English | 中文

phi

A minimal terminal coding agent harness in Go — a sibling to Pi.

• Sub-agents — spawn isolated jobs and watch the full run unfold in the TUI / job logs, without stuffing every turn into the parent context
• Hashline edits — edit by line + content hash (same idea as oh-my-pi): the model points at anchors instead of rewriting whole files; stale hashes are rejected so over-edits and silent corruption stop here
• Permission gate — Gate / Ask before destructive tools fire; safety is not optional when an agent can touch your tree
• MCP without context death — configure as many MCP servers as you want; their tool schemas never enter the model prompt. The system prompt lists server names only (like the Skills catalog); the agent uses three meta-tools (mcp_list / mcp_inspect / mcp_call) to discover and call on demand. Same Gate / Ask / Hooks path as built-in tools. See MCP
• Any model — OpenAI-compatible or Anthropic, no vendor lock-in

• Quick start
• Footprint
• Configuration
• Interactive mode
• Commands
• Sessions
• Headless mode
• Skills
• Permissions
• Hooks
• MCP
• Tools
• Project layout

Quick start

Install the latest release (macOS / Linux):

curl -fsSL https://raw.githubusercontent.com/pulseaiclub/phi/main/scripts/install.sh | bash

Windows (PowerShell 5.1+):

irm https://raw.githubusercontent.com/pulseaiclub/phi/main/scripts/install.ps1 | iex

First launch needs a model. Open the config editor (creates ~/.phi layout and writes ~/.phi/config.yaml):

phi config

Or set env vars for a one-off run:

export PHI_MODEL=gpt-4o
export PHI_API_KEY=sk-...

Then start the TUI:

phi

Or build from source (Go 1.26.3+, see go.mod):

make build          # produces ./phi
make install        # build and install into $GOBIN

On first start, phi automatically creates ~/.phi/{bin,skills,hooks,session}. Search tools (fd, rg) download into ~/.phi/bin in the background when missing.

The TUI gives the model four core tools — read, write, edit, and bash — plus grep, glob, list, and fetch. The model uses these to fulfill your requests.

Footprint

phi aims to stay cheap to run and cheap to hack on. Numbers below are for a stripped release build (CGO_ENABLED=0, -ldflags="-s -w"), measured on macOS arm64 unless noted.

Configuration

phi reads ~/.phi/config.yaml (standard YAML). Environment variables override it for one-off runs. phi config opens an HTML editor for the same file in your browser.

# ~/.phi/config.yaml
models:
  - name: gpt-4o            # model name; "claude-*" routes to the Anthropic API
    api_key: sk-...         # or set PHI_API_KEY
    base_url: https://api.openai.com/v1   # default; PHI_BASE_URL overrides
    context_window: 128000  # optional
    default: true           # the model used at startup; first entry wins if absent
  - name: claude-sonnet-4-20250514   # extra models; switchable at runtime
    api_key: sk-ant-...
    base_url: https://api.anthropic.com
    context_window: 200000

skill_path: ~/.phi/skills # where SKILL.md files are loaded from

agents:
  enabled: true           # default; set false to disable agent_* sub-agent tools

permissions:
  mode: interactive       # interactive | readonly | autopilot | headless-strict
  bash:
    default: ask          # ask | allow | deny
    allow:
      - "go test ./..."
    deny:
      - "rm -rf *"
  fetch:
    default: allow
    allowed_hosts:
      - "github.com"

Environment overrides:

Provider routing: a base URL containing anthropic or a model name starting with claude uses the Anthropic Messages API; everything else uses the OpenAI-compatible /chat/completions path.

Workspace layout

~/.phi/
├── config.yaml   # global configuration
├── bin/          # downloaded search tools (fd, ripgrep)
├── skills/       # SKILL.md skill directories
├── hooks/        # tool-loop hook scripts (hook.json + run)
├── jobs/         # sub-agent job artifacts (meta, logs, result.md)
└── session/      # persisted sessions, one dir per working directory
    └── <encoded-cwd>/

Interactive mode

phi (or phi tui) starts the TUI: a chat transcript on top, an editor at the bottom, and a footer with the current activity. When a newer release is available, the footer shows a hint like 0.2.0 available · phi update.

Assistant output is rendered as Markdown (CommonMark/GFM): headings, emphasis, strikethrough, links, blockquotes, lists, task checkboxes, and tables are styled with the active theme; fenced code blocks get a frame and per-language syntax highlighting. Structural markers (#, `, *) are stripped.

The editor supports:

• @ — fuzzy file mention picker (type @ and start typing a path)
• / — slash command picker (/sessions, /resume, /clear)
• !command — run a shell command locally and stream its output into the transcript (see Commands)
• Ctrl+K — command palette: settings → model / theme / permissions / agents, skills, hooks

Keyboard shortcuts

Themes: Dark, Darcula, Pink, and Terminal (default), switchable from the palette under settings → theme.

Commands

In the TUI, !command runs locally via bash -c — outside the agent loop. It doesn't count toward agent busy state, and the running command can be cancelled with Esc without touching an in-flight agent turn.

Sessions

Sessions persist automatically per working directory under ~/.phi/session/<encoded-cwd>/ as JSONL trajectories.

• phi sessions list — list session id, mtime, and preview for the current directory
• /sessions in the TUI — same, in-app
• /resume <id> — continue a session (id or unique prefix)
• /clear — start a fresh session (new id, empty transcript)
• phi run --session <id> / phi run --continue-last — resume headlessly

Headless mode

phi run -p "fix the failing test in internal/tools"

Runs one agent loop without a TUI. Human logs go to stderr; with --jsonl, machine-readable events go to stdout, one JSON object per line.

Flags:

Exit codes: 0 success · 1 runtime/LLM error · 2 max rounds reached · 3 config/usage error.

In the interactive TUI, exhausting the tool-round budget prompts Continue / Stop. Headless phi run has no confirmation UI, so it exits with code 2.

In headless mode, permission ask decisions are denied (there is no approval UI), so readonly-style safety applies without extra flags.

Skills

Skills are directories containing a SKILL.md file with YAML frontmatter and a Markdown body. They are loaded from ~/.phi/skills/ (or skill_path / PHI_SKILL_PATH) and injected into the agent's context, letting you give the model reusable procedures:

---
name: My Skill
 description: What this skill does
license: MIT
compatibility: claude, openai
---
Instructions the agent should follow when this skill is relevant.

In the TUI, add skills from the palette (skills → list), then submit the message with the selected skills applied.

Permissions

Tool execution is gated by a permission policy, so the agent can run read-only by default and ask before anything destructive. Configure it under permissions: in ~/.phi/config.yaml.

Modes:

Per-tool rules: bash.default / bash.allow / bash.deny (exact command prefix matching) and fetch.default / fetch.allowed_hosts. Global keys: workspace_only_writes (default true), ask_timeout_sec, and dangerously_allow_all (default false).

In the TUI, an approval dialog replaces the editor with options to approve, deny with feedback, or allow all for the session / for every session. The palette's settings → permissions entry toggles session-wide bypass.

Hooks

Hooks run custom logic around each tool call — before the permission gate and after execution. Use them for organization policy, audit trails, or rewriting tool input, without changing phi's binary or config.yaml.

Each hook is a directory containing a hook.json manifest and an executable:

{
  "name": "guard-bash",
  "event": "pre_tool",
  "match": "bash",
  "run": "./run.sh",
  "fail_closed": true
}

Hooks load from ~/.phi/hooks/ and <cwd>/.phi/hooks/; a project hook with the same name replaces the user hook. In the TUI, list or reload them via Ctrl+K → hooks. In readonly permission mode, only fail_closed hooks run so slow audit hooks don't stall exploration. Full guide: doc/hooks.md.

MCP

Configure 100 MCP servers. Pay ~0 schema tokens until you call one.

Most MCP hosts dump every tools/list schema into the model context before you ask a question — browser stacks alone can burn 50k+ tokens. phi does not.

Instead the agent gets three meta-tools, and the system prompt lists configured server names (no schemas):

Flow: pick a server from the prompt → mcp_list(server=…) → mcp_inspect → mcp_call. Subprocesses start lazily on first use. Calls still go through PreHooks → Gate / Ask → Run → PostHooks.

phi mcp add browsermcp -- npx @browsermcp/mcp@latest
phi mcp doctor
# In the TUI, the model can use configured servers without guessing MCP exists

Config: ~/.phi/mcp.json (project <cwd>/.phi/mcp.json overrides by name). Disable with PHI_MCP=off. Stdio and HTTP in v1.

Full guide: doc/mcp.md.

Sub-agents

Sub-agent tools (agent_spawn, agent_task, …) are on by default. To keep a session lean, disable them in ~/.phi/config.yaml:

agents:
  enabled: false

Or toggle for the current session via the palette: settings → agents. When disabled, those tools are not registered and the model cannot spawn jobs.

Sub-agents themselves use a role (explore default | review | worker):

Default stays explore (read-only). Prefer worker only after the parent has a concrete plan.

Tools

Built-in tools the model can call (see internal/tools/):

Sub-agent transcripts live under ~/.phi/jobs/<id>/ and are not injected into the parent context — only the wait/task summary is.

Fast search tools (fd, ripgrep) are downloaded on first startup into ~/.phi/bin when missing.

See Project layout for the source tree map.

See CONTRIBUTING.md for development setup, code style, and commit conventions.