apfel

The free AI already on your Mac.

Apple Silicon Macs ship a built-in LLM via Apple FoundationModels. apfel exposes it as a UNIX tool and a local OpenAI-compatible server. 100% on-device. No API keys, no cloud.

| Mode | Command | What you get | |------|---------|--------------| | UNIX tool | apfel "prompt" / echo "text" \| apfel | Pipe-friendly answers, file attachments, JSON output, exit codes | | OpenAI-compatible server | apfel --serve | Drop-in local http://localhost:11434/v1 backend for OpenAI SDKs |

apfel --chat - interactive REPL.

Tool calling works in all contexts. On-device context window: 4096 tokens on macOS 26 and on M1/M2 Macs, 8192 on M3+ Macs with 12 GB+ running macOS 27 - read at runtime, see Limitations.

!apfel CLI

Requirements & Install

macOS 26 Tahoe+, Apple Silicon (M1+), Apple Intelligence enabled. On macOS 27, Macs with an M3 or newer chip and 12 GB+ run Apple's larger on-device model with an 8192-token window; M1/M2 Macs and every Mac on macOS 26 have 4096. apfel --model-info shows yours.

brew install apfel

Update:

brew upgrade apfel

Build from source (Command Line Tools with macOS 26.4 SDK / Swift 6.3, no Xcode):

git clone https://github.com/Arthur-Ficial/apfel.git && cd apfel && make install

Nix, same-day tap, Mint, mise, troubleshooting: docs/install.md.

Quick Start

UNIX tool

Quote prompts with ! in single quotes (zsh/bash history expansion): apfel 'Hello, Mac!'.

# Single prompt
apfel "What is the capital of Austria?"
Permissive mode - reduces guardrail false positives for creative/long prompts
apfel --permissive "Write a dramatic opening for a thriller novel"
Stream output
apfel --stream "Write a haiku about code"
Pipe input
echo "Summarize: $(cat README.md)" | apfel
Attach file content to prompt
apfel -f README.md "Summarize this project"
Attach multiple files
apfel -f old.swift -f new.swift "What changed between these two files?"
Attach a PDF or image - on-device text extraction, OCR, and "what the image is about"
apfel -f report.pdf "Summarize the key findings"
apfel -f receipt.jpg "What is the total?"
Pipe a file straight in (PDF, image, or text)
cat report.pdf | apfel "Summarize this"
Combine files with piped input
git diff HEAD~1 | apfel -f CONVENTIONS.md "Review this diff against our conventions"
Only the code - no prose, no markdown fences (pipe-safe, exit 7 if empty)
apfel --code "a python function that deduplicates a list" > dedupe.py
apfel --code "shell one-liner to find the 10 largest files here" | pbcopy
JSON output for scripting
apfel -o json "Translate to German: hello" | jq .content
Guaranteed schema-valid JSON output (guided generation)
apfel --schema person.schema.json "Extract the person: Alice is 30." | jq .name
One-shot multi-turn: conversation JSON in, next assistant turn out
jq '. += [{"role":"user","content":"and in German?"}]' conv.json | apfel --messages -
Preflight token budget before a large prompt
apfel --count-tokens -f README.md "Summarize this"
System prompt
apfel -s "You are a pirate" "What is recursion?"
System prompt from file
apfel --system-file persona.txt "Explain TCP/IP"
Quiet mode for shell scripts
result=$(apfel -q "Capital of France? One word.")

OpenAI-compatible server

apfel --serve                              # foreground
brew services start apfel                  # background (like Ollama)
brew services stop apfel
APFEL_TOKEN=$(uuidgen) APFEL_MCP=/path/to/tools.py brew services start apfel

curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"apple-foundationmodel","messages":[{"role":"user","content":"Hello"}]}'

from openai import OpenAI
client = OpenAI(base_url="http://localhost:11434/v1", api_key="unused")
resp = client.chat.completions.create(
    model="apple-foundationmodel",
    messages=[{"role": "user", "content": "What is 1+1?"}],
)
print(resp.choices[0].message.content)

Background service details: docs/background-service.md.

Quick testing chat

apfel --chat is a small REPL for testing prompts or MCP servers. For a GUI chat app, see apfel-chat.

apfel --chat
apfel --chat -s "You are a helpful coding assistant"
apfel --chat --mcp ./mcp/calculator/server.py      # chat with MCP tools
apfel --chat --debug                                # debug output to stderr

Ctrl-C exits. Context is trimmed automatically (docs/context-strategies.md).

Demos

Real shell scripts that wrap apfel: cmd (English to shell command), oneliner, mac-narrator, wtd, explain, naming, port, gitsum. They are bundled in the binary - no repo clone. Write them out wherever you installed apfel from (homebrew-core, the tap, or source):

apfel demos ./apfel-demos

That writes every demo (executable) plus a README.md into ./apfel-demos. Then run them from there:

./apfel-demos/cmd "find all .log files modified today"
$ find . -name "*.log" -type f -mtime -1
./apfel-demos/cmd -x "show disk usage sorted by size"   # -x = execute after confirm
./apfel-demos/cmd -c "list open ports"                   # -c = copy to clipboard

Re-run apfel demos after brew upgrade apfel to refresh. Sources: demo/. Walkthroughs and a copy-paste cmd shell function for your .zshrc: docs/demos.md.

MCP Tool Support

Attach Model Context Protocol servers with --mcp. apfel discovers, invokes, and returns.

apfel --mcp ./mcp/calculator/server.py "What is 15 times 27?"

mcp: ./mcp/calculator/server.py - add, subtract, multiply, divide, sqrt, power, round_number    ← stderr
tool: multiply({"a": 15, "b": 27}) = 405                                          ← stderr
15 times 27 is 405.                                                                ← stdout

Use -q to suppress tool info.

apfel --mcp ./server_a.py --mcp ./server_b.py "Use both tools"
apfel --serve --mcp ./mcp/calculator/server.py
apfel --chat --mcp ./mcp/calculator/server.py

Ships with a calculator at mcp/calculator/ (docs/mcp-calculator.md).

Remote MCP servers (Streamable HTTP, MCP spec 2025-03-26):

apfel --mcp https://mcp.example.com/v1 "what tools do you have?"
bearer token - prefer env var (flag is visible in ps aux)
APFEL_MCP_TOKEN=mytoken apfel --mcp https://mcp.example.com/v1 "..."
mixed local + remote
apfel --mcp /path/to/local.py --mcp https://remote.example.com/v1 "..."

Security: prefer APFEL_MCP_TOKEN over --mcp-token (ps aux). apfel refuses bearer tokens over plaintext http://.

apfel-run: optional config layer

apfel itself has no config file - flags + env vars, like any UNIX tool. If you want a TOML config (many MCPs, profiles, team configs in git), apfel-run is an MIT wrapper that adds one via execve drop-in.

brew install Arthur-Ficial/tap/apfel-run
apfel-run config init                 # starter ~/.config/apfel/config.toml
alias apfel=apfel-run                 # optional, every apfel flag still works

OpenAI API Compatibility

Base URL: http://localhost:11434/v1

| Feature | Status | Notes | |---------|--------|-------| | POST /v1/chat/completions | Supported | Streaming + non-streaming | | POST /v1/responses | Supported | OpenAI Responses API: string/message input, instructions, streaming (canonical event sequence), text.format (incl. json_schema), function tools (non-streaming). Stateful features (previous_response_id, store: true, background, reasoning, hosted tools) return honest 501s | | GET /v1/models | Supported | Returns apple-foundationmodel | | GET /health | Supported | Model availability, context window, languages, reported model capabilities (capabilities, capabilities_reported) | | GET /v1/logs, /v1/logs/stats | Debug only | Requires --debug | | Tool calling | Supported | Native ToolDefinition + JSON detection. See docs/tool-calling-guide.md | | response_format: json_object | Supported | System-prompt injection; markdown fences stripped from output | | response_format: json_schema | Supported | Guaranteed schema-conforming output via FoundationModels DynamicGenerationSchema; works with stream: true | | temperature, top_p, max_tokens, seed | Supported | Mapped to GenerationOptions. top_p is nucleus sampling; temperature: 0 maps to greedy (deterministic). Omitting max_tokens uses the remaining context window (drop-in OpenAI semantics) - see Default response cap | | stream: true | Supported | SSE; final usage chunk only when stream_options: {"include_usage": true} (per OpenAI spec) | | usage | Supported | Real tokenizer counts on macOS 26.4+ (chars/4 estimate below). macOS 27: the runtime's own accounting incl. Apple's chat-template framing (about 46 extra prompt_tokens on a one-line prompt); macOS 26: apfel counts the message text. Details in docs/openai-api-compatibility.md | | finish_reason | Supported | stop, tool_calls, length | | Context strategies | Supported | x_context_strategy, x_context_max_turns, x_context_output_reserve extension fields | | CORS | Supported | Enable with --cors | | POST /v1/completions | 501 | Legacy text completions not supported | | POST /v1/embeddings | 501 | Embeddings not available on-device | | logprobs=true, n>1, stop, presence_penalty, frequency_penalty | 400 | Rejected explicitly. n=1 and logprobs=false are accepted as no-ops | | Multi-modal (images) | Supported on macOS 27 | image_url parts as base64 data URLs go to the on-device model natively (20 MB base64 cap, longest side downscaled to 4096 px). Remote http(s) URLs and file paths are always 400 - apfel never fetches images and never reads files for an HTTP client. On macOS 26: 400 with a clear error. See docs/openai-api-compatibility.md | | Authorization header | Supported | Required when --token is set. See docs/server-security.md |

Full API spec: openai/openai-openapi.

Default response cap (max_tokens)

When max_tokens is omitted, CLI and OpenAI-compatible server behave identically: the value flows through as nil and the model uses whatever room is left in the context window. This is drop-in OpenAI semantics - no arbitrary fallback constant.

The on-device context window holds input and output combined: 4096 tokens on macOS 26 and on M1/M2 Macs, 8192 tokens on macOS 27 with an M3 or newer chip and 12 GB+. apfel reads the real size at runtime via SystemLanguageModel.contextSize - check yours with apfel --model-info. If generation runs into the ceiling, the response ends cleanly with finish_reason: "length" and the partial content is returned (server: HTTP 200; CLI: exit 0 with a stderr warning). Pass max_tokens explicitly when you want a tighter latency budget or a known cap for your client.

Examples

# Omitted: uses remaining window, finish_reason: "stop" or "length"
curl -sS http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"apple-foundationmodel",
       "messages":[{"role":"user","content":"Reply SKIP, MOVE, or RENAME."}]}'
Explicit cap (recommended for tight latency budgets)
curl -sS http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"apple-foundationmodel","max_tokens":128,
       "messages":[{"role":"user","content":"Summarise: ..."}]}'

Picking a value

| Use case | max_tokens | |----------------------------------------|---------------| | Single-word / classification reply | 16 - 32 | | One-line instruction | 64 - 128 | | Short paragraph | 256 - 512 | | Long paragraph / structured JSON | 1024 - 2048 | | As long as the context window allows | omit it |

Keep input_tokens + max_tokens comfortably below the context window (4096 tokens on macOS 26 and on M1/M2 Macs, 8192 on macOS 27 with an M3 or newer chip and 12 GB+ (Apple's AFM 3 Core Advanced model)). If the prompt itself exceeds the window, generation cannot start and the request fails with [context overflow] (HTTP 400 / CLI exit 4). The validator rejects non-positive values (max_tokens <= 0).

CLI parity

CLI and server share one rule: omitted = use remaining window. No constant to drift. Override with --max-tokens N or APFEL_MAX_TOKENS=N.

apfel "Reply SKIP."                    # uses remaining window
apfel --max-tokens 64 "Reply SKIP."    # explicit cap
APFEL_MAX_TOKENS=2048 apfel "..."      # via env var

Permissive guardrails for the server

apfel --serve --permissive makes the server use Apple's .permissiveContentTransformations guardrails for every request the process handles. Same flag, same semantics as the CLI's --permissive (docs/PERMISSIVE.md). There is no per-request override - the server operator decides for the whole process.

apfel --serve --permissive             # every request uses permissive guardrails

Limitations

| Constraint | Detail | |------------|--------| | Context window | 4096 tokens on macOS 26 and on M1/M2 Macs, 8192 tokens on macOS 27 with an M3 or newer chip and 12 GB+ (input + output combined). Not a hardcoded constant - apfel reads SystemLanguageModel.contextSize at runtime; apfel --model-info prints the live value | | Platform | macOS 26+, Apple Silicon only | | Model | One model (apple-foundationmodel, ~3B params on-device), not configurable | | Guardrails | Apple's safety system may block benign prompts. --permissive reduces false positives (docs/PERMISSIVE.md) | | Speed | On-device, not cloud-scale - a few seconds per response | | No embeddings | Not available on-device | | Vision | Image input reaches the model only on macOS 27 (server: data URLs; CLI: -f photo.jpg and piped images, on top of the OCR text). On macOS 26 images are text-only via OCR/classification. The 3B model's scene descriptions are coarse - OCR remains the reliable channel for reading print | | Training data / knowledge cutoff | Apple has not published a precise cutoff for the on-device model. When pushed to name one, the model confabulates a different date each sample (e.g. "October 2023", "April 2023"). Treat all model self-reports about its own training as unreliable. | | No current date or real-time awareness | The model does not know today's date and has no network/clock access. If asked, it will either refuse or invent a date. Inject the current date via system prompt when arithmetic depends on it (see workaround below). |

Workaround for date-dependent prompts - inject the current date as a system message:

apfel -s "Today is $(date '+%B %d, %Y')." "Write a one-line release note dated today."

apfel --chat -s "Today is $(date '+%B %d, %Y'). You are a helpful assistant."

Note: even with an injected date the 3B model can still hallucinate (especially when asked directly about its own training cutoff). The injection helps generative prompts that use the date; it does not override the model's self-report reflex.

Background: #158.

Reference Docs

Guides to use apfel from Python, Node.js, Ruby, PHP, Bash/curl, Zsh, AppleScript, Swift, Perl, AWK - see docs/guides/index.md. Empirically tested; runnable proof at apfel-guides-lab.

• docs/install.md - install, troubleshooting, and Apple Intelligence setup
• docs/cli-reference.md - every flag, exit code, and environment variable
• docs/file-extraction.md - -f and piped PDF / image / text extraction (OCR + image understanding)
• docs/background-service.md - brew services and launchd usage
• docs/openai-api-compatibility.md - /v1/* support matrix in depth
• docs/server-security.md - origin checks, CORS, tokens, and --footgun
• docs/context-strategies.md - chat trimming strategies
• docs/mcp-calculator.md - local and remote MCP usage
• docs/tool-calling-guide.md - detailed tool-calling behavior
• docs/integrations.md - third-party tool integrations (opencode, etc.)
• docs/local-setup-with-vs-code.md - local review with apfel + a second edit/apply model in VS Code
• docs/demos.md - longer walkthroughs of the shell demos
• docs/EXAMPLES.md - 50+ real prompts with unedited output
• docs/swift-library.md - ApfelCore Swift Package for downstream developers
• docs/coreai-impact.md - why apfel runs on FoundationModels, not Apple's new Core AI (the Core ML successor)
• docs/apfel-vs-fm.md - apfel compared with Apple's own fm CLI (macOS 27), feature by feature with measured numbers

Architecture

CLI (single/stream/chat) ──┐
                           ├─→ FoundationModels.SystemLanguageModel
HTTP Server (/v1/*) ───────┘   (100% on-device, zero network)
                                ContextManager → Transcript API
                                SchemaConverter → native ToolDefinitions
                                TokenCounter → real token counts (SDK 26.4)

Swift 6.3 strict concurrency. Three targets: ApfelCore (pure logic, unit-testable, also available as a Swift Package product - linked under Reference Docs above), apfel (CLI + server), and apfel-tests (pure Swift runner, no XCTest).

Build & Test

make test                                # release build + all unit/integration tests
make preflight                           # full release qualification
make install                             # build release + install to /usr/local/bin
make build                               # build release only
make version                             # print current version
make release                             # patch release
make release TYPE=minor                  # minor release
make release TYPE=major                  # major release
swift build                              # quick debug build (no version bump)
swift run apfel-tests                    # unit tests
python3 -m pytest Tests/integration/ -v  # integration tests
apfel --benchmark -o json                # performance report

.version is the single source of truth. Only make release bumps versions. Local builds do not change the version.

The apfel tree

Projects built on apfel. Each ships as its own repo + Homebrew formula.

| Project | What it does | Install | |---------|--------------|---------| | apfel | The root. On-device FoundationModels CLI + OpenAI-compatible server. | brew install apfel | | apfel-chat | macOS chat client: streaming markdown, speech I/O, Apple Vision image analysis. | brew install Arthur-Ficial/tap/apfel-chat | | apfel-clip | Menu-bar AI actions on the clipboard: summarize, translate, rewrite. | brew install Arthur-Ficial/tap/apfel-clip | | apfel-quick | Instant AI overlay: press a key, ask, answer, dismiss. | brew install Arthur-Ficial/tap/apfel-quick | | apfelpad | Formula notepad - on-device AI as an inline cell function. | brew install Arthur-Ficial/tap/apfelpad | | apfel-mcp | Token-budget-optimized MCPs for the small on-device window: url-fetch, ddg-search, search-and-fetch. | brew install Arthur-Ficial/tap/apfel-mcp | | apfel-gui | SwiftUI debug inspector: request timeline, MCP protocol viewer, TTS/STT. | brew install Arthur-Ficial/tap/apfel-gui | | apfel-run | UNIX wrapper adding a persistent MCP registry + TOML config on top of apfel. | brew install Arthur-Ficial/tap/apfel-run | | apfel-tag | On-device content tagging CLI: pipe text in, get tags/topics/emotions out. | brew install Arthur-Ficial/tap/apfel-tag | | apfel-server-kit | Swift package for ecosystem tools: discover, spawn, and stream from a local apfel --serve. | Swift Package |

Community Projects

Built something on top of apfel? Open an issue and it can be added here.

| Project | What it does | Links | |---------|-------------|-------| | apfelclaw by @julianYaman | Local AI agent that reads files, calendar, mail, and Mac status via read-only tools | github - site | | fruit-chat by @bhaskarvilles | Browser-based chat UI that talks to apfel --serve over the OpenAI-compatible API | github | | local-claude by @lucaspwo | Claude Code wrapper that swaps in apfel as a local backend via a small Anthropic-OpenAI proxy | github | | apfeller by @hasit | App manager for local shell apps built around apfel | github - site - catalog | | apfel-for-raycast by @eggsy | Raycast command bar extension: ask, translate, explain files and directories, conversation history, custom system prompts. On-device via apfel CLI. | store - github |

Contributing

Issues and PRs welcome on any Arthur-Ficial/apfel* repo.

#agentswelcome - AI agent PRs are fine. Read the repo's CLAUDE.md, run the tests, credit the tool in a Co-Authored-By trailer. Same bar as humans: clean code, passing tests, honest limits. Most agent-friendly entry point: apfel-mcp (contribution rules).

License

MIT