Aiden
A durable local work engine that executes and verifies multi-step computer tasks.
Per-dimension scores and reasoning
Evidence shows risk-tiered approvals, trust levels, budgets, scoped permissions, leases, and generation fencing, plus a 10-module security moat (moat/) gating tool calls, including dangerous-command classifier, SSRF-safe URL fetcher, secret/PII pre-write scanner. User confirmation via approval panels, destructive actions ask even at highest trust. Data flow transparency via tracking of Jobs, Attempts, Effects, Evidence, Verification, Verdicts, and Proof, but no detailed logs or audit trail. Sensitive data: API keys stored in local .env, OAuth tokens with 0600 permissions, but no explicit encryption or leak protection. Dependency security: npm audit CI step (high severity, continue-on-error) and overrides, but no vulnerability scan results. External effects: browser and terminal execution, but controlled via approvals and sandbox (Docker), not all external side effects specified. Rollback: /undo command and conflict evidence, but no full rollback mechanism. Source attribution: author and repo clear, but publisher unverified.
Self-consistency: README, package.json, and CI config consistent on version (4.19.1) and features, but skill/tool counts differ slightly (76 vs 69+). Dependency availability: dependency list complete, overrides present, but no lockfile or integrity checks. Failure messages: error recovery system (TCE) and failure-injection coverage, but no specific failure message examples.
Audience and scenarios: README lists multiple use cases (codebase mode, research, browser workflows, etc.), but no detailed scenario guides. Capability boundaries: trust levels and approvals, but not all tool capability boundaries listed. Trigger precision: triggers (file, schedule, email, webhook), but no precise trigger condition docs. Environment fit: supports Windows, Linux, WSL, macOS API mode, but no detailed system requirements.
Information architecture: README, docs/ directory, CHANGELOG.md, but no full architecture docs. Install notes: npm install and npx commands, but no detailed install steps or troubleshooting. Naming stability: version numbers stable, but no naming convention docs. Examples and FAQ: example prompts, but no FAQ. Known limitations: known issue (terminal resize), but no comprehensive limitations list. License: AGPL-3.0, LICENSE file present, complete. Versioning and changelog: CHANGELOG.md and version numbers, but no detailed changelog. Maintenance responsibility: author and contact, but no maintenance plan.
Output usability: evidence, verification, and verdicts, but no detailed output format specs. Marginal value: unique features (evidence tracking, recovery), but no comparison with other tools. Cost-benefit: budgets and cost tracking, but no pricing or cost analysis.
Claim traceability: README claims partially supported (e.g., skill count), but no detailed evidence. Cross-source corroboration: CI and tests present, but no independent verification. Fact-inference separation: evidence and verification, but no explicit separation of facts and inferences.
- Publisher identity unverified; treat with caution.
- Dependency security scan results not provided, and npm audit is continue-on-error, potentially masking vulnerabilities.
- API server defaults to unauthenticated (loopback); set AIDEN_API_KEY.
- Skill and tool counts inconsistent across docs; verify.
- Rollback mechanism incomplete; only /undo command provided.
What does this agent do, and when should you use it?
Aiden is an autonomous work engine for computer tasks, exposed through an operator TUI, one-shot CLI mode, an opt-in daemon, channel adapters, and MCP. It represents a request as a Job and runs it through fenced Attempts with generations, leases, budgets, and cancellation state. Consequential operations are recorded as Effects, while Evidence, Verification, Verdicts, and Proof separate execution from a claim of success. Its Codebase Mode snapshots repository and VCS state before making conflict-safe changes and recording structured test, build, and diagnostic results. The runtime uses Node.js, TypeScript, SQLite, Playwright, and Chromium, with stated support for Windows, Linux, WSL, and macOS API mode.
Start the interactive Operator TUI with aiden, or submit a one-shot task with aiden -q "...". Aiden can read and write files, execute shell commands and code, manage processes, search and fetch web content, automate a browser through Playwright/Chromium, call APIs and MCP servers, and work with supported applications and channels. In Codebase Mode, it snapshots repository root and VCS state, branch, HEAD, dirty and untracked files, manifests, lockfiles, and build/test commands before mutations; each change is tied to an Attempt, source snapshot, and Effect, then read back with hashes and diffs. TestRun, BuildRun, and Diagnostic records retain commands, parsed outcomes, artifacts, failures, and verification state, while Git branch, stage, commit, fetch, pull, push, and tag actions are reconcilable Effects. With AIDEN_DAEMON=1, file watchers, webhooks, email polling, and schedules can dispatch agent turns through a SQLite-backed trigger bus.
- A maintainer of a Node.js repository can ask Aiden to snapshot the repo, diagnose failing tests, apply safe edits, rerun relevant tests and builds, and retain verification evidence.
- An operations user handling a browser-based back-office process can have it navigate supported sites, fill forms, handle tabs and dialogs, and capture outcomes.
- A Windows, Linux, or WSL user can use the terminal interface for file organization, system inspection, and repeatable shell procedures.
- A team that needs work triggered by incoming text files can enable the daemon and register a watched inbox with
aiden trigger add file. - A researcher who needs a saved deliverable can ask it to search, compare sources, collect findings, and produce a structured Markdown report.
- A user breaking down a larger job can delegate focused work to isolated sub-agents with intersected toolsets and evidence-required returns.
What are this agent's strengths and limitations?
- Completion is evidence-oriented: Jobs, Attempts, Effects, Evidence, Verification, Verdicts, and Proof are tracked as distinct states rather than trusting a command exit code or model response.
- Codebase Mode snapshots repository and VCS state before writes and blocks stale bases, unsafe paths, symlink escapes, and conflicting user edits.
- The documented model path spans 19 providers, Ollama, custom OpenAI-compatible endpoints, and ChatGPT OAuth subscription routing where supported.
- Interactive, one-shot, daemon, channel, MCP, and browser surfaces share durable execution authority and restart/recovery protections.
- Adoption requires Node.js ≥18, npm, provider credentials, and a suitable model; model quality and budget materially affect complex-work outcomes.
- It can read and write files, execute shell commands, access the web, and connect to external services, so permission scope, credentials, and backups need deliberate handling.
- macOS support is stated as API mode, and some capabilities are explicitly described as experimental or platform-dependent.
- The documented known issue is cosmetic activity-row artifacts during aggressive terminal resizing, although the project reports no duplicate execution or durable-state corruption.
- Daemon, email, webhook, and browser workflows need further configuration; daemon mode is off by default.
How do you install or deploy this agent?
Requires Node.js ≥18, npm, and authentication for the provider you choose. Install and start:
npm install -g aiden-runtime
aidenThe first launch opens setup, where you select a provider, retrieve available models where supported, and run a connection probe. To run without a global installation:
npx aiden-runtime@latestVerify with aiden --version; use aiden doctor to diagnose providers, authentication, configuration, and environment.
How do you use this agent?
Give Aiden a complete goal in the TUI, or run:
aiden -q "Inspect this repository, explain how it works, identify the three highest-risk areas, and support the findings with repository Evidence."Use /model to select a provider and model, or set AIDEN_DEFAULT_PROVIDER and AIDEN_DEFAULT_MODEL. For triggered background work:
export AIDEN_DAEMON=1
aiden trigger add file --path ~/Documents/inbox --label "watch-inbox" --include "*.txt"
aidenInspect triggered activity with aiden runs list. The daemon binds to 127.0.0.1 by default; a non-loopback bind requires AIDEN_API_KEY.
FAQ
Am I locked into one model provider?
How are risky actions controlled?
What happens if a task is interrupted or an external result is uncertain?
Can I use it without a background service?
aiden for the terminal interface or aiden -q "..." for a one-shot task. The daemon is opt-in.