Filo

Filo is a high-performance AI coding assistant written in modern C++.

It began as the handyman for Lampo: the private, on-device AI workspace for Apple Silicon. Lampo ships on the Mac App Store under App Sandbox, so it cannot freely reach the filesystem, shell, or developer tooling. Filo was built to sit beside it and do that work: open projects, read and write files, search code, apply patches, and run commands when Lampo’s models need real hands on the machine.

After those early days, Filo grew into a fully fledged terminal coding agent that stands on its own. You can run it end-to-end without Lampo: interactive TUI, skills, session resume, and non-interactive prompter mode for scripts and CI. The Lampo partnership remains first-class when you want it.

Because Filo is written in modern C++, it can be compiled with embedded llama.cpp (FILO_ENABLE_LLAMACPP=ON) and run local GGUF models in-process, no separate inference server required. Pair that with any remote providers you configure, and Filo’s built-in smart router can blend them: prefer local for everyday work, fall back or escalate to remote when a task needs more capacity, and keep spend/quota guardrails under your control.

It runs in multiple runtime modes:

interactive terminal app (TUI)
non-interactive prompter mode for scripts/CI
MCP server over stdio
HTTP daemon exposing MCP and/or compatible chat API endpoints

Switch models with --model or /model — local, hybrid, or remote, no lock-in.

A real terminal interface	Full FTXUI TUI with streaming tool output, slash-command autocomplete, session resume, context mentions, and clipboard image paste.
Embedded local AI	C++ core can link `llama.cpp` for in-process GGUF inference; Ollama over localhost is first-class too. The HTTP daemon binds `127.0.0.1` by default.
Smart hybrid routing	In-process policies (`smart`, `fallback`, `latency`, `load_balance`) mix local and remote backends with automatic fallback, spend/quota guardrails, and complexity-based tier routing.
MCP server and client	Expose Filo’s coding tools to hosts such as Lampo, or connect Filo to external MCP servers over stdio or Streamable HTTP.
Prompter for automation	Single-shot and streaming modes for scripts and CI, with `text`, `json`, and `stream-json` output formats.
Agent Skills	On-demand instruction packages under `.filo/skills`, slash-command activation, and compatibility with common skill roots.
Embedded Python	Persistent in-process interpreter tool with optional `FILO_PYTHON_VENV` isolation.

Quick Start

Prerequisites

CMake >= 3.28
C++26 compiler (GCC 15+ or Clang 17+ recommended)
OpenSSL
Python 3 (required when FILO_ENABLE_PYTHON=ON, which is the default)

Build And Run

Linux

cmake --preset linux-debug
cmake --build --preset linux-debug
ctest --preset linux-debug --output-on-failure
cmake --build --preset linux-debug --target run_integration_tests
./build/Linux/linux-debug/filo

macOS

cmake --preset xcode-debug
cmake --build --preset xcode-debug
ctest --preset xcode-debug --output-on-failure
cmake --build --preset xcode-debug --target run_integration_tests
./build/Darwin/xcode-debug/Debug/filo

Install

# Linux
cmake --install build/Linux/linux-debug --prefix "$HOME/.local"

# macOS (Xcode preset — select the configuration explicitly)
cmake --install build/Darwin/xcode-debug --config Debug --prefix "$HOME/.local"

The executable is installed to <prefix>/bin/filo. Verify with filo --version.

Configure providers and credentials from the TUI (/settings, /model, /login) or your preferred environment. Local backends such as Ollama (http://localhost:11434 by default) and embedded llama.cpp need no cloud keys.

What Makes Filo Different

Local-first architecture

Local providers are first-class: Ollama over localhost and embedded llama.cpp for in-process GGUF inference (FILO_ENABLE_LLAMACPP=ON).
Router guardrails can exempt providers flagged as local (enforce_on_local: false), keeping embedded local backends available when remote limits are hit.
The daemon listens on 127.0.0.1 by default.

Embedded smart routing

In-process router engine with policy rules and strategies: smart, fallback, latency, load_balance.
Automatic fallback chains with per-candidate retries.
Guardrails for spend and quota reserves (max_session_cost_usd, token/request/window reserve ratios).
Auto-classifier that scores prompt complexity and routes to fast/balanced/powerful tiers.

Embedded Python runtime

Built-in python tool executes code inside an embedded interpreter.
Interpreter state persists across calls (variables/imports/functions carry over).
Optional venv isolation via FILO_PYTHON_VENV.

Feature highlights

C++26 core with streaming-first provider protocols
TUI built with FTXUI
Context mentions (@file, quoted paths, and escaped paths like @My\ Folder/file.txt)
Agent Skills support with .filo/skills and on-demand activation
Ctrl+V clipboard paste support (text paste and clipboard-image insertion as @"<path>")
First-class thread management (Ctrl+N new thread, /threads for active runtimes; Ctrl+H/Ctrl+J aliases on enhanced-keyboard terminals) — project-named tabs let you switch without tmux/screen while hidden threads keep working. On an empty prompt, Ctrl+D archives and closes the current idle secondary thread; the main thread retains double-Ctrl+D app exit. In /threads, C closes any idle non-main thread without deleting its saved session. /sessions remains the saved-conversation manager.
Session persistence and resume
Global + workspace config layering
MCP dispatcher shared across stdio and HTTP transports
OAuth and API-key credentials

Runtime Modes

Mode	Command
Interactive TUI	`filo`
Prompter (single-shot)	`filo --prompt "Summarize this diff"`
MCP over stdio	`filo --mcp stdio` or `filo --mcp stdio --headless`
MCP over TCP (HTTP `/mcp` endpoint)	`filo --mcp --headless --port 8080` or `filo --mcp tcp --headless --port 8080`
API gateway only	`filo --api --headless --port 8080`
MCP + API gateway	`filo --mcp --headless --api --port 8080`

Daemon / transport

--mcp without a value defaults to tcp.
--mcp tcp starts the HTTP daemon and exposes MCP on /mcp.
--mcp stdio runs an MCP server over standard input/output (headless).
--daemon is still accepted as a deprecated alias for --mcp tcp.
Set FILO_MCP_BEARER_TOKEN to require Authorization: Bearer <token> on /mcp.
For LAN worker deployments, use --host 0.0.0.0 only with a bearer token and network access controls.
The API gateway is off by default to keep daemon startup minimal and local-first.
--api starts the same HTTP daemon and exposes compatible chat proxy endpoints:
- GET /v1/models
- POST /v1/chat/completions
- POST /v1/messages
Combine --api with --mcp if you want both /mcp and /v1/* on one port.
Model routing in API gateway endpoints:
- policy/<policy_name> routes via Filo smart router policy.
- <provider>/<model> routes directly to a configured provider/model.
- <provider> routes to that provider's default configured model.

Useful CLI flags

--version print the Filo version and exit
--mcp [tcp|stdio] run as MCP server (default transport: tcp)
--daemon deprecated alias for --mcp tcp
--api enable optional chat API proxy mode
filo --auth <provider> [login|logout] authenticate or sign out and exit
--list-sessions list resumable sessions
--model <MODEL|PROVIDER|PROVIDER/MODEL> select a model for this process only without changing saved defaults
-r, --resume [id|index|name] resume a saved session (names are set with /rename)
--prompter force non-interactive mode
--prompt, -p prompt text
--output-format, -o one of text, json, stream-json
--input-format one of text, stream-json
--include-partial-messages include deltas in stream-json
-c, --continue continue the latest project-scoped session (TUI + prompter)
--work-dir, -w add a workspace directory; the first one is primary and later ones are additional allowed directories
--sandbox [read-only|workspace-write|off] controls landrun (default: off); bare --sandbox enables workspace-write. read-only blocks workspace mutations from both native file tools and child processes while preserving a writable private temp root. Opt-in secure modes use native Landlock + seccomp on Linux and the native Seatbelt SPI on macOS—never sandbox-exec—and deny child network access.

Prompter examples

# Direct prompt
filo --prompt "Review this patch for regressions"

# Stdin only
git diff | filo

# Prompt + stdin context
cat README.md | filo --prompt "Summarize the key setup steps"

# JSON output for automation
filo -p "Generate release notes from these commits" -o json

# Stream JSON events
filo -p "Explain the architecture" -o stream-json --include-partial-messages

# Continue latest project-scoped session
filo --continue -p "Now apply the follow-up refactor"

# Multi-project workspace: ../Lampo is primary, ../filo is additional
filo -w ../Lampo -w ../filo

Lampo

Lampo is a private, on-device AI workspace for Apple Silicon (macOS). Filo and Lampo integrate in both directions:

Direction	What it enables
Filo → Lampo (MCP tools)	Lampo’s local models can call Filo’s coding tools — filesystem, shell, search, patches — outside Lampo’s App Sandbox.
Lampo → Filo (prompt editor)	Edit Filo’s current draft in Lampo’s Prompter UI (`Ctrl+G`), then return the saved text to the TUI.

1) Filo as an MCP tools server for Lampo

Filo’s MCP server (filo-mcp) exposes local coding tools so a host such as Lampo can act on the real filesystem and run shell commands. Server instructions describe the preferred workflow: search before reading, line-sliced reads for large files, search_replace / apply_patch for edits, and persistent shell state across calls.

Coding-oriented tools include (MCP registration set):

Read / search: read_file, list_directory, file_search, grep_search
Write / edit: write_file, search_replace, apply_patch, replace, delete_file, move_file, create_directory
Shell: run_terminal_command
Workspace / orchestration: get_workspace_config, delegate_task
Web (when enabled): web_search, fetch_url
Skills: activate_skill when instruction skills are installed

Paths may be absolute or relative to the active workspace. Use --work-dir / -w to set the primary project and optional additional allowed roots.

Streamable HTTP (Filo daemon)

Best when Filo should keep running independently of Lampo:

# Preferred modern flags (binds 127.0.0.1:8080 by default)
filo --mcp tcp --headless --host 127.0.0.1 --port 8080

# Deprecated but still accepted alias for --mcp tcp:
# filo --daemon --headless --host 127.0.0.1 --port 8080

Health check:

curl http://127.0.0.1:8080/ping

In Lampo, set Transport to Streamable HTTP and the endpoint to:

http://127.0.0.1:8080/mcp

Optional auth for the HTTP MCP endpoint:

export FILO_MCP_BEARER_TOKEN="your-secret"
filo --mcp tcp --headless --host 127.0.0.1 --port 8080

Clients must then send Authorization: Bearer <token>. Prefer localhost unless remote access is required; if you bind 0.0.0.0, combine a bearer token with firewall rules and never expose MCP to the public internet unprotected.

macOS: transparent daemon with launchctl

To keep Filo available for Lampo without a terminal window, install a user LaunchAgent that runs the same MCP TCP command at login and restarts it if it exits.
# 1) Write ~/Library/LaunchAgents/com.filo.mcp.plist (adjust ProgramArguments paths)
# 2) Load it:
launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/com.filo.mcp.plist
# later:
launchctl bootout gui/$(id -u) ~/Library/LaunchAgents/com.filo.mcp.plist
Minimal plist shape: Label com.filo.mcp, RunAtLoad + KeepAlive true, ProgramArguments = absolute path to filo plus --mcp tcp --headless --host 127.0.0.1 --port 8080, optional StandardOutPath / StandardErrorPath under ~/Library/Logs. Point Lampo at http://127.0.0.1:8080/mcp as usual.

2) Lampo as Filo’s prompt editor

On macOS, Filo can open the current draft in Lampo’s Prompter instead of $VISUAL / $EDITOR. Filo implements Lampo’s client-neutral Prompter CLI protocol natively — no adapter script is required.

Interactive setup

Install Lampo from the Mac App Store
In Filo’s TUI, open /settings
Choose Lampo as the prompt editor backend

Usage

Press Ctrl+G to edit the current draft in Lampo
Ctrl+X is also available as an alternate external-editor shortcut
Filo shows opening / editing state, then restores the saved text into the input box when you save in Lampo
Cancelling in Lampo leaves Filo’s draft unchanged

Lampo is opt-in, available only in macOS builds, and does not change the behaviour of other editor backends.

Settings

Most day-to-day options live in the interactive TUI. Open /settings for user/workspace preferences (start mode, approval mode, UI chrome, prompt editor, auto-compaction, tool compression). Model selection is separate via /model.

Preferences persist to ~/.config/filo/settings.json (user) and ./.filo/settings.json (workspace). Workspace values override user values.

Setting	Key	Options
Start Mode	`default_mode`	`BUILD`, `DEBUG`, `RESEARCH`, `EXECUTE`
Approval Mode	`default_approval_mode`	`prompt`, `yolo`
Default Router Policy	`default_router_policy`	configured policy names
Prompt Editor	`prompt_editor`	`system` (uses `$VISUAL` / `$EDITOR`), `lampo` (macOS), or an editor command
Startup Banner	`ui_banner`	`show`, `hide`
Footer	`ui_footer`	`show`, `hide`
Model Badge	`ui_model_info`	`show`, `hide`
Context Meter	`ui_context_usage`	`show`, `hide`
Message Timestamps	`ui_timestamps`	`show`, `hide`
Activity Spinner	`ui_spinner`	`show`, `hide`
Reasoning	`ui_reasoning`	`show`, `hide`
Auto-Compaction	`auto_compact_threshold`	`0` (off), `25000`, `50000`, `100000`, `200000`
Tool Compression	`context_compression`	`off`, `light`, `full`, `ultra`

Useful slash commands:

Command	What it does
`/settings`	Interactive preferences panel (user or workspace scope)
`/dir`	Add an extra directory, or change the primary working directory
`/model`	Switch model / provider / router target
`/auth` · `/login`	Authenticate with a provider
`/logout`	Sign out of a provider OAuth session
`/compression` · `/compress`	Tool output compression (`off`, `light`, `full`, `ultra`)
`/effort`	Model effort (`auto`, `low`, `medium`, `high`, `max`)
`/yolo`	Toggle auto-approval for sensitive tools
`/mcp`	List or manage external MCP tool servers
`/profile`	List, switch, or clear named profiles
`/threads` · `/new`	Switch, rename, create, or close active threads
`/sessions` · `/resume` · `/continue` · `/rename`	Manage and resume saved conversation sessions
`/goal` · `/todo` · `/memory`	Session goal, todos, and durable memory
`/help`	Full command and keyboard shortcut list

Skills without an entry_point also appear as slash commands: /<skill-name> [arguments].

Enable Embedded `llama.cpp`

Build with embedded local GGUF inference:

cmake --preset linux-debug -DFILO_ENABLE_LLAMACPP=ON
cmake --build --preset linux-debug

Then select a llamacpp provider/model from the TUI (/model) after configuring the local backend.

Contributing

See CONTRIBUTING.md for development setup, PR guidelines, and code style.

License

Apache License 2.0. See LICENSE.

Name		Name	Last commit message	Last commit date
.github/workflows		.github/workflows
.vscode		.vscode
cmake		cmake
docs		docs
src		src
tests		tests
.gitignore		.gitignore
CMakeLists.txt		CMakeLists.txt
CMakePresets.json		CMakePresets.json
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE		LICENSE
NOTICE		NOTICE
README.md		README.md

Navigation Menu

Repository files navigation

Filo

Quick Start

Prerequisites

Build And Run

Linux

macOS

Install

What Makes Filo Different

Local-first architecture

Embedded smart routing

Embedded Python runtime

Feature highlights

Runtime Modes

Daemon / transport

Useful CLI flags

Prompter examples

Lampo

1) Filo as an MCP tools server for Lampo

Streamable HTTP (Filo daemon)

2) Lampo as Filo’s prompt editor

Settings

Enable Embedded `llama.cpp`

Contributing

License

About

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

Filo

Quick Start

Prerequisites

Build And Run

Linux

macOS

Install

What Makes Filo Different

Local-first architecture

Embedded smart routing

Embedded Python runtime

Feature highlights

Runtime Modes

Daemon / transport

Useful CLI flags

Prompter examples

Lampo

1) Filo as an MCP tools server for Lampo

Streamable HTTP (Filo daemon)

2) Lampo as Filo’s prompt editor

Settings

Enable Embedded llama.cpp

Contributing

License

About

Resources

Contributing

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

Enable Embedded `llama.cpp`