pi-go

A terminal-based coding agent built on Google ADK Go. It connects to multiple LLM providers, runs sandboxed tools, integrates LSP, and ships with a process-based subagent system.

• Multi-provider LLM — Claude (Anthropic), GPT/O-series (OpenAI), Gemini (Google), OpenCode, and Ollama (local or cloud) for models
• Sandboxed tools — read, write, edit, shell, grep, find, tree, and git operations. All tools are restricted to the project directory via os.Root.
• Interactive TUI — Bubble Tea v2 with Markdown rendering (Glamour), slash commands, and theming
• Session persistence — JSONL append-only event logs with branching, compaction, and resume
• Model roles — Named configurations (default, smol, slow, plan, commit) selectable via CLI flags
• Subagents — Process-based multi-agent system with types: explore, plan, designer, reviewer, task, quick_task
• LSP — JSON-RPC client for Go, TypeScript/JS, Python, and Rust, with auto-format and diagnostics hooks
• AI Git tools — Repository overview, file diffs, hunk parsing, and LLM-generated conventional commits (/commit)
• RPC server — Unix socket JSON-RPC 2.0 for IDE/editor integration
• Memory Palace — 4-layer contextual memory with SQLite storage, semantic embeddings (all-MiniLM-L6-v2), temporal knowledge graph, and project/conversation miners
• Extensions — Hooks (shell callbacks), skills (.SKILL.md instructions), and Model Context Protocol (MCP) servers
• Skills audit — Security scanning for hidden Unicode characters, BiDi attacks, and supply-chain threats in skill files (pi audit)

cmd/pi/             Entry point — CLI parsing, output mode selection
internal/
├── agent/          ADK agent setup, retry logic, runner
├── cli/            Cobra CLI flags, output modes (interactive, print, json, rpc)
├── config/         Global and project config (roles, hooks, MCP, themes)
├── audit/          Security scanner for skills (hidden Unicode, supply-chain threats)
├── extension/      Hooks, skills, MCP server integration
├── lsp/            LSP JSON-RPC client, language registry, manager, hooks
├── palace/         Memory Palace — drawers, layers, KG, miners, embedder, search
├── provider/       LLM providers implementing genai model interface
├── rpc/            Unix socket JSON-RPC 2.0 server
├── session/        JSONL persistence, branching, compaction
├── subagent/       Process spawner, orchestrator, concurrency pool
├── tools/          Sandboxed tools (read, write, edit, bash, grep, find, git, lsp)
└── tui/            Bubble Tea v2 UI, slash commands, commit workflow

User input → CLI → Agent → LLM provider → Tool calls → Sandbox → Response → TUI
                     ↕           ↕            ↕
              Session store   Palace       LSP servers
              (JSONL events)  (memory,   (format, diagnostics)
                              KG, search)

See ARCHITECTURE.md for detailed documentation.

curl -fsSL https://raw.githubusercontent.com/dimetron/pi-go/main/scripts/install.sh | bash

This script detects your OS/arch, downloads the latest release binary, and installs it to /usr/local/bin (or ~/.local/bin if needed).

go install github.com/dimetron/pi-go/cmd/pi@latest

Make sure your GOPATH/bin is in your PATH. The binary will be installed as pi.

git clone https://github.com/dimetron/pi-go.git
cd pi-go
go install ./cmd/pi

Download the latest release for your platform from the Releases page.

• Go 1.25+
• At least one LLM provider API key or a running Ollama instance

Set the API key for your provider as an environment variable. The provider is inferred from the model name, so --model is usually the only routing you need.

export ANTHROPIC_API_KEY="sk-ant-..."
export OPENAI_API_KEY="sk-..."
export GEMINI_API_KEY="..."
export OPENCODE_API_KEY="..."
export OLLAMA_API_KEY="..."   # only for Ollama Cloud (:cloud suffix)

A name with no recognized prefix is rejected rather than guessed at — reach for the ollama/ prefix or the :cloud suffix to name an Ollama model explicitly.

make build      # build the pi binary
make test       # run unit tests
make lint       # golangci-lint (vet, staticcheck, errcheck, …)
make e2e        # run E2E integration tests
make clean      # remove binary

# Default interactive mode
pi

# Select a model by prefix
pi --model claude:sonnet
pi --model openai:gpt-4o
pi --model gemini:gemini-2.5-pro
pi --model ollama/gemma4:12b-mlx
pi --model opencode/kimi-k3
pi --model minimax-m3:cloud # automatically detect ollama if :cloud

# Use model roles
pi --smol          # fast, cheap model
pi --slow          # most capable model
pi --plan          # planning-oriented model

# Additional options
pi --continue      # continue last session
pi --session <id>  # resume specific session
pi --system "..." # custom system instructions
pi --url "..."    # custom API endpoint URL

# Non-interactive modes
pi --mode print "explain this codebase"
pi --mode json "list all TODO comments"
pi --mode socket --socket /tmp/pi-go.sock  # JSON-RPC 2.0 over a Unix socket
pi --mode rpc                              # pi-compatible NDJSON over stdio (for pi-acp)

A 4-layer contextual memory system that gives the agent persistent awareness across sessions.

Layers:

CLI commands:

# Setup
pi memory model download         # download all-MiniLM-L6-v2 embedding model
pi memory model status           # check model path and status
pi memory init [dir]             # create palace.db + generate mempalace.yaml

# Ingest
pi memory mine <dir>             # mine source files into drawers
pi memory mine --convos <dir>    # mine conversation files (JSONL/text)

# Query
pi memory status                 # palace overview (drawers, wings, rooms, KG)
pi memory search <query>         # semantic or keyword search
pi memory wake-up                # print L0+L1 context for system prompt
pi memory recent [project]       # recent memory observations

# Knowledge Graph
pi memory kg query <entity>      # query triples involving an entity
pi memory kg add <s> <p> <o>     # add a fact triple
pi memory kg timeline <entity>   # chronological timeline of facts

Configuration via mempalace.yaml in the project root:

wing: my-project
rooms:
  - name: auth
    patterns: ["internal/auth/**"]
    keywords: [jwt, token, session]
  - name: api
    patterns: ["internal/api/**"]
    keywords: [handler, endpoint, route]

When the Palace is enabled, the agent also gains tool access: palace-search, palace-add-drawer, palace-kg-query, palace-kg-add, palace-diary-write, palace-traverse, and more.

# Scan all skill files for hidden Unicode characters
pi audit

# Scan with verbose output (include info-level findings)
pi audit -v

# Output as JSON for CI pipelines
pi audit --format json --output report.json

# Auto-remove dangerous characters (creates .bak backups)
pi audit --strip

# Preview what would be removed
pi audit --strip --dry-run

# Scan a specific file
pi audit --file path/to/SKILL.md

Skills are automatically scanned on load — skills with critical findings (Unicode tags, BiDi overrides, variation selector attacks) are blocked from loading.

Pi reads configuration from ~/.pi-go/config.json (global) and .pi-go/config.json (project-local):

• Model roles — Map role names to specific model strings
• Hooks — Shell commands triggered on tool events (e.g., post-write formatting)
• MCP servers — External tool servers via Model Context Protocol
• Themes — Terminal color schemes via theme config field
• Base URLs — Per-provider endpoints via the baseURLs field

Self-hosted or LAN endpoints can be declared in config instead of exported in every shell:

{
  "roles": {
    "default": { "model": "ollama/gemma-4-e4b:latest", "provider": "ollama" }
  },
  "baseURLs": {
    "ollama": "http://192.168.1.10:11434"
  }
}

Precedence is --url flag, then environment variable, then baseURLs config. The matching env vars are ANTHROPIC_BASE_URL, OPENAI_BASE_URL, GEMINI_BASE_URL, MISTRAL_BASE_URL, XAI_BASE_URL, OPENCODE_BASE_URL, and OLLAMA_HOST. A per-shell or CI override still takes effect. An empty env var does not mask a configured value.

Ollama's per-request options are left at the server's own defaults, except for an output cap. Each knob below is opt-in: unset means the option is not sent at all, so Ollama's default stays in force. An unparseable value is ignored rather than fatal — a typo in an env var should not take down an otherwise healthy session.

These matter for models prone to repetition collapse, where a turn stops making progress and restates the same phrase until it hits a limit. num_predict only bounds how far such a turn runs; it does not stop it degenerating. The penalty window is the knob that targets the cause, and the default window is narrow: Ollama penalises repeats across the last 64 tokens only, while observed degenerate turns cycle on phrases of roughly 25–55 tokens, so a full cycle can fall outside the window the penalty can see.

# Widen the repetition window and penalise repeats harder.
export PI_OLLAMA_REPEAT_LAST_N=512
export PI_OLLAMA_REPEAT_PENALTY=1.2

Both apply to local Ollama and Ollama Cloud — they share one request path. Raising these trades diversity for repetition control, and a value that helps one model can degrade another, so tune per model rather than setting them globally.

For OpenAI-compatible APIs with model names that Pi cannot infer from a prefix, explicitly set the role provider to openai and point OPENAI_BASE_URL at the custom endpoint:

export OPENAI_API_KEY="your-api-key"
export OPENAI_BASE_URL="https://api.example.com/v1"

{
  "roles": {
    "default": {
      "model": "Qwen3.5-397B-A17B-FP8",
      "provider": "openai"
    }
  }
}

Then run Pi normally:

pi

You can also pass the endpoint per invocation:

OPENAI_API_KEY="your-api-key" pi --model Qwen3.5-397B-A17B-FP8 --url https://api.example.com/v1

When --url or OPENAI_BASE_URL is set, unknown model names are treated as custom OpenAI-compatible models. Setting provider: "openai" in config avoids relying on model-prefix detection.

Pi supports the Model Context Protocol. Use it to extend the agent with external tools. Configure servers in ~/.pi-go/config.json:

{
  "mcp": {
    "servers": [
      {
        "name": "tavily-search",
        "url": "https://mcp.tavily.com/mcp/?tavilyApiKey=${TAVILY_API_KEY}"
      },
      {
        "name": "filesystem",
        "command": "npx",
        "args": [
          "-y",
          "@modelcontextprotocol/server-filesystem",
          "/tmp"
        ]
      }
    ]
  }
}

Or in standalone ~/.pi-go/mcp.json (Claude Desktop compatible format):

{
  "mcpServers": {
    "tavily-search": {
      "url": "https://mcp.tavily.com/mcp/?tavilyApiKey=${TAVILY_API_KEY}"
    },
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/tmp"
      ]
    }
  }
}

Supported transports:

• HTTP/Streamable — url field for cloud-based MCP servers
• Stdio — command + args for local subprocess servers

Environment variable substitution: Pi automatically expands ${ENV_VAR} patterns in server URLs using .pi-go/.env

Pi can run as an Agent Client Protocol (ACP) server. Use it from any IDE that supports ACP.

Add pi to Zed's agent_servers in your settings:

{
  "agent_servers": {
    "pi": {
      "type": "custom",
      "command": "pi",
      "args": ["acp-server", "--model", "glm-5.2:cloud"],
      "env": {}
    }
  }
}

Then invoke via Zed's agent panel (⌘⇧A / Ctrl+Shift+A) and select "pi". The agent runs in the current Zed project directory with full access to pi's tools and memory.

See LICENSE for details.