You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit b7a515d
Browse filesBrowse the repository at this point in the historyBrowse files
Copy file name to clipboardExpand all lines: README.md
+8-7Lines changed: 8 additions & 7 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -26,7 +26,7 @@ Meet the Agent! family's newest members: two cross-platform CLIs with the **exac
26
26
27
27
<ahref="https://github.com/AgentiLoop/AgentiLoopCLI"><imgsrc="docs/agentiloop-cli-ad-v2.svg"width="900"alt="AgentiLoop CLI: two versions, Rust and Go, with the exact same capabilities. macOS, Windows, Linux." /></a>
28
28
29
-
*Built by AgentiLoop Agent! for Mac. Yes, the agent wrote its own little siblings.* 🤖✨ **v0.0.5 is out:** a bigger toolbox — glob/grep search, web_fetch, todos, AGENTS.md, /undo, custom slash commands and --json. Download it today!
29
+
*Built by AgentiLoop Agent! for Mac. Yes, the agent wrote its own little siblings.* 🤖✨ **v0.0.10 is out:** a bigger toolbox — glob/grep search, web_fetch, todos, AGENTS.md, /undo, custom slash commands and --json. Download it today!
30
30
31
31
<ahref="https://github.com/AgentiLoop/AgentiLoopCLI"><imgsrc="docs/agentiloop-rust-cli-button.svg"width="222"height="44"alt="Rust CLI on GitHub" /></a>
32
32
<ahref="https://github.com/AgentiLoop/AgentiLoopGo"><imgsrc="docs/agentiloop-go-cli-button.svg"width="212"height="44"alt="Go CLI on GitHub" /></a>
@@ -173,6 +173,7 @@ Companies and LLM providers who want to support the project: see [docs/SPONSORSH
173
173
| Provider | Sponsor | Cost | Best for |
174
174
|---|:---:|---|---|
175
175
|**A2Agent**|| Cheap | DeepSeek, GLM, Kimi, MiniMax and Qwen via one OpenAI-compatible key at a fraction of official pricing |
176
+
|**Alibaba DashScope**|| Cheap | Qwen 3.8 via DashScope; Model Studio `sk-` and QwenCloud `sk-ws-` keys |
176
177
|**Apple Intelligence**|| Free, on-device | Triage, summaries, token compression (brain icon, not available on the provider picker) |
177
178
|**Claude**|| Per-token (API key) or subscription (OAuth) | Long autonomous tasks, extended thinking, prompt caching |
178
179
|**Codex**|| ChatGPT subscription | OpenAI models via ChatGPT OAuth — no API key, no per-token charges |
@@ -190,7 +191,6 @@ Companies and LLM providers who want to support the project: see [docs/SPONSORSH
190
191
|**OpenAI**|| Paid | General purpose, tool calling, vision, `reasoning_effort`|
191
192
|**OpenRouter**|| Paid | 200+ models, one key; Claude routed via Anthropic protocol |
192
193
|**OrcaRouter**|| Paid | 190+ models via one OpenAI-compatible key at provider cost; auto/fusion/fallback routing |
193
-
|**Alibaba DashScope**|| Cheap | Qwen 3.8 via DashScope; Model Studio `sk-` and QwenCloud `sk-ws-` keys |
Full per-action reference: [docs/TECHNICAL.md](docs/TECHNICAL.md).
217
217
218
218
## AgentScript — Swift scripts with full TCC
219
219
220
-
AgentScripts are plain Swift files in `~/Documents/AgentScript/agents/Sources/Scripts/`. Agent! compiles each one to a `.dylib` with SwiftPM (`Package.swift` lists every script plus the 51 ScriptingBridge app bridges), then `dlopen`s it with Agent!'s own TCC grants — Accessibility, Automation, Calendar, Contacts, Mail, Photos, etc. The LLM manages them with `agent_script` (`create` / `edit` / `run` / `delete` / `restore` / `pull`); ~35 examples ship in the folder (`Hello`, `TodayEvents`, `NowPlaying`, `CheckMail`, `CreateDmg`, `ArchiveXcode`, …).
220
+
AgentScripts are plain Swift files in `~/Documents/AgentScript/agents/Sources/Scripts/`. Agent! compiles each one to a `.dylib` with SwiftPM (`Package.swift` lists every script plus the 51 ScriptingBridge app bridges), then `dlopen`s it with Agent!'s own TCC grants — Accessibility, Automation, Calendar, Contacts, Mail, Photos, etc. The LLM manages them with `agent_script` (`create` / `edit` / `run` / `delete` / `restore` / `pull`); ~70 examples ship in the folder (`Hello`, `TodayEvents`, `NowPlaying`, `CheckMail`, `CreateDmg`, `ArchiveXcode`, …).
221
221
222
222
**Entry point** — no top-level code, no `exit()`; `stdout` is returned to the LLM, the return value is the exit status:
223
223
@@ -405,14 +405,15 @@ Runs the task loop in unattended cycles — on the main tab or any LLM tab — u
405
405
|---|---|
406
406
|`/auto <goal>`| Work toward the goal until the LLM reports it reached |
407
407
|`/auto 4h <goal>`| Same, but stop after 4 hours (`30m`, `1.5h` also work) |
408
+
|`/auto --worktree <goal>`| Work in an isolated git worktree and branch (`-w`); `--shared` forces the shared checkout. By default a worktree is used only when another tab is already running `/auto` on the same project |
408
409
|`/auto`| Review the project first, then ask you for the goal via `ask_user`|
409
410
|`/auto history`| List previous auto-pilot goals |
410
411
|`/auto last` / `/auto #N`| Restart the most recent / Nth goal from history |
411
412
|`/auto add <folder>`| Add a parity folder — every change to the project is mirrored there |
412
413
|`/auto remove <folder>`| Remove a parity folder |
413
414
|`/auto status` / `/auto stop`| Show the session / end it after the current cycle (immediately if between cycles) |
414
415
415
-
Each cycle is a normal task; its summary is appended to `.agent/autopilot/progress.md` and fed into the next cycle's prompt. The LLM ends the session by starting its final summary with `AUTOPILOT: GOAL REACHED`. Cycles that end with no summary (errors, cancellations) never end the session — the next cycle just waits longer (15s, 30s, 60s… up to 5 min). `/auto` commands are recorded in prompt history (arrow-up) and goals are saved to the `/auto history` list. Active sessions survive app restarts: quitting (or a crash) pauses them, and on the next launch each one resumes on its tab with the next cycle.
416
+
Each cycle is a normal task; its summary is appended to `.agent/tabs/<key>/progress.md` and fed into the next cycle's prompt. The LLM ends the session by starting its final summary with `AUTOPILOT: GOAL REACHED`. Cycles that end with no summary (errors, cancellations) never end the session — the next cycle just waits longer (15s, 30s, 60s… up to 5 min). `/auto` commands are recorded in prompt history (arrow-up) and goals are saved to the `/auto history` list. Active sessions survive app restarts: quitting (or a crash) pauses them, and on the next launch each one resumes on its tab with the next cycle. The session also ends when the LLM starts its summary with `AUTOPILOT: BLOCKED` (it needs you: input, a hands-on test, or your uncommitted changes), or after 3 completed cycles in a row with no new git commit, so it never loops on "nothing to do".
Agent! is the result of three years of building agentic AI apps — ANIE, Game Changer, BattleScript, XCF MCP Server and Client, D1F, and about eight original Swift packages. The missing piece was an intelligent autonomous loop; once achieved, the best of those projects came together into Agent!. It has written video games ([Boss-Man](https://github.com/AgentiLoop/bossman)), created apps, written poetry into Pages via AppleScript, generated disk images and attached them to GitHub releases. Where Claude Code relies on ~65 third-party NPM packages, Agent! is 100% native, uses very little RAM, and ships Xcode automation, Swift Syntax 6.2 analysis, Accessibility, AppleScript, AgentScript/ScriptingBridge, Safari automation and MCP support out of the box.
429
+
Agent! is the result of three years of building agentic AI apps — ANIE, Game Changer, BattleScript, XCF MCP Server and Client, D1F, and about eight original Swift packages. The missing piece was an intelligent autonomous loop; once achieved, the best of those projects came together into Agent!. It has written video games ([Boss-Man](https://github.com/AgentiLoop/bossman) · [Umbravault](https://github.com/AgentiLoop/Umbravault) · [BvS](https://github.com/AgentiLoop/bvs) · Go · Chess), created apps, written poetry into Pages via AppleScript, generated disk images and attached them to GitHub releases. Where Claude Code relies on ~65 third-party NPM packages, Agent! is 100% native, uses very little RAM, and ships Xcode automation, Swift Syntax 6.2 analysis, Accessibility, AppleScript, AgentScript/ScriptingBridge, Safari automation and MCP support out of the box.
Copy file name to clipboardExpand all lines: readme/README_de.md
+8-7Lines changed: 8 additions & 7 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -26,7 +26,7 @@ Lerne die neuesten Mitglieder der Agent!-Familie kennen: zwei plattformübergre
26
26
27
27
<ahref="https://github.com/AgentiLoop/AgentiLoopCLI"><imgsrc="../docs/agentiloop-cli-ad-v2_de.svg"width="900"alt="AgentiLoop CLI: zwei Versionen, Rust und Go, mit exakt denselben Fähigkeiten. macOS, Windows, Linux." /></a>
28
28
29
-
*Erstellt mit AgentiLoop Agent! für Mac. Ja, der Agent hat seine eigenen kleinen Geschwister geschrieben.* 🤖✨ **v0.0.5 ist da:** ein größerer Werkzeugkasten mit glob/grep-Suche, web_fetch, Todos, AGENTS.md, /undo, eigenen Slash-Befehlen und --json. Jetzt herunterladen!
29
+
*Erstellt mit AgentiLoop Agent! für Mac. Ja, der Agent hat seine eigenen kleinen Geschwister geschrieben.* 🤖✨ **v0.0.10 ist da:** ein größerer Werkzeugkasten mit glob/grep-Suche, web_fetch, Todos, AGENTS.md, /undo, eigenen Slash-Befehlen und --json. Jetzt herunterladen!
30
30
31
31
<ahref="https://github.com/AgentiLoop/AgentiLoopCLI"><imgsrc="../docs/agentiloop-rust-cli-button_de.svg"width="220"height="44"alt="Rust CLI auf GitHub" /></a>
32
32
<ahref="https://github.com/AgentiLoop/AgentiLoopGo"><imgsrc="../docs/agentiloop-go-cli-button_de.svg"width="203"height="44"alt="Go CLI auf GitHub" /></a>
@@ -173,6 +173,7 @@ Unternehmen und LLM-Anbieter, die das Projekt unterstützen möchten: siehe [doc
173
173
| Anbieter | Sponsor | Kosten | Am besten für |
174
174
|---|:---:|---|---|
175
175
|**A2Agent**|| Günstig | DeepSeek, GLM, Kimi, MiniMax und Qwen mit einem OpenAI-kompatiblen Schlüssel zu einem Bruchteil der offiziellen Preise |
176
+
|**Alibaba DashScope**|| Günstig | Qwen 3.8 über DashScope |
176
177
|**Apple Intelligence**|| Kostenlos, auf dem Gerät | Triage, Zusammenfassungen, Token-Kompression (Gehirn-Symbol, nicht in der Anbieterauswahl verfügbar) |
177
178
|**Claude**|| Pro Token (API-Key) oder Abo (OAuth) | Lange autonome Aufgaben, Extended Thinking, Prompt-Caching |
178
179
|**Codex**|| ChatGPT-Abo | OpenAI-Modelle über ChatGPT-OAuth — kein API-Key, keine Token-Kosten |
@@ -190,7 +191,6 @@ Unternehmen und LLM-Anbieter, die das Projekt unterstützen möchten: siehe [doc
Vollständige Referenz pro Aktion: [docs/TECHNICAL.md](../docs/TECHNICAL.md).
217
217
218
218
## AgentScript — Swift-Skripte mit vollem TCC
219
219
220
-
AgentScripts sind einfache Swift-Dateien in `~/Documents/AgentScript/agents/Sources/Scripts/`. Agent! kompiliert jede davon per SwiftPM zu einer `.dylib` (`Package.swift` listet jedes Skript plus die 51 ScriptingBridge-App-Bridges) und lädt sie per `dlopen` mit Agent!s eigenen TCC-Freigaben — Bedienungshilfen, Automation, Kalender, Kontakte, Mail, Fotos usw. Das LLM verwaltet sie über `agent_script` (`create` / `edit` / `run` / `delete` / `restore` / `pull`); ~35 Beispiele liegen im Ordner (`Hello`, `TodayEvents`, `NowPlaying`, `CheckMail`, `CreateDmg`, `ArchiveXcode`, …).
220
+
AgentScripts sind einfache Swift-Dateien in `~/Documents/AgentScript/agents/Sources/Scripts/`. Agent! kompiliert jede davon per SwiftPM zu einer `.dylib` (`Package.swift` listet jedes Skript plus die 51 ScriptingBridge-App-Bridges) und lädt sie per `dlopen` mit Agent!s eigenen TCC-Freigaben — Bedienungshilfen, Automation, Kalender, Kontakte, Mail, Fotos usw. Das LLM verwaltet sie über `agent_script` (`create` / `edit` / `run` / `delete` / `restore` / `pull`); ~70 Beispiele liegen im Ordner (`Hello`, `TodayEvents`, `NowPlaying`, `CheckMail`, `CreateDmg`, `ArchiveXcode`, …).
221
221
222
222
**Einstiegspunkt** — kein Top-Level-Code, kein `exit()`; `stdout` geht zurück ans LLM, der Rückgabewert ist der Exit-Status:
223
223
@@ -405,14 +405,15 @@ Führt die Aufgabenschleife in unbeaufsichtigten Zyklen aus (im Haupt-Tab oder i
405
405
|---|---|
406
406
|`/auto <goal>`| Arbeitet auf das Ziel hin, bis das LLM es als erreicht meldet |
407
407
|`/auto 4h <goal>`| Ebenso, aber Stopp nach 4 Stunden (`30m` und `1.5h` gehen auch) |
408
+
|`/auto --worktree <goal>`| Arbeitet in einem isolierten Git-Worktree und Branch (`-w`); `--shared` erzwingt den gemeinsamen Checkout. Standardmäßig wird ein Worktree nur genutzt, wenn bereits ein anderer Tab `/auto` im selben Projekt ausführt |
408
409
|`/auto`| Prüft zuerst das Projekt und fragt dann per `ask_user` nach dem Ziel |
409
410
|`/auto history`| Listet frühere Autopilot-Ziele auf |
410
411
|`/auto last` / `/auto #N`| Startet das neueste / das N-te Ziel aus dem Verlauf neu |
411
412
|`/auto add <folder>`| Fügt einen Paritätsordner hinzu: jede Projektänderung wird dorthin gespiegelt |
412
413
|`/auto remove <folder>`| Entfernt einen Paritätsordner |
413
414
|`/auto status` / `/auto stop`| Zeigt die Sitzung / beendet sie nach dem aktuellen Zyklus (sofort, wenn zwischen Zyklen) |
414
415
415
-
Jeder Zyklus ist eine normale Aufgabe; seine Zusammenfassung wird an `.agent/autopilot/progress.md` angehängt und in den Prompt des nächsten Zyklus übernommen. Das LLM beendet die Sitzung, indem es seine finale Zusammenfassung mit `AUTOPILOT: GOAL REACHED` beginnt. Zyklen ohne Zusammenfassung (Fehler, Abbrüche) beenden die Sitzung nie, der nächste Zyklus wartet nur länger (15 s, 30 s, 60 s … bis 5 min). `/auto`-Befehle landen im Prompt-Verlauf (Pfeil hoch), Ziele in der Liste von `/auto history`. Aktive Sitzungen überstehen App-Neustarts: Beenden (oder ein Absturz) pausiert sie, und beim nächsten Start setzt jede in ihrem Tab mit dem nächsten Zyklus fort.
416
+
Jeder Zyklus ist eine normale Aufgabe; seine Zusammenfassung wird an `.agent/tabs/<key>/progress.md` angehängt und in den Prompt des nächsten Zyklus übernommen. Das LLM beendet die Sitzung, indem es seine finale Zusammenfassung mit `AUTOPILOT: GOAL REACHED` beginnt. Zyklen ohne Zusammenfassung (Fehler, Abbrüche) beenden die Sitzung nie, der nächste Zyklus wartet nur länger (15 s, 30 s, 60 s … bis 5 min). `/auto`-Befehle landen im Prompt-Verlauf (Pfeil hoch), Ziele in der Liste von `/auto history`. Aktive Sitzungen überstehen App-Neustarts: Beenden (oder ein Absturz) pausiert sie, und beim nächsten Start setzt jede in ihrem Tab mit dem nächsten Zyklus fort. Die Sitzung endet auch, wenn das LLM seine Zusammenfassung mit `AUTOPILOT: BLOCKED` beginnt (es braucht dich: eine Eingabe, einen Praxistest oder deine nicht committeten Änderungen), oder nach 3 abgeschlossenen Zyklen in Folge ohne neuen Git-Commit, damit es nie bei „nichts zu tun“ hängen bleibt.
Agent! ist das Ergebnis von drei Jahren Entwicklung agentischer KI-Apps — ANIE, Game Changer, BattleScript, XCF MCP Server and Client, D1F und rund acht originale Swift-Pakete. Das fehlende Puzzlestück war eine intelligente autonome Schleife; sobald sie erreicht war, floss das Beste dieser Projekte in Agent! zusammen. Es hat Videospiele geschrieben ([Boss-Man](https://github.com/AgentiLoop/bossman)), Apps erstellt, Gedichte per AppleScript in Pages verfasst, Disk-Images erzeugt und an GitHub-Releases angehängt. Wo Claude Code auf ~65 NPM-Pakete von Drittanbietern setzt, ist Agent! zu 100 % nativ, braucht sehr wenig RAM und bringt Xcode-Automatisierung, Swift-Syntax-6.2-Analyse, Accessibility, AppleScript, AgentScript/ScriptingBridge, Safari-Automatisierung und MCP-Unterstützung direkt mit.
429
+
Agent! ist das Ergebnis von drei Jahren Entwicklung agentischer KI-Apps — ANIE, Game Changer, BattleScript, XCF MCP Server and Client, D1F und rund acht originale Swift-Pakete. Das fehlende Puzzlestück war eine intelligente autonome Schleife; sobald sie erreicht war, floss das Beste dieser Projekte in Agent! zusammen. Es hat Videospiele geschrieben ([Boss-Man](https://github.com/AgentiLoop/bossman) · [Umbravault](https://github.com/AgentiLoop/Umbravault) · [BvS](https://github.com/AgentiLoop/bvs) · Go · Schach), Apps erstellt, Gedichte per AppleScript in Pages verfasst, Disk-Images erzeugt und an GitHub-Releases angehängt. Wo Claude Code auf ~65 NPM-Pakete von Drittanbietern setzt, ist Agent! zu 100 % nativ, braucht sehr wenig RAM und bringt Xcode-Automatisierung, Swift-Syntax-6.2-Analyse, Accessibility, AppleScript, AgentScript/ScriptingBridge, Safari-Automatisierung und MCP-Unterstützung direkt mit.
0 commit comments