxy for everyone. Use Method 1 when you need the proxy to stay alive independently, or when launching agents in arbitrary order.

All agents share the same hot-switch endpoint — switching the backend via /_proxy/mode affects every connected agent immediately. Cost tracking aggregates across all sessions into a single ~/.hi/.cost.json.

Status line live model update

hi auto-discovers your existing statusLine configuration in ~/.claude/settings.json and replaces it with the built-in hi statusline command. After a backend switch, the model name in the status bar updates automatically (within 120s). No manual config needed.

Pass --preserve-statusline to skip this override and keep your existing statusLine command untouched:

hi launch --preserve-statusline
hi cc --preserve-statusline

How it works

hi launch
  ├── Discover settings.json statusLine.command
  ├── Save original to ~/.hi/.statusline-command
  └── Replace with hi statusline

Claude Code status bar refresh (every 120s)
  ├── Run hi statusline
  ├── Query proxy → get active backend model
  ├── Patch model field in stdin JSON
  └── Delegate to original script for cost, context, cache stats

Effect

Backend	Status bar shows
deepseek	`🤖 deepseek-v4-pro`
claude	`🤖 claude-sonnet-4-6`

Test manually

echo '{"model":{},"workspace":{},"context_window":{},"cost":{}}' | hi statusline
# 📁 tmp | 🤖 deepseek-v4-pro | 🧠 ctx:-- | 💰 $0.010

Control endpoints

# Status
curl -s http://127.0.0.1:18799/_proxy/status | python3 -m json.tool
# {
#   "active_backend": "deepseek",
#   "backends": ["claude", "deepseek"],
#   "requests": 47,
#   "uptime_seconds": 3600
# }

# Cost tracking
curl -s http://127.0.0.1:18799/_proxy/cost | python3 -m json.tool
# {
#   "backends": {
#     "deepseek": {
#       "input_tokens": 23223,
#       "output_tokens": 166,
#       "cost": 0.0104,
#       "anthropic_equivalent": 0.0722
#     }
#   },
#   "total_cost": 0.0104,
#   "anthropic_equivalent": 0.0722,
#   "savings": 0.0618
# }

# Switch backend
curl -sX POST http://127.0.0.1:18799/_proxy/mode -d 'backend=deepseek'
# {"mode":"deepseek","previous":"claude"}

Cost persistence

Cost data persists across proxy restarts. Token usage is accumulated in memory and flushed to disk on a regular cadence:

Trigger	Behavior
`Record()` per request	Memory-only update (no disk write)
Background goroutine	Flush to `~/.hi/.cost.json` every 30s
Proxy shutdown	Final flush via `Close()`
Next startup	Load cumulative data from file

This avoids excessive disk I/O during heavy sub‑agent workloads while ensuring data survives process restarts and system reboots.

# Check cumulative cost across all sessions
cat ~/.hi/.cost.json | python3 -m json.tool

Debugging

Log levels

Level	Shows	Use case
`debug`	Everything (env resolution, model remap, upstream URL)	Troubleshooting
`info`	Request trace, backend switch, cost stats (default)	Daily use
`warn`	Warnings + errors	Watch for issues
`error`	Errors only	Minimal noise

Request log examples (debug level)

[hi] INFO  15:07:03 #2 POST deepseek /v1/messages 200 634ms
[hi] INFO  15:07:03 #2 tokens: in=114 out=120
[hi] DEBUG 15:07:03   -> upstream: POST https://api.anthropic.com/v1/messages
[hi] DEBUG 15:07:03   <- status=200 content-type=application/json
[hi] DEBUG 15:07:03 Model remap: claude-sonnet-4-6 → deepseek-v4-pro

hi status output

Config:  /home/fish/.hi.yaml
Claude:  detected at /home/fish/.claude/settings.json
Proxy:   http://127.0.0.1:18799
Active:  deepseek
Log:     /tmp/hi.log (level=info)
Backends:
    claude
  * deepseek

Building from source

Requires Go 1.21+.

git clone https://github.com/mars-base/hi.git
cd hi
make build
make install

Cost comparison

Backend	Input/M	Output/M
Claude Opus	$3.00	$15.00
DeepSeek V4	$0.42	$0.83

Heavy usage (25 days/month): $30-80 vs $200 — 60-85% savings.

The proxy tracks and reports savings in real time:

total_cost: 0.0104              ← actual spend
anthropic_equivalent: 0.0722    ← what Anthropic would cost
savings: 0.0618                 ← 85.6% saved

License

CC BY-NC 4.0 — free for personal and non-commercial use. Commercial use requires explicit permission.

Name		Name	Last commit message	Last commit date
Latest commit History 57 Commits
.github/workflows		.github/workflows
cmd/hi		cmd/hi
pkg		pkg
.gitignore		.gitignore
.goreleaser.yaml		.goreleaser.yaml
LICENSE		LICENSE
Makefile		Makefile
README.md		README.md
README_CN.md		README_CN.md
go.mod		go.mod
go.sum		go.sum
index.html		index.html
install-windows.cmd		install-windows.cmd
install-windows.ps1		install-windows.ps1
install.sh		install.sh

Folders and files

Latest commit

History

Repository files navigation

hi

Why hi?

Quick start

Prerequisites

Install (Linux / macOS)

Install (Windows)

Run

Command reference

CLI options

Watch logs

Supported Models

Configuration

API key resolution priority

Model name mapping

Adding custom backends

Hot-switching backends

Switch log output

Multi‑agent workflows

Method 1: hi proxy + hi cc (recommended)

Method 2: hi launch + bare claude

Status line live model update

How it works

Effect

Test manually

Control endpoints

Cost persistence

Debugging

Log levels

Request log examples (debug level)

hi status output

Building from source

Cost comparison

License

About

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases 15

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Method 1: `hi proxy` + `hi cc` (recommended)

Method 2: `hi launch` + bare `claude`

Packages