# Aielia > Open-source AI agent that takes on real, multi-step work — research across a list, working through your files, drafting and sending messages — while an 11-layer harness controls what it may do, what it believes, and what a job may cost. The model proposes; the harness decides. Apache 2.0. Made by 3IVIS, built on Build A Harness (https://buildaharness.com). Aielia (`@buildaharness/aielia` on npm) is a general-purpose, everyday-use AI agent you can hand real, multi-step work to. Use it in a browser tab (https://myaielia.com/try — choose Aielia Chat, or Aielia Write for articles (alpha); bring your own model key, nothing to install), in a terminal (install with `curl -fsSL https://myaielia.com/install.sh | sh`, or `irm https://myaielia.com/install.ps1 | iex` on Windows; or run `npx @buildaharness/aielia` if you have Node), or as a native desktop app. All three front ends share one assistant class and one harness; they differ only in the storage backend (filesystem for CLI/desktop, IndexedDB for the browser). It is an early-stage public alpha (v0.3.x), not an audited security product. ## What the harness controls The harness is the process every model call goes through on every real turn, so limits hold whatever the model wants to do — they are not instructions in a prompt. - **What it may do, right now.** Every read-only tool call is checked against a live control state (normal / cautious / blocked; a deterministic ALLOW / DENY / REQUIRE_APPROVAL) before it runs. The state tightens as tool failures accumulate. - **What it may believe.** Only high-confidence model inferences reach durable memory on their own (medium ones wait for user confirmation, low ones stay session-only); each remembered fact keeps its source (user-asserted, model-inferred, observed, externally verified); contradictions are flagged. Content fetched from the web or a shell command is wrapped as untrusted data before it re-enters the conversation. - **Whether the answer holds up.** A reviewer pass runs before a reply goes out and verification partly does (2 of its 9 checks return a result in chat). Replies that used tools carry evidence and a status: verified, unverified, contradicted, or no evidence. - **What a job may cost.** For an explicit list of items, each item gets its own bounded search (calibrated after the first items), three dead ends in a row stop that item, a large batch asks for confirmation before it runs, and a hard per-turn ceiling caps total tool calls. Every item ends found, not found, or truncated while still productive. - **What happens when it goes wrong.** Failures are classified and met with a recovery strategy; the harness checkpoints mid-turn, so a crash resumes the interrupted turn instead of restarting it. - **What can't be undone.** `write_file`, `run_shell_command`, and `send_email` stage the exact action as a pending proposal and apply only on explicit approval with the matching pending-action ID (an opt-in `dangerouslySkipPermissions` setting, off by default, removes the prompt). One classification call every turn also flags consequential messages (send, delete, pay, post) in any language before the harness or any tool runs (one short classification call is spent); a classifier error resolves to "requires approval", never to low-risk. Shell access is opt-in and off by default; file tools are limited to one workspace folder with every change approved. Both exist only in the CLI and desktop app; email sending exists only in the CLI. The browser build is chat-only. Web search needs your own search key. ## Pages - [How it works](https://myaielia.com/how-it-works): the eleven layers applied to a chat message, the risk gate, the trust boundary, crash resume, the three front ends, and every way to run it. - [Install](https://myaielia.com/install): CLI, desktop and browser setup on macOS, Windows and Linux, including how to open the unsigned builds and verify checksums. - [Try it in your browser](https://myaielia.com/try): choose the hosted browser version of Aielia: Aielia Chat (https://myaielia.com/chat) for multi-step jobs, or Aielia Write (https://myaielia.com/write, alpha) for writing an article with claims checked against their sources. Bring your own API key; it stays in your browser. - [Set up OpenRouter](https://myaielia.com/setup-openrouter): get a model key, add prepaid credit and cap what the key can spend, in seven steps. - [Set up Brave Search](https://myaielia.com/setup-brave-search): get a search key and set a monthly usage limit so Aielia can find sources online, in seven steps. - [Harness comparison](https://myaielia.com/harness-comparison): Claude Code, Codex CLI, Cline, Hermes Agent, Kilo Code, OpenClaw and Pi compared layer by layer against a harness-first assistant, including what each one lacks (Pi is minimal by design), with references. - [Faceoff](https://myaielia.com/harness-faceoff): the same 16 everyday coding tasks run through Aielia, Claude Code, Codex CLI, Pi, Cline and Kilo Code on one model (DeepSeek V4 Flash), with side-by-side transcripts and a count of where each failed. One run per agent per task; Aielia is the site's own assistant. - [Harness architecture](https://myaielia.com/harness-architecture): the 11-layer control loop in detail — world model, control state, verification, recovery and cross-run learning — plus the one-loop proposer with mid-turn steering, the Trajectory Supervisor and the hierarchical goal graph. - [Harness evaluation](https://myaielia.com/harness-evaluation): measured results for the Trajectory Supervisor on a purpose-built multi-turn corpus (3 seeds, Claude Sonnet 5): on injected stalls it recovered two tasks, at higher cost. Every run has its own transcript page. ## Accounts, keys, and privacy There is no Aielia account. You bring your own key (Anthropic, OpenAI, OpenRouter). Keys are stored in plaintext — `localStorage` in the browser build, a local JSON config file for the CLI. The desktop app stores the model API key in the OS keychain instead (macOS Keychain; Linux Secret Service via secret-tool; a DPAPI-protected file on Windows). ## For developers Aielia is the reference application for Build A Harness: an open-source visual canvas, FlowSpec (a runtime-neutral JSON format), four framework adapters (LangGraph, CrewAI, Mastra, Microsoft Agent Framework), and Langfuse observability. Architecture, the build-your-own guide, framework comparisons, and the published feature-evaluation results live at https://buildaharness.com. Source: https://github.com/3IVIS/buildaharness (assistant in `packages/aielia`).