PatchPilot

PatchPilot is a local-only coding-agent TUI for repository work. It runs the model on infrastructure you control, keeps file and shell actions visible, and gives you explicit control over every risky operation.

PatchPilot 2.0 removes the cloud-provider integrations from the 1.x line and focuses on making local models reliable: native tool calling with a repair fallback, streaming output, context pressure controls, real child agents, a compact terminal shell, and model discovery across common local runtimes.

What it supports

• Ollama on the current machine, LAN, or Tailscale.
• OpenAI-compatible local endpoints from LM Studio/Bionic, MLX, llama.cpp, and vLLM.
• Native function calling when the runtime supports it, with a validated JSON-envelope fallback.
• Nine focused agent tools for reading, searching, editing, shell work, public URL retrieval, todos, and child-agent delegation.
• plan, build, and bypass modes with workspace boundaries and scoped approvals.
• Streaming answers, tool-argument progress, model throughput, context pressure, token totals, and local-vs-hosted cost estimates.
• Persistent JSONL sessions, resumable summaries, context pinning/compaction, prompt history, and session recaps.
• Isolated explore and general child agents with bounded steps and narrow tool allowlists.
• Model discovery, selection, and unloading across supported local runtimes.
• Flow, fullscreen, and legacy terminal layouts with reduced-motion support.

All file and shell tools still run on the machine where PatchPilot is launched. Connecting to a remote Ollama host moves only model inference.

Requirements

• Node.js 22 or newer.
• Git.
• At least one running model runtime:
  • Ollama, or
  • an OpenAI-compatible local server such as LM Studio/Bionic, MLX, llama.cpp, or vLLM.

No cloud-model account or provider API key is required. PATCHPILOT_LOCAL_API_KEY exists only for local endpoints that you configured to require authentication.

Install

npm install -g @jx-grxf/patchpilot
patchpilot --version

Start a model, then launch PatchPilot inside a repository:

ollama pull qwen2.5-coder:7b
cd /path/to/repository
patchpilot

You can also provide the first task directly:

patchpilot "inspect this repository and identify the riskiest untested path"

First launch opens guided setup. Run /onboarding at any time to select a runtime, endpoint, model, default mode, and child-agent preference.

Local runtimes

Runtime	Provider value	Default endpoint
Ollama	ollama	http://127.0.0.1:11434
LM Studio / Bionic	local-openai	http://127.0.0.1:1234/v1
MLX	local-openai	http://127.0.0.1:8080/v1
llama.cpp	local-openai	http://127.0.0.1:8080/v1
vLLM	local-openai	http://127.0.0.1:8000/v1

Discover every reachable runtime and usable model:

patchpilot models
patchpilot models --json

Examples:

patchpilot --provider ollama --model devstral:24b
PATCHPILOT_LOCAL_URL=http://127.0.0.1:1234/v1 \
  patchpilot --provider local-openai --model qwen3-coder-30b

For remote Ollama inference:

patchpilot --provider ollama \
  --ollama-url http://192.168.1.50:11434 \
  --model qwen3-coder:30b

Inside the TUI, /connect scans remembered, LAN, and reachable Tailscale candidates. /connect local returns to localhost. PatchPilot verifies Ollama candidates before listing them.

CLI

patchpilot [task...] [options]
patchpilot init [--workspace <path>]
patchpilot cleanup [cache|sessions|temp|all] [--workspace <path>]
patchpilot doctor [options]
patchpilot models [--all] [--json]
patchpilot sessions [--workspace <path>]
patchpilot resume [session-id] [--workspace <path>]

Important run options:

Option	Meaning
--workspace <path>	Workspace root; defaults to the current directory.
--provider ollama|local-openai	Select the local model protocol.
--model <name>	Select the runtime model.
--ollama-url <url>	Select a local or remote Ollama endpoint.
--steps <count>	Maximum primary-agent steps.
--apply	Start with workspace writes enabled.
--allow-shell	Start with shell execution enabled.
--subagents	Enable isolated child-agent delegation.

--apply --allow-shell starts in bypass mode. Use it only for a trusted workspace; high-risk shell syntax can still require confirmation.

TUI commands

Type / to browse the complete command palette. Common commands are:

Command	Meaning
/mode plan|build|bypass	Switch permission mode. Tab cycles the same modes.
/provider ollama|local-openai	Switch the model protocol.
/models [query|number]	Refresh, search, or select models.
/connect [host|local]	Select an Ollama compute host.
/eject [model|all]	Unload Ollama or LM Studio/Bionic model instances.
/agents on|off	Toggle child-agent delegation.
/config and /set	Inspect or change persisted runtime settings.
/context ... and /compact ...	Inspect, pin, export, clear, or compact session context.
/status and /usage	Show runtime, permission, token, tool, and savings telemetry.
/diff	Show the current Git diff.
/sessions, /resume, /recap, /new	Manage local sessions.
/doctor	Check Node, Git, endpoint, and selected-model readiness.
/update	Check npm/GitHub and offer to install an exact newer release.
/experimental	Configure file analysis, memory, child agents, and shell metacharacters.
/theme	Choose the flow, fullscreen, or legacy interface.

Press Esc to stop a running task. The flow shell keeps transcript output in normal terminal scrollback; the fullscreen and legacy layouts keep bounded internal scroll regions.

Safety model

• plan advertises only read-only tools.
• build exposes editing and shell tools but requests approval before side effects.
• bypass allows configured writes and shell commands without routine per-call approval.
• File operations are constrained to the selected workspace.
• Common secret and credential paths are denied by default.
• Public URL fetching blocks loopback, LAN, link-local, and cloud-metadata targets, including DNS resolutions into those ranges.
• Child agents cannot spawn grandchildren or use shell commands. explore is read-only; general can use bounded file edits.
• Aborting a run propagates to the active model request and child agent before sibling writes continue.

Review the diff and test the result before committing generated changes. See SECURITY.md for vulnerability reporting.

Configuration

Guided setup and /set persist runtime configuration in ~/.patchpilot/.env. The repository's .env.example documents supported variables.

The most important values are:

PATCHPILOT_PROVIDER=ollama
PATCHPILOT_MODEL=qwen2.5-coder:7b
PATCHPILOT_OLLAMA_URL=http://127.0.0.1:11434
PATCHPILOT_LOCAL_URL=http://127.0.0.1:1234/v1
PATCHPILOT_NUM_CTX=32768
PATCHPILOT_NUM_PREDICT=16384
PATCHPILOT_DEFAULT_MODE=build
PATCHPILOT_SUBAGENTS=0

Thinking stays with the selected model and runtime. PatchPilot detects known thinking-capable model families but does not expose a manual reasoning-effort command.

Development

npm ci
npm run dev -- "summarize this repository"
npm run typecheck
npm test
npm run build
npm audit --omit=dev --audit-level=moderate
npm audit signatures
npm pack --dry-run

The project uses strict TypeScript with NodeNext ESM, React/Ink for the TUI, Zod at protocol boundaries, and Vitest for tests. Build output goes to dist/ and is not committed.

Architecture details live in docs/architecture.md. Product invariants live in docs/product-context.md.

Releases

Release notes are in docs/releases. The current release is v2.0.1.

A signed v* tag triggers CI verification, npm publication, package creation, and the GitHub Release workflow.

License

PatchPilot is released under the MIT License.