conduit
A local-first, provider-aware coding agent for the terminal.
Conduit began as a Go port of Claude Code. It still keeps that compatibility where it matters: Claude OAuth, Anthropic wire format, the familiar tool surface, project memory, sessions, plugins, and MCP. It has also grown into its own agent runtime: Conduit-owned config, multi-account switching, role-based model assignment, OpenAI-compatible providers such as Gemini, and local/MCP-backed models can all live in one TUI.
The goal is simple: keep the Claude Code ergonomics, make the runtime fast and hackable, and let you choose where each kind of work runs.
<img src="https://api.star-history.com/svg?repos=icehunter/conduit&type=Date" width="600" />
</picture>
Why conduit?
| Claude Code | conduit |
|---|
| Cold start | ~800ms | <30ms |
| Idle memory | ~200MB | <80MB |
| Binary size | ~206MB | ~30MB |
| Dependencies | Node.js + Bun runtime | None (static binary) |
| Multi-account | Restart required | /account panel, live switch |
| Providers | Claude-focused | Claude subscription, Anthropic API, OpenAI-compatible, local MCP |
| Role routing | Single active model | Default, main, background, planning, implement |
| Token savings | External rtk subprocess | In-process RTK filter |
Features
Core agent
- Full Claude Code parity — see
PARITY.md for the current audited feature matrix
- Provider-aware routing — assign models per role from
/models
- OpenAI-compatible providers — use API-key providers such as Gemini through a local credential alias
- Local model support — route turns to MCP-backed local providers, including model servers on another machine
- Streaming SSE — real-time token-by-token output with cost tracking
- Auto-compact — context window managed automatically; manual
/compact
- Thinking mode — extended reasoning via
/effort low|normal|high|max
- Fast mode —
/fast toggles Haiku for faster/cheaper turns
- Coordinator mode — Claude as orchestrator managing parallel sub-agents
- Council mode — multi-model debate across N configured providers; parallel critique rounds with synthesis, convergence detection, member roles, and peer voting
- Exponential backoff — automatic retry on 429 with jitter
Tools (30+ built-in)
| Tool | What it does |
|---|
BashTool | Run shell commands with RTK token compression |
FileReadTool | Read files with line-range support |
FileWriteTool | Write files with diff preview |
FileEditTool | Surgical edits by exact string replacement |
GrepTool | Regex search across files |
GlobTool | File discovery by pattern |
AgentTool | Spawn sub-agents for parallel work |
WebFetchTool | Fetch URLs with content extraction |
WebSearchTool | Web search via Anthropic's native search API |
NotebookEditTool | Edit Jupyter notebooks cell by cell |
REPLTool | Subprocess code execution for JavaScript, Python, and Bash snippets |
TodoWriteTool | Structured task tracking |
TaskCreate/Get/List/Update/Stop | Sub-agent task lifecycle |
EnterPlanMode / ExitPlanMode | Propose-only mode (no file edits) |
EnterWorktree / ExitWorktree | Git worktree switching |
AskUserQuestion | Pause for user input mid-turn |
ConfigTool | Read/write conduit settings |
MCPTool | Call any MCP server tool |
ListMcpResources / ReadMcpResource | MCP resource access |
SkillTool | Execute installed skills |
SleepTool | Pause between actions |
LSPTool | Language server queries: hover, definition, references, diagnostics |
ToolSearchTool | Search available tools by name or description |
SyntheticOutputTool | Inject synthetic tool results (coordinator signalling) |
MCP (Model Context Protocol)
- stdio, HTTP, SSE transports — connect any MCP server
- Conduit-owned MCP config —
~/.conduit/mcp.json overlays imported Claude MCP settings
- Plugin system — install from marketplace (
claude-plugins-official), enable/disable plugins, and manage plugin MCP servers
- Resource listing and reading —
ListMcpResources, ReadMcpResource
- MCP panel —
/mcp opens full browser with per-server tool lists and reconnect
- Plugin panel —
/plugin browses marketplace, manages installed plugins
Memory & context
- CLAUDE.md loading — walks from cwd to root, injects as system blocks
@include directives — compose CLAUDE.md files across directories- Auto-memory — summarises conversations to
MEMORY.md; injected on next session
- Memory scanning —
/memory scan finds stale or contradicted facts
- Relevant memory search — surfaces related memories by keyword on session start
TUI
- Bubble Tea v2 — native Shift+Enter via Kitty keyboard protocol
- Full GFM markdown — tables, task lists, strikethrough, blockquotes, code blocks with syntax
- Light and dark themes —
/theme picker with 6 built-in palettes (dark, light, daltonized variants, ANSI variants), custom theme support
- Output styles —
/output-style for different response modes (default, Explanatory, Learning; plugins can add more)
- Custom keybindings —
~/.conduit/keybindings.json remaps actions, with Claude keybindings imported as fallback
- Status bar — provider, model, mode, context %, cost, and rate-limit indicator
- Coordinator footer — live sub-agent progress during multi-agent runs
Accounts and providers
- Platform keychain storage — macOS Keychain, Linux libsecret, Windows Credential Manager via
go-keyring
/account panel — add, switch, remove, and delete accounts without restarting- Conduit-owned account registry — account metadata lives in
~/.conduit/conduit.json; credentials stay in the keychain
- Provider registry — Claude subscription, Anthropic API, OpenAI-compatible, and MCP providers can be assigned to roles
- Provider context windows — custom providers can set
contextWindow; OpenAI-compatible streams request usage metadata when the server supports it
- Role presets — default, main, background, planning, and implement roles make it easy to run expensive work on one model and smaller work somewhere else
- Provider account OAuth plan — see docs/provider-account-oauth-plan.md for the Copilot and ChatGPT/Codex account-provider design
- Provider account wire checks — see PROVIDER_COMPATIBILITY.md and run
make wire for Copilot and ChatGPT/Codex drift guards
In-process RTK (token savings)
Conduit ships with RTK (Rust Token Killer) ported to Go as a zero-overhead in-process filter. Every BashTool result is passed through the filter before being sent to the model.
Typical savings:
git status → 90% reductioncargo test → 85% (test failures only)npm install → 80%eslint → 75%
View your savings: type /rtk gain inside the TUI.
Hooks
- PreToolUse / PostToolUse / SessionStart / Stop — shell, HTTP, prompt, and agent hooks
- Async hooks — PostToolUse hooks run in background without blocking the next turn
- Desktop notifications — notify on turn complete (macOS native + Linux
notify-send)
Session management
- JSONL transcripts — every session saved to
~/.conduit/projects//.jsonl, with existing Claude history imported on resume
/resume — fuzzy-search all past sessions, load with full history/search — search across all session transcripts for a term- Cost persistence — accumulated cost tracked per-session, shown in
/status
- Session recording —
/record captures terminal output in asciicast v2 format
Installation
Homebrew (macOS / Linux)
brew install icehunter/tap/conduit
Go install
go install github.com/icehunter/conduit/cmd/conduit@latest
Pre-built binary
Download from Releases for your platform:
curl -L https://github.com/icehunter/conduit/releases/latest/download/conduit-darwin-arm64 -o conduit
chmod +x conduit
mv conduit /usr/local/bin/
Build from source
git clone https://github.com/icehunter/conduit
cd conduit
make build
# binary at ./conduit
Quick start
# Start an interactive session
conduit
# Sign in with your Claude Max / Teams subscription (inside the TUI)
/login
# Assign providers/models by role
/models
# Configure accounts, providers, plugins, stats, and usage
/config
# One-shot prompt (no TUI)
conduit --print "explain this codebase"
# Continue last session
conduit --continue
# Resume a specific session
conduit --resume
Slash commands
| Command | Description |
|---|
/help | List all commands |
/commands | Open the slash command picker |
/model | Switch model (claude-opus-4-7, claude-sonnet-4-6, …) |
/models | Open the role-aware provider/model picker |
/fast | Toggle fast mode (⚡ badge in status bar) |
/effort | Set thinking budget: low / normal / high / max |
/compact | Summarise and compress conversation history |
/clear | Clear conversation and start fresh |
/plan | Enter plan mode — Claude proposes, doesn't edit |
/council | Run a multi-model debate regardless of current mode |
/council-history | List past debate transcripts for this project |
/permissions | Manage tool permission mode |
/account | Add, switch, or remove accounts |
/login | Sign in to Claude |
/logout | Sign out current account |
/config | Full settings panel (Status · Config · Stats · Usage · Accounts · Providers) |
/status | Current model, mode, session, cost, context usage |
/theme | Switch colour theme |
/output-style | Switch output formatting style |
/mcp | Browse MCP servers and their tools |
/plugin | Browse and manage plugins |
/memory | View, search, and scan MEMORY.md |
/resume | Load a past session |
/search | Search all session transcripts |
/record | Start / stop session recording (asciicast v2) |
/files | Files read or written this session |
/diff | Git diff of files edited this session |
/context | Current context window usage breakdown |
/color | Toggle ANSI colour output on/off |
/rename | Set a title for the current session |
/rewind | Roll back the last n turns (default 1) |
/commit | Create a git commit with a generated message |
/cost | Token and cost breakdown |
/usage | Rate limit quota and burn rate |
/toggle-usage | Toggle Claude plan usage footer and background fetching |
/stats | Per-tool usage counts |
/session | Session ID, path, message count, duration |
/tasks | Active sub-agent tasks |
/agents | Active sub-agents |
/thinkback | Show last thinking blocks |
/copy | Copy last assistant response to clipboard |
/rtk | RTK token-savings commands |
/doctor | Diagnose: auth, MCP connectivity, settings |
/keybindings | Show active keybinding map |
/buddy | Meet your companion |
/coordinator | Toggle coordinator mode |
/hooks | Manage PreToolUse / PostToolUse hooks |
/init | Create a CLAUDE.md for this project |
/review | Compact and summarise what was done |
/exit | Quit |
Configuration
Settings file
~/.conduit/conduit.json — Conduit's primary user-global settings file, managed via /config or edited directly. On first run, Conduit can import compatible Claude settings from ~/.claude/settings.json; project-local settings in .claude/ and .conduit/ still layer on top, with .conduit/settings.local.json winning last.
Key settings:
{
"model": "claude-sonnet-4-6",
"permissions": {
"defaultMode": "default",
"allow": ["Bash(git status:*)"]
},
"theme": "dark",
"outputStyle": "default",
"env": {
"MY_TOKEN": "..."
}
}
CLAUDE.md
Place a CLAUDE.md in any directory — conduit reads from the file's directory up to the repo root and your home directory, injecting each as a system context block. Use @path/to/file inside to include another file inline.
Custom keybindings
~/.conduit/keybindings.json:
{
"compact": ["ctrl+k"],
"command:clear": ["ctrl+l"],
"submit": ["enter"]
}
Hooks
// ~/.conduit/conduit.json
{
"hooks": {
"PreToolUse": [
{
"matcher": "BashTool",
"hooks": [{ "type": "command", "command": "echo pre-bash" }]
}
],
"PostToolUse": [
{
"matcher": "*",
"hooks": [{ "type": "command", "command": "notify.sh", "async": true }]
}
]
}
}
Hook types: command (shell), http (POST JSON to URL), prompt (inject result as context), agent (spawn sub-agent).
MCP servers
// ~/.conduit/mcp.json
{
"mcpServers": {
"context7": {
"command": "npx",
"args": ["-y", "@upstash/context7-mcp"]
},
"playwright": {
"command": "npx",
"args": ["@playwright/mcp@latest"]
}
}
}
Themes
Built-in themes (matching Claude Code's palette set):
| Name | Description |
|---|
dark | Default dark theme |
light | Light theme |
dark-daltonized | Dark with colour-blindness-friendly palette (alias: dark-accessible) |
light-daltonized | Light with colour-blindness-friendly palette (alias: light-accessible) |
dark-ansi | Dark using terminal's 16 ANSI colours |
light-ansi | Light using terminal's 16 ANSI colours |
Switch with /theme, or set "theme": "dark-ansi" in settings.json.
Custom themes can be defined under "themes" in settings.json:
{
"themes": {
"my-theme": {
"primary": "#c0caf5",
"secondary": "#565f89",
"accent": "#7aa2f7",
"background": "#1a1b26"
}
}
}
RTK — in-process token compression
Conduit ships with RTK (Rust Token Killer) — originally a Rust CLI by rtk-ai — ported to Go and fused directly into the process. It intercepts BashTool results before they reach the model and applies command-aware compression rules.
Inside the TUI:
/rtk gain # cumulative savings across all sessions
/rtk discover # scan transcripts for missed compression opportunities
RTK covers 70 command patterns including: git, cargo, npm/yarn/pnpm, pytest, eslint, dotnet, aws CLI, ls, and structured log output.
Multi-agent / coordinator mode
/coordinator
Enables coordinator mode: Claude acts as an orchestrator that spawns sub-agents via AgentTool. Sub-agents run in parallel, each with their own tool context and history. The coordinator footer in the status bar shows live agent progress.
Building from source
Requirements: Go 1.26
make build # builds to ./conduit
make test # go test ./...
make test-race # go test -race ./...
make lint # golangci-lint run
Claude Code compatibility
Conduit maintains wire compatibility with Claude Code: Anthropic OAuth, API headers and billing block, plugin format (commands, skills, hooks, agents), MCP protocol, and the tool surface the model expects. Claude-compatible settings can be imported or layered for projects, while Conduit-owned user state lives under ~/.conduit.
Features intentionally not implemented (Claude-internal or bridge-dependent):
- VS Code / JetBrains bridge — use real Claude Code for IDE integration
- Remote agents / ULTRAPLAN — Anthropic-hosted, bridge-dependent
- Voice recording — requires cgo + platform audio
- Team swarm (SendMessageTool, TeamCreateTool) — Anthropic teammate mailbox
- GrowthBook / KAIROS — Anthropic-internal
Reference map: PARITY.md
License
MIT — see LICENSE.
Conduit is an independent project and is not affiliated with or endorsed by Anthropic. Claude® is a trademark of Anthropic, PBC.