nanocode-cli 0.8.0

A small terminal coding agent written in Python

nanocode

A small terminal coding agent written in Python.

简体中文

Chinese blog

nanocode is pre-1.0 software. Commands, configuration, and tool behavior may change before a stable release.

Features

• Live turn control: Add follow-up input while the agent is still working, without losing the current tool flow.
• File-state brain: Reads and edits build a current, line-numbered view of the files that matter now.
• Stale-edit protection: line:hash anchors reject edits when the target code has drifted.
• Project-aware navigation: Use the symbol index to jump through outlines, references, implementors, call chains, and changed files quickly.
• Recoverable context: Tool output stays bounded in the prompt, while raw tr.N results remain recallable.
• Session recovery: Resume saved work with nanocode --resume, including restored conversation history.
• Cache-aware context: Stable sections stay early and noisy working state stays late to improve prompt-cache reuse.
• Focused working memory: Note separates goal, plan, and known facts from noisy execution logs.
• MCP integration: Connect to remote (HTTP) or local (stdio) Model Context Protocol servers and call their tools.
• Terminal-first workflow: Model selection, history search, confirmations, live command output, appended input, and status all stay in one CLI.

Install

Install with uv:

uv tool install nanocode-cli

Upgrade:

uv tool upgrade nanocode-cli

For local development:

uv sync --extra dev
uv run nanocode

Quick Start

Create a config file:

nanocode --init-config

Edit ~/.nanocode/config.toml, then start:

nanocode

Useful arguments:

• --config <path>: use a TOML config file.
• --init-config: create a default config file.
• --resume [UID]: resume a saved session; without UID, resumes latest.
• --yolo: skip confirmations for mutating tools.
• --mcp <selector>: choose which configured MCP servers to enable.
• --debug: write model I/O debug traces.
• -v, --version: show the version.

During a running turn, type into the +> prompt to add follow-up input for the next model request.

Sessions

nanocode saves recoverable non-empty sessions under [paths] data_dir and prints a restore command on exit:

Resume with: nanocode --resume <session-id>

Resume by id, or use the latest saved session:

nanocode --resume <session-id>
nanocode --resume latest
nanocode --resume last

Restored sessions replay the visible conversation history once; tool summaries are shown, raw tool result bodies are not. /status shows the active session id. Old session files are cleaned up after runtime.session_retention_days days, defaulting to 7; set it to 0 to disable cleanup.

CLI

Commands:

• /help: show commands and tools.
• /status: show runtime status, including the active session id.
• /context [PATH]: show the model's context frame — environment, memory (goal, plan, known facts, check), and file state; PATH shows that file's current in-context lines.
• /skills: list installed skills (load with Skill(name) or reference inline with $name).
• /config: show active config.
• /api [auto|chat|anthropic]: show or set provider API format.
• /debug [on|off]: toggle model I/O debug traces.
• /compact: compact context now.
• /index [force]: sync or rebuild the code symbol index.
• /mcp [tools|login|logout|refresh] ...: manage MCP servers and tools.
• /provider [NAME]: show or set provider.
• /model [MODEL]: show or set model.
• /reason: choose reasoning effort.
• /strict: toggle strict tool-call schemas (OpenAI / DeepSeek only).
• /set KEY VALUE: set supported provider/runtime values for the current session.
• /yolo: toggle tool confirmations.
• /exit, /quit: exit.

Interactive selectors support j/k, arrows, / search, Enter, and Esc. Input supports history, completion, and Ctrl-R history search.

Tools:

• File: Read, LineCount, List, Find, Search.
• Code index: InspectCode.
• Edit: Edit creates or patches file content.
• Shell: Bash, Git.
• Tool results: Recall.
• Working notes: Note.
• Ask the user: Question.
• MCP: MCP.
• Skills: Skill loads a skill's full instructions on demand (offered whenever any skill exists — the built-in nanocode-help means it is normally always available).

Read, Search, and InspectCode return line anchors where useful. Edit uses current line:hash anchors to reject stale edits.

Configuration

Default config location:

~/.nanocode/config.toml

Main fields:

• [provider] active = "name"
• [provider.<name>]: url, key, model, api, prompt_cache_key, available_models, reasoning, chat_reasoning, temperature, max_tokens, strict_tools, timeout
• [paths] data_dir
• [runtime] shell_timeout, max_agent_steps, max_context_tokens, max_parallel_tools, check_updates, update_check_interval_hours, session_retention_days, yolo, debug, tips

api = "auto" chooses between Chat Completions and Anthropic Messages using provider/model profiles. prompt_cache_key = "auto" derives a stable key from provider, model, workspace, and tool schema names.

strict_tools = true (toggle with /strict) constrains tool-call arguments to each tool's JSON schema. It only takes effect on hosts that support it (OpenAI and DeepSeek) and on the Chat Completions path; it is a no-op elsewhere. For DeepSeek, enabling it routes requests to the /beta endpoint. Tools whose schemas can't be represented under strict function calling fall back to non-strict automatically.

Runtime flags such as --yolo, --debug, and --mcp apply to resumed sessions too. Saved sessions do not carry their old runtime config forward.

MCP

nanocode connects to Model Context Protocol servers and exposes their tools through the MCP tool. Configure each server under [mcp.<name>]. A server is either url (remote) or command (local), never both.

Remote server over streamable HTTP:

[mcp.example]
url = "https://example.com/mcp"
bearer_token_env_var = "EXAMPLE_MCP_TOKEN"  # optional; sends Authorization: Bearer
enabled = true

[mcp.oauth_example]
url = "https://example.com/mcp"
auth = "oauth"                              # browser login via /mcp login <server>
enabled = true

Local server over stdio:

[mcp.filesystem]
command = "npx"
args = ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"]
env = { SOME_TOKEN = "value" }              # optional; merged over the inherited environment
enabled = true

The HTTP auth options (auth, bearer_token_env_var, env_http_headers) apply to url servers only. env_http_headers maps a header name to the environment variable holding its value.

Manage servers at runtime:

• /mcp: list configured servers and connection status.
• /mcp tools [server]: list discovered tools.
• /mcp refresh [server]: rediscover servers.
• /mcp login <server> / /mcp logout <server>: OAuth login and logout.

Skills

Skills are reusable instruction packs the agent can pull in on demand. Each skill is a folder with a SKILL.md:

.nanocode/skills/                 # project skills (travel with the repo)
  release-notes/
    SKILL.md
    scripts/
      collect_commits.py
~/.nanocode/skills/               # personal skills (all projects)

SKILL.md has name/description frontmatter and a Markdown body of instructions:

---
name: release-notes
description: Draft a CHANGELOG entry from commits since the last release.
---
Run `python "{skill_dir}/scripts/collect_commits.py" <last-tag>` to gather commits,
then group them by type and write entries in the house style.

• Discovery: .nanocode/skills/ (project) and ~/.nanocode/skills/ (user). On a name clash the project skill wins.
• How the model sees them: only a compact SKILLS index (name + description) sits in context; the full body is loaded on demand when the model calls Skill(name). A repeated load of the same skill collapses to a pointer so the instructions are not re-billed. When no skills are installed, nothing is added to the prompt.
• Reference one inline: type $name in your message (Tab-completes) to nudge the model to use that skill; its instructions are injected for that turn.
• Bundled scripts: {skill_dir} (or ${SKILL_DIR}) in the body expands to the skill's absolute folder path, so the model can run bundled scripts via Bash (subject to normal confirmation unless /yolo).
• Inspect: /skills lists installed skills; the status bar and /status show the count.
• Built-in: a nanocode-help skill ships by default and carries a self-contained manual — authored prose on how to use nanocode, its features, and common problems, plus command/tool/config lists assembled from nanocode's own /help text, tool descriptions, and config keys. So "how do I / what does X / why is Y" questions are answered from the manual without searching the source, and the lists can't drift from the running version. Drop a nanocode-help skill of your own to override it.

Providers

The following providers have been tested with nanocode:

• deepseek: DeepSeek API
• opencode: OpenCode API
• aliyun: Alibaba Cloud Tongyi Qianwen API via Chat Completions
• llama.cpp: Local inference via llama.cpp server

Context Model

Each model request is built manually from explicit messages. Stable context comes first, conversation stays as messages, working memory follows, and the latest file state is appended at the end.

model request
+--------------------------------------------------+
| system                                           |
|   concise agent contract and tool rules          |
+--------------------------------------------------+
| user                                             |
|   Environment                                    |
+--------------------------------------------------+
| user/assistant                                   |
|   conversation, compacted summaries, tools       |
+--------------------------------------------------+
| user                                             |
|   Memory: goal, plan, known, code index, date    |
+--------------------------------------------------+
| user                                             |
|   FILE STATE: latest Read/Edit file view         |
+--------------------------------------------------+

Core rules:

• Environment holds only stable host facts (cwd, os, arch, detected commands) so it stays byte-identical across turns and keeps the conversation prefix cacheable; volatile state like the code index status lives in the late Memory section instead.
• Mid-turn assistant text and appended user input are kept as conversation.
• Earlier conversation is compacted into an explicit summary when the context grows too large.
• FILE STATE is updated by successful Read and Edit tools and shows current listed file ranges, with recent files first.
• Newer file lines overwrite older lines; edit invalidations clear stale ranges.
• File lines are checked against current file stat or line hash before being shown.
• Successful Read and Edit tool messages point to FILE STATE instead of repeating file bodies.
• Other tool outputs are bounded in conversation messages and can be recalled by tr.N.

Safety

nanocode can edit files and run shell commands in the environment where it is started. It does not provide sandbox protection. Run it inside your own sandbox, container, VM, or other isolated environment when needed.