AgenC Core

An open source AI coding agent for your terminal. It works with any model provider.

<p align="center"> <img src="docs/media/agenc-demo.gif" width="880" alt="AgenC Core in a terminal with DeepSeek V4.1 Flash. It finds two regressions in a checkout service, fixes them, and runs the tests until all four pass, for less than a cent."> </p>

curl -fsSL https://get.agenc.ag/install.sh | sh

Then run agenc onboard. Windows and npm installs are on get.agenc.ag, and you can verify the download.

• With DeepSeek V4.1 Flash it scored 76.2% on Terminal-Bench 2.1 for $34. Every entry on the Harbor Hub leaderboard spent at least $134. Full results.
• Bring any model: OpenAI, Anthropic, Grok, Gemini, DeepSeek, OpenRouter and more, or local models through Ollama and LM Studio.
• Work in the terminal UI, run it headless with agenc -p in scripts and CI, or start background agents and follow them from your phone.

<p align="center"> <a href="https://get.agenc.ag/terminal-bench/"><img src="docs/media/terminal-bench-2-1-cost.png" width="720" alt="Chart of score against cost on Terminal-Bench 2.1. AgenC Core with DeepSeek V4.1 Flash sits at the top, 76.2% for $34. The leaderboard entries all cost $134 or more."></a> </p>

AgenC Core is the implementation repository for the agenc CLI. A local daemon owns agent/session lifecycle, permissions, provider calls, and command execution. The interactive TUI, headless --print / --no-tui CLI, background agents, channel gateway, and remote phone bridge are all clients of that daemon.

| Package | Path | Role | | --- | --- | --- | | @tetsuo-ai/agenc 0.19.8 | packages/agenc/ | Public launcher binary | | @tetsuo-ai/runtime 0.19.8 | runtime/ | Daemon, TUI, tools, providers, tests | | @tetsuo-ai/agenc-sdk 0.3.0 | packages/agenc-sdk/ | Typed embedding SDK (daemon protocol) |

Documentation map: docs/INDEX.md. Architecture: docs/ARCHITECTURE.md. CLI reference: docs/reference/cli.md.

Contents

• Features
• Project status
• Requirements
• Quick start
• Usage
• Configuration
• Architecture
• Development
• Security
• Contributing
• License

Features

• Daemon-backed process model — launcher → daemon (app-server) → TUI / print / agents / gateway clients. One control plane per AGENC_HOME.
• Two front-ends — interactive React TUI and headless one-shot mode (agenc -p|--print, agenc --no-tui).
• Background agents — fire-and-forget sessions (agenc agent start|list| attach|stop|logs), attachable and resumable, independent of the foreground UI.
• Channel gateway — Telegram, Discord, Slack, WebChat, and stdio channels; pairing, bindings, in-channel approvals. See docs/gateway.md.
• Budget-bounded execution — daemon-owned reservations, hierarchical caps, concurrency, cancellation, and evidence across interactive/background work, including heartbeat / cron / hooks. Design: docs/design/execution-admission-kernel.md.
• Durable scheduled prompts — require Linux directory descriptors and a local OS-account home for protected locks. Stop older scheduler/gateway processes before upgrading; see storage and platform requirements.
• Guided onboarding — agenc onboard includes AgenC account sign-in, X / xAI sign-in for Grok, provider API keys, and local providers without requiring slash commands; follow-on acts: identity, channel, autonomy, recap (personas, channels, budget/heartbeat/webhooks).
• Remote control — pair iOS or Android with agenc remote on|off|status; co-drive chats, receive background completion/attention events, and settle permissions from the phone. Android can route an explicit, physically approved @ledger SOL transfer to a Ledger Flex. See docs/remote-control.md and the Ledger security contract.
• Ledger Wallet CLI — Ledger hardware-wallet requests route through a bundled skill and a read-only CLI status tool. If the official wallet-cli is missing, AgenC asks before downloading anything; an approved /ledger install (or agent install action) resolves @ledgerhq/wallet-cli@latest, verifies the platform package's SHA-512 integrity, and installs it under AGENC_HOME.
• Built-in tools — Bash, file read/write/edit, transactional apply_patch, web fetch/search, LSP, MCP, sub-agents; read-before-write and atomic-patch safety.
• Browser automation — the Browser tool drives an isolated Chromium over a CDP pipe with stable accessibility refs (navigate, snapshot, click, type, screenshot, tabs). All egress routes through an SSRF proxy that blocks private/loopback/metadata addresses by default; dedicated profile, never your real one.
• MCP — outbound MCP client and MCP server (agenc mcp serve|…), including enterprise XAA (SEP-990).
• Layered safety — permission modes, opt-in OS sandbox (bubblewrap/Landlock on Linux, Seatbelt on macOS), agenc security audit [--fix].
• 17 built-in providers — default provider grok; fresh-config session model grok-4.6 (fresh config and direct-provider map). Grok 4.6 has a 500k context window, low/medium/high/xhigh reasoning (catalog defaultReasoningLevel is high; defaultConfig().reasoning_effort is medium and bootstrap copies that onto the session), vision, tools, and structured output; Grok 4.5 and Grok 4.3 remain selectable in the catalog; also openai, anthropic, ollama, lmstudio, openai-compatible, openrouter, groq, venice, deepseek, meta, gemini, mistral, nvidia-nim, minimax, github, amazon-bedrock, agenc. See docs/reference/providers.md.
• Grok OAuth: sign in with X during onboarding, via /grok-login, or agenc grok-login for subscription Grok access without an API key (docs/grok-oauth.md).
• Embedding SDK — @tetsuo-ai/agenc-sdk for socket / subprocess embedding. See docs/sdk.md.
• Durable sessions — append-only rollout logs + SQLite state; --continue / --resume. Optional local trajectory export for training data (docs/trajectory-training-data.md).

Project status

Current version in tree: 0.19.8. Runtime and launcher are versioned 0.19.8; the embedding SDK package is 0.3.0. The public launcher is @tetsuo-ai/agenc. The root workspace is non-publishable ("private": true); the GitHub source repository is public so npm can issue verifiable provenance. Type-clean: 0 @ts-nocheck. MIT licensed (LICENSE).

Shipped in this line: durable recovery and transactional compaction, scalable workflow and CSV scheduling, bounded project and memory indexes, provider-aware token accounting, multi-channel gateway, Browser tool, budget-bounded autonomy, Ledger verification, and the independently versioned SDK.

Requirements

• Standalone installations carry a private, verified Node.js 26.5.0 runtime (module ABI 147); no host Node installation is required. Source checkouts, the npm launcher, and the SDK require Node.js >=26.5 <27; see the supported-host matrix.
• npm 11.17.0 (exactly pinned by packageManager and devEngines).
• ripgrep (rg) on PATH for file search (agenc doctor reports status).
• Linux source builds require a C11 compiler available as cc to build the bundled process-containment broker. Published standalone/runtime artifacts already include that broker.
• A provider before real model calls. Default: xAI via XAI_API_KEY (also accepts GROK_API_KEY); default model grok-4.6 (AGENC_MODEL overrides). Inspect with agenc providers, agenc login, agenc config.

Quick start

Five-minute path: docs/quickstart.md. Install details: docs/install.md. Migrations: OpenClaw · Hermes.

curl -fsSL https://get.agenc.ag/install.sh | sh

Or:

npm install -g @tetsuo-ai/agenc
agenc --help

From this checkout:

npm ci
npm run build
node runtime/bin/agenc --help
npm run start                 # interactive TUI

First-run / re-run setup:

agenc onboard                 # provider, key, theme, first chat
agenc onboard --status        # non-interactive posture for scripts
agenc security audit
agenc doctor

Print / one-shot:

agenc --no-tui "summarize this repository"
agenc --print --output-format stream-json "summarize this repository"

Usage

Full flag and subcommand reference: docs/reference/cli.md. agenc help / agenc help <command> print live help (top-level formatCliHelpText is not exhaustive; dispatch is).

Command surface

agenc [options] [PROMPT]
agenc -p|--print [options] [PROMPT]
agenc --no-tui [options] [PROMPT]
agenc help [command]
agenc init [--force]
agenc doctor [--json]
agenc onboard [--status [--json] | --reset]
agenc onboard identity|channel|autonomy|recap
agenc update [--check] [--pin <x.y.z>] [--json]
agenc security audit [--json] [--fix]
agenc gateway run [--stdio] [--webchat] [--heartbeat] [--hooks]
agenc gateway status [--json]
agenc gateway pairing list [--json] | pairing revoke <channel> <peerId>
agenc gateway install-service
agenc budget status [--json]
agenc remote on|off|status
agenc login | logout | whoami
agenc providers [--json] [--no-local-check]
agenc config <show|get|set|unset|validate|edit|path>
agenc plugin|plugins <command>
agenc permissions <list|approve|revoke>
agenc state export <agent-id> | import
agenc trajectories export [--format sft|dpo] [--dir <path>] [--out <file>]
agenc daemon start [--foreground] | stop | status | reload | restart
agenc agent start|list|attach|stop|logs
agenc mcp <serve|add|list|get|remove|…|doctor|xaa>

Common session flags:

-p, --print
--output-format <text|json|stream-json>
--input-format <stream-json>
--no-tui
--bare
--light
-c, --continue
-r, --resume <session-id>
--profile <name>
--provider <name>
--model <id|provider:id>
--permission-mode <mode>
--autonomous
--dangerously-bypass-approvals-and-sandbox
--image <file|url|data-url>

Light mode is experimental and off by default. Use agenc --light (or agenc -p --light "your task") for a new session that starts with eight tools for reading, editing, writing, shell execution, polling, file/content search, and tool discovery. It reduces the tool descriptions sent initially; lower total token use, faster completion, and equivalent task performance are not guaranteed.

Light keeps the standard system instructions and complete parameter documentation. Other enabled tools are discovered through system.searchTools; agent and skill catalogs appear when their tools are loaded. If policy disables discovery, the remaining permitted tools stay visible. Project instructions, memory, approvals, sandboxing, hooks, deadlines, and completion checks remain in effect. The mode belongs to the session, survives resume,