Go terminal harness for coding agents and MCP

English | 中文

A lean, high-performance terminal coding agent harness in Go — a sibling to Pi.

Docs: pulseaiclub.github.io

• Fast and small — ~12 MB release binary, ~21 MB idle RSS, ~40 ms to first frame; no Node / Electron / Python runtime
• 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 whole-file @file path#TAG plus line LINE#HASH anchors (same idea as oh-my-pi): the model points at anchors instead of rewriting whole files; stale tags/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
• Extensions (Go or Rust) — native binaries speak the PXB binary protocol over stdin/stdout; official author SDKs for Go (ext/go) and Rust (ext/rust): LLM tools, slash commands, event intercepts, confirm dialogs — no reflection; JSON at the SDK edges via serde_json. See Extensions
• In-TUI diff review — /diff opens a full-screen git review (working tree / staged / HEAD): syntax-highlighted hunks, line notes, then a sends notes to the agent. See Diff review
• Any model — OpenAI-compatible or Anthropic, no vendor lock-in

• Docs
• Quick start
• Footprint
• Configuration
• Interactive mode
• Diff review
• Commands
• Sessions
• Headless mode
• Skills
• Permissions
• Extensions
• 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, find, and ls. The model uses these to fulfill your requests. External HTTP fetch is available via MCP when configured.

Footprint

Lean is not enough — phi is built to feel instant and stay cheap under load. Numbers below are for a stripped release build (CGO_ENABLED=0, -ldflags="-s -w"), measured on macOS arm64 unless noted.

Metric	phi
Release binary	~12 MB
Idle RSS (1 session)	~21 MB
10 idle sessions (total RSS)	~196 MB (~20 MB each)
Time to first frame	~40 ms (27–65 ms)
Cold go build (empty GOCACHE)	~5.5 s
Warm rebuild	~0.7 s
Go source (excl. tests)	~22k LOC / 107 files
Go packages	32
Direct module deps	6 (15 modules total)
Linked runtimes	system libs only (no Node / Electron / Python)

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
  models:                 # optional per-role defaults; omit → inherit parent model
    explore: cheap-model
    review: strong-model
    worker: coding-model

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

Recommended model: DeepSeek Flash

phi + DeepSeek Flash — the best pairing: grounded, low hallucination, cache hit rates near 100%.

Measured data:

39 LLM rounds, same session — prompt 16k→40k, hit rate 95–100% (avg 98.7%).

Round	Prompt tokens	Cached tokens	Cache hit
1	16,176	15,872	98.1%
10	20,163	20,096	99.7%
20	27,604	26,624	96.4%
30	35,245	35,072	99.5%
39	39,794	39,552	99.4%

xychart-beta
    title "Cache hit % (39 rounds)"
    x-axis [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39]
    y-axis "Hit %" 95 --> 100
    line [98.1, 98.8, 97.5, 98.1, 98.1, 98.8, 99.1, 99.9, 98.7, 99.7, 99.5, 99.5, 96.3, 99.4, 97.0, 99.7, 99.9, 99.3, 98.3, 96.4, 98.9, 98.7, 97.2, 99.8, 95.0, 99.7, 99.2, 98.9, 99.3, 99.5, 99.5, 99.4, 99.2, 98.6, 99.4, 99.6, 100.0, 97.8, 99.4]

Environment overrides:

Variable	Overrides
PHI_API_KEY	models[].api_key (default model)
PHI_MODEL	models[].name (default model)
PHI_BASE_URL	models[].base_url (default model)
PHI_SKILL_PATH	skill_path

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
├── extensions/   # PXB binaries + phi.yaml
├── 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 muted language caption 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, /diff)
• ? — shortcut help picker (lists /, !, @, and key bindings; Esc closes)
• !command — run a shell command locally and stream its output into the transcript (see Commands)
• Ctrl+K — command palette: settings → model / theme / permissions / agents (incl. per-role models), skills, hooks

Keyboard shortcuts

Key	Action
Ctrl+C	Quit phi
Esc	Cancel the running agent / close pickers
Ctrl+K	Toggle the command palette
Ctrl+A	Jump to the start of the line
Ctrl+E	Jump to the end of the line
Ctrl+U	Clear the composer input, images, and skills
Ctrl+Shift+C	Copy the selected transcript text

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

Diff review

/diff is a full-screen git review inside the TUI — read the change, leave line notes, then hand them to the agent without leaving the terminal.

Command	What opens
/diff	Working tree (git diff)
/diff staged	Staged changes
/diff HEAD	Last commit (git show)

Slash-picker Enter inserts /diff plus a trailing space into the composer (same pattern as /resume); submit to open. Inside the overlay: j/k move, s side-by-side, i add/edit a note, x delete, a send notes to the agent, ? help, q / Esc close. Notes persist under .phi/review.json.

Commands

Command	Description
phi / phi tui	Start the interactive TUI
phi run -p "…"	Run one agent loop headlessly (see below)
phi update	Download and install the latest GitHub release
phi update --check	Query the latest release without installing
phi sessions list	List persisted sessions for this directory
/sessions	List sessions for this directory (TUI)
/resume <id>	Resume a session by id or unique prefix (TUI)
/clear	Start a fresh empty session (TUI)
/diff	Full-screen git review — see Diff review
!command	Run a shell command locally, stream output into the transcript; Esc cancels it

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:

Flag	Description
-p, --prompt STRING	Prompt to run (required)
--jsonl	Emit JSONL events to stdout
--yolo	Skip all permission checks for this run (benchmarks / CI only)
--max-rounds N	Cap tool rounds (default 64)
--timeout DURATION	Limit the agent run wall-clock time (e.g. 10m; disabled by default)
--session ID	Resume a persisted session by id or unique prefix
--continue-last	Resume the newest persisted session for this directory
--session-dir DIR	Override the session storage directory
--tools LIST	Enable only these comma-separated built-in tools

--tools accepts built-in names such as read,ls,grep. MCP and agent tools still append when configured; the flag only scopes the built-in toolset.

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. For benchmarks that need arbitrary shell (pytest, npm test, …), pass --yolo to skip the permission gate for that run only.

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:

Mode	Behavior
interactive	Default. ask decisions prompt in the TUI.
readonly	Deny writes / bash; read tools still work.
autopilot	Fold ask → allow, run unattended.
headless-strict	Fold ask → deny (used by phi run).

Per-tool rules: bash.default / bash.allow / bash.deny (exact command prefix matching). 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.

Extensions

Extensions are native binaries speaking the PXB binary protocol over stdin/stdout (author SDKs: Go github.com/pulseaiclub/phi/ext/go/phi and Rust ext/rust, phi-ext). They subscribe to tool/session events, register LLM tools, and add slash commands.

go get github.com/pulseaiclub/phi/ext/go@v0.21.0

package main

import (
	"github.com/pulseaiclub/phi/ext/go"
	"github.com/pulseaiclub/phi/ext/go/phi"
)

func main() {
	m := phi.New("hello", "0.1.0")
	m.OnToolCall(func(ev ext.ToolCallEvent) *ext.ToolCallResult {
		// return &ext.ToolCallResult{Block: true, Reason: "..."}
		return nil
	})
	_ = m.Run()
}

Install under ~/.phi/extensions/<name>/ with a phi.yaml pointing at the binary. In the TUI: Ctrl+K → extensions. Disable with PHI_EXTENSIONS=off. Full guide: doc/extensions.md.

Codec throughput on Apple Silicon (release, single-threaded):

Implementation	Hello encode+decode	Frame write+read (in-memory)	Allocs
Rust PXB (phi-ext)	~0.12 µs	~0.06 µs	—
Go PXB (ext/go/pxb)	~0.11 µs	~0.05 µs	3 / op
Go JSON lines	~1.2 µs	—	15 / op

The ~10× gap over JSON lines is the protocol (fixed header + tagged fields), not the language — Rust and Go are within noise on the same codec work. Real extension latency is dominated by process spawn and pipe RTT anyway. Re-probe: cargo run --release --example bench in ext/rust.

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):

Tool	Role
mcp_list	List tool names on one server (compact text)
mcp_inspect	Fetch a slim parameter summary for one tool
mcp_call	Run server + tool + args

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_wait, …) 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.

Per-role model defaults (optional) under agents.models pick which configured model name each role uses when spawned. Omitted roles inherit the parent session model. Switch for the current session only via settings → agents → models → explore|review|worker (same session-only semantics as settings → model; does not write config.yaml).

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

Role	Tools	Use for
explore	no write/edit; bash except hard denies	Multi-hop recon / map structure
review	same as explore	Diffs / checks; report only
worker	full tools except nesting; bash except hard denies	Scoped, self-contained edits

Default stays explore (no edits). Prefer worker when the task is to implement a scoped change in an isolated context.

Tools

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

Tool	Purpose
bash	Run a shell command in the working directory
read	Read a file
write	Write a file (gated by permissions)
edit	Targeted edit of a file
grep	Regex search across files
find	File patterns (fd)
ls	Directory listing
agent_spawn	Start an isolated sub-agent job (async)
agent_wait	Wait for a job; returns short summary only
agent_list	List jobs
agent_cancel	Cancel a running job

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.