AgenC Core
Open-source terminal AI coding agent that works with any model and posts a strong Terminal-Bench score at low cost.
- Source repo
- tetsuo-ai/agenc-core
- Stars
- ★ 75
- Last updated
- today
- License
- MIT
- Primary language
- TypeScript
- FA score
- 74/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platform
- Cost
- Free software; you pay for model usage
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- Developers who want an agent to find regressions, patch them, and rerun tests until green from the terminal.
- Not a fit if
- Developers who want a GUI IDE rather than a terminal
- Users expecting zero-configuration setup
- Source review
- 74/100 · Some gaps
What does this agent do, and when should you use it?
AgenC Core is the implementation repository behind the agenc CLI. A local daemon owns agent and session lifecycle, permissions, provider calls, and command execution, while the interactive React TUI, headless print mode (agenc -p / --no-tui), background agents, channel gateway, and remote phone bridge all act as clients of that single control plane per AGENC_HOME. The agent ships built-in tools for Bash, file read/write/edit, transactional apply_patch, web fetch/search, LSP, MCP, sub-agents, and an isolated Chromium browser driven over a CDP pipe. It supports 17 built-in providers (default grok) plus local models through Ollama and LM Studio, and layers permission modes, an opt-in OS sandbox, and agenc security audit on top. MIT licensed, versioned 0.19.19 for runtime and launcher, with a separately versioned embedding SDK @tetsuo-ai/agenc-sdk 0.3.0.
You run agenc with a prompt in a terminal; the launcher starts the app-server daemon, which owns sessions, permissions, and provider calls. The agent then exercises built-in tools: Bash for commands, read/write/edit for files, transactional apply_patch, web fetch/search, LSP, MCP, and a Browser tool that drives an isolated Chromium through a CDP pipe with an SSRF-blocking egress proxy. Sessions persist as append-only rollout logs plus SQLite state and can be resumed with --continue / --resume. Background agents are managed by agenc agent start|list|attach|stop|logs. The gateway (agenc gateway run) binds Telegram, Discord, Slack, WebChat, and stdio channels with pairing and in-channel approvals, and agenc budget status reports budget-bounded execution. Headless runs emit text, json, or stream-json for scripts and CI, agenc mcp serve exposes the product as an MCP server, and @tetsuo-ai/agenc-sdk provides a typed socket/subprocess client for embedding.
- Developers who want an agent to find regressions, patch them, and rerun tests until green from the terminal.
- CI and scripting setups that call agenc -p --output-format stream-json for automated summaries or edits.
- Teams that prefer local models through Ollama or LM Studio instead of sending code to a cloud provider.
- Solo developers and operators who start background agents and follow them from a paired phone.
- Engineering teams embedding the daemon via @tetsuo-ai/agenc-sdk or consuming it as an MCP server.
- Groups collaborating with an agent through the Telegram/Discord/Slack gateway channels.
How do you install or deploy this agent?
The official installer is the recommended path; standalone installs bundle a private Node.js 26.12.1 runtime:
curl -fsSL https://get.agenc.ag/install.sh | shAlternatively install via npm (requires Node.js >=26.5 <27):
npm install -g @tetsuo-ai/agenc
agenc --helpOr build from a checkout:
npm ci
npm run build
node runtime/bin/agenc --help
npm run startMake sure ripgrep is on PATH; Linux source builds also need a C11 compiler available as cc.
How do you use this agent?
Run guided onboarding first, then supply provider credentials (xAI is the default):
agenc onboard
export XAI_API_KEY=your_key
agenc doctorInteractive and headless invocations:
agenc
agenc --no-tui "summarize this repository"
agenc --print --output-format stream-json "summarize this repository"Daemon and background agents:
agenc daemon status
agenc agent start "fix the failing parser test"
agenc agent listConfigure providers, permissions, and budgets in config.toml through agenc config set. AGENC_HOME (default ~/.agenc) must live on a local, single-host filesystem.
What are this agent's strengths and limitations?
- A daemon-plus-clients model gives the TUI, headless mode, background agents, and gateway one shared control plane.
- Broad model coverage: 17 built-in providers plus local runtimes such as Ollama and LM Studio, so no single-vendor lock-in.
- Concrete cost evidence: the README reports 76.2% on Terminal-Bench 2.1 for $34 with DeepSeek V4.1 Flash.
- Layered safety: permission modes, opt-in OS sandboxing (bubblewrap/Landlock, Seatbelt), agenc security audit, and transactional apply_patch.
- Runs as a CLI, embeds through @tetsuo-ai/agenc-sdk, and can expose itself as an MCP server with agenc mcp serve.
- MIT licensed with a type-clean codebase (0 @ts-nocheck), SBOM tooling, and verifiable npm provenance.
- Tight runtime requirements: source and npm installs need Node.js >=26.5 <27 and npm pinned at 11.17.0.
- Durable scheduled prompts require Linux directory descriptors and a local OS-account home, and upgrades require stopping older scheduler/gateway processes first.
- AGENC_HOME cannot sit on NFS, SMB, or multi-host container volumes, which constrains deployment.
- The authoritative npm test gate needs Docker Engine 25.0+, a specific Linux kernel, and seccomp support.
- The surface area is large (gateway, remote control, Ledger wallet, browser tool), so setup and configuration cost more than a single-purpose tool.
- The default provider is xAI/Grok, so real model calls need a key or a configured local provider.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| AgenC Core This agent | 74 · Some gaps | CLIFree + model costs | ★ 75 | today | TypeScript | — |
| FuXi Terminal AI Coding Agent | 37 · Major gaps | CLIFreemium | ★ 3.2k | 6d ago | Python | OpenAI API · Claude API |
| Orca | 59 · Major gaps | CLIFree + model costs | ★ 510 | today | Rust | — |
| agentty | 44 · Major gaps | CLIFree + model costs | ★ 617 | today | C++ | OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows a fail-closed sandbox design, per-spawn least privilege, secret scrubbing, SSRF proxy, lockfile pinning and reproducible builds, so least_privilege, sensitive_data_handling and dependency_security score high. User confirmation is largely documented rather than code-proven, deducted; data-flow transparency covers only some paths; external effects (background agents, gateway, remote control, Ledger transfer) are broad with limited default-behavior detail, deducted; rollback is only a single revert plus immutable releases with no user-facing recovery flow, deducted; source attribution cites SBOM/attestation but is not directly visible in the files, deducted.
Failure messaging is thorough (stable diagnostic codes, doctor reporting), high score. Self-consistency has tension: README says 17 providers while the description says 25, and multiple versions coexist (0.19.19/0.3.0), deducted; dependency availability hinges on Node 26.x and a private runtime, narrowing host compatibility, deducted.
Capability boundaries are very clearly documented (tool catalog, sandbox coverage matrix, platform readiness), high score. Audience/scenarios span terminal, headless, background and gateway but docs skew maintainer-facing, deducted; trigger precision relies on tool search and deferred loading, described but not code-corroborated, deducted; environment fit is constrained by Linux/macOS sandboxing and Windows fail-closed, deducted.
Information architecture and install notes are excellent, and known limitations (no Windows restricted token, macOS process-tree gaps) are candidly disclosed, high scores. Naming stability deducted for inconsistent provider counts and version statements; examples/FAQ are thin, deducted; license is only asserted MIT in README with no LICENSE body shown, deducted; versioning/changelog is scattered, deducted; maintenance responsibility is low because the publisher is unverified and only contribution flow is shown.
Output usability offers TUI/print/stream-json forms but lacks concrete output samples, deducted; marginal value cites a Terminal-Bench comparison that is self-reported, deducted; cost-benefit rests on a single benchmark figure without independent corroboration, low score.
Docs cite many primary sources and issue/PR numbers, so claim_traceability is fair; but benchmark scores and cost comparisons point only to the vendor site, limiting cross_source_corroboration; fact/inference separation is broadly reasonable yet some performance claims are not clearly distinguished, all deducted.
- README states 17 built-in providers while the repository description says 25; this inconsistency should be resolved against code.
- The Terminal-Bench 76.2%/$34 result points only to the vendor site and lacks independent corroboration; do not treat it as verified.
- Windows has no restricted-token isolation and restricted execution fails closed; macOS process-tree termination has snapshot gaps; assess platform risk.
- Publisher identity is unverified, leaving maintenance responsibility and update path unclear.
- Installation relies on curl | sh and a private Node 26.x runtime; supply-chain and host compatibility need independent review.