Aiden is the autonomous-agent project that doesn’t want your data: local-first, BYOK (bring your own keys), Windows-native, and shipped solo. It runs your browser, your terminal, your files, and keeps going across long tasks without phoning home. Six hundred stars in four months from people who wanted an agent engine they could actually own.
- 621 stars
- 113 forks
- TypeScript
- AGPL-3.0 License
Local-first means actually local-first
The architectural choices are the differentiator. Aiden runs entirely on your machine: the model calls go to whatever provider you configure (Groq, Anthropic, OpenAI, Gemini, Ollama for fully local), but your files, terminal, browser sessions, and memory never leave your disk. SQLite is the persistence layer. No telemetry, no cloud relay, no “we’ll sync your context for you”.
The license reinforces the stance: AGPL-3.0, not MIT. The maintainer explicitly wants derivatives to stay open. That’s a strong signal about who this is for: operators who want to own the agent, not rent one from a vendor.
76 skills, 121 tools, 19 providers, and one human
The scope is wider than most solo projects survive. Skills (high-level capabilities like “operate the browser” or “run shell workflows”), tools (atomic actions skills compose), provider integrations (19 supported including all the majors and Ollama for fully-local inference), and channels (Discord, Slack, Telegram, Email, Webhook, plus four more), all wired into a coherent runtime. The numbers aren’t marketing; they’re in the boot card.
Browser control uses Playwright with accessibility-tree snapshots and stable element handles. Clicks land on referenced elements, not coordinate guesses. Terminal execution is sandboxed when Docker is available. Memory splits into three namespaces (user, project, session) so Aiden doesn’t conflate personal preferences with task-specific context.
Sub-agents, durable jobs, and verification before claiming success
The architecture that ships in v4.13+ is the real headline. Every task gets a job-card. Aiden verifies work before claiming success, resumes after restart, avoids duplicate sends, retries recoverable failures, and stops to ask for permission on destructive actions. Sub-agents must return evidence (a file, a test result, an ID) and Aiden re-checks that evidence before trusting it.
The “trust dial” lets you set how autonomous Aiden gets: Observer (watches and suggests), Assistant (acts with approval), Partner (acts and reports). Destructive actions and spending still prompt, even at the highest tier. It’s the right design: autonomy and safety aren’t pretending to be the same slider.
Built solo, on a real release cadence
The README is explicit about this: Built solo. Still rough in spots. Four major versions in four months. v4.10 made it inspectable (event ledger, task lifecycle, trace propagation, eval harness). v4.12 added the MCP client. v4.13 landed durable autonomy. v4.15 just shipped terminal-lifecycle cleanup and queueable input. The commit cadence is the proof: this isn’t a launch-day repo that went quiet.
That solo-build context is also the honest caveat. A 800-star agent engine maintained by one person has bus-factor risk. The AGPL license and the active fork count (113) are partial mitigations, but if you’re betting production infrastructure on Aiden, factor in the maintainer’s continued involvement.
Run it
60-second install:
npm install -g aiden-runtime
aidenPick a provider (Groq is free + fastest), paste your key, start chatting. For autonomous triggers, daemon mode watches a folder and acts:
export AIDEN_DAEMON=1
aiden trigger add file --path ~/Documents/inbox --label "watch-inbox" --include "*.txt"
aidenDrop a file in the watched folder and Aiden acts on it. aiden runs list shows every agent turn.
Aiden is worth a serious look if you’ve been uncomfortable with how much of the agent stack lives on someone else’s server. Install it, point it at Ollama or Groq, give it one real task, and watch the operator TUI. Whether you keep using it or not, you’ll come away with a clearer opinion on what local-first agents should look like.