letscode

A minimal, OpenAI-compatible coding agent for the terminal — written in Python. Point it at any OpenAI-API-compatible endpoint (Ollama, Fireworks, OpenRouter, vLLM, llama.cpp's llama-server, …) and get a streaming agent loop with the four tools that cover 95% of coding sessions — read, write, edit, bash — plus skills, slash commands, and a plugin system you can extend.

Status: v0.6, alpha.

The full docs live in docs/ and build via Zensical — run make docs-serve for a live preview, or make docs for a static build.

Or go to https://letscode.hop3.abilian.com for the online doc.

What it does

• Streaming agent loop with parallel tool execution.
• Four built-in tools: read, write, edit, bash. Restrictable with --tools / --no-tools.
• Skills in the agentskills.io format (SKILL.md + YAML frontmatter). Drop a file in ~/.letscode/skills/<name>/ and the agent picks it up — the model invokes it as a tool, or you invoke it as /skill:<name>.
• Sessions persisted as JSONL under ~/.letscode/sessions/. -c continues the most recent for the current directory; /tree, /fork, /clone branch and duplicate them.
• Type while it's working. Enter stops the run; type then Enter stops it and redirects to the new instruction. Ctrl+C also stops; press it twice to exit.
• Slash commands: /help, /quit, /reset, /model, /tools, /skills, /plugins, /reload, /tree, /fork, /clone, /compact, /steer, /follow-up, /queue, /verbose, /footer, plus /skill:<name>. Plugins can add more.
• TOML config in ~/.letscode/config.toml (user) + .letscode/config.toml (project, overrides). Run letscode --init to scaffold a commented starter. API keys come from env vars only.
• First-run nice: letscode --version, letscode --init, and an actionable message if you forget the API key — never a traceback.
• Plugin system based on pluggy: tools, commands, skills, frontends, plus lifecycle hooks. See docs/plugins/.
• RPC mode (letscode --mode rpc) — NDJSON over stdio, for IDE / host-app integration.
• Rich terminal — Markdown-rendered assistant text, syntax-highlighted read panels, per-turn footer with model / tokens / cost / context-pct (and cache-hit count when the provider reports one).

The deliberate non-goals — native Anthropic / Bedrock / Vertex provider modules, MCP in core, sub-agents in core, plan mode, built-in todos — are kept outside core on purpose; each can be a plugin. Full list in the roadmap.

Install

uv tool install letscode       # or: pipx install letscode

Then letscode --version and letscode --init to scaffold a config.

To hack on it instead:

git clone https://github.com/abilian/letscode
cd letscode
uv sync
uv run letscode --version

Quickstart

Option A — local with Ollama (free)

brew install ollama
ollama serve &
ollama pull qwen2.5-coder:7b

uv run letscode \
  --base-url http://localhost:11434/v1 \
  --api-key ollama \
  --model qwen2.5-coder:7b \
  "create hello.py that prints 'hi from letscode' and run it"

The first call downloads the model (~4.5 GB for the 7B). After that, everything runs locally; any non-empty string works as the API key.

Option B — Fireworks (hosted)

The reference test endpoint is Fireworks running MiniMax M2.5:

export OPENAI_BASE_URL=https://api.fireworks.ai/inference/v1
export OPENAI_API_KEY=$FIREWORKS_API_KEY
export LETSCODE_MODEL=accounts/fireworks/models/minimax-m2p5

uv run letscode "explain what this directory contains"

Any OpenAI-API-compatible endpoint works the same way — OpenRouter, vLLM, llama.cpp's llama-server, deepinfra, together.ai, etc. Set OPENAI_BASE_URL, OPENAI_API_KEY, and LETSCODE_MODEL (or pass --base-url, --api-key, --model).

Modes

letscode                                  # interactive — streaming TUI
letscode "what does main.py do?"          # one-shot, rich rendering, then exit
letscode -p "summarize this directory"    # print mode — plain stdout, pipe-friendly
letscode -c "now do step 2"               # continue most recent session for cwd
letscode --version                        # print the version (no key needed)
letscode --init                           # scaffold ~/.letscode/config.toml (no key needed)
letscode --mode rpc                       # NDJSON over stdio, for IDE/host-app integration

Print-mode exit codes: 0 (ok), 1 (provider error), 130 (Ctrl+C), 2 (usage error).

Tools

The four built-ins, enabled by default:

Tool	What it does
read	Read a file (with optional line range).
write	Create or overwrite a file atomically.
edit	Apply a search-and-replace edit.
bash	Run a shell command with a timeout.

Restrict with --tools=read,bash (comma-separated), or disable all with --no-tools.

Skills

Drop a SKILL.md file in any of these paths:

~/.letscode/skills/<name>/SKILL.md     # user-global
~/.agents/skills/<name>/SKILL.md       # agentskills.io standard
~/.pi/agent/skills/<name>/SKILL.md     # additional path
~/.claude/skills/<name>/SKILL.md       # additional path
.letscode/skills/<name>/SKILL.md       # project-local (walked up from cwd)
.agents/skills/<name>/SKILL.md
.pi/skills/<name>/SKILL.md
.claude/skills/<name>/SKILL.md

Format:

---
name: code-review
description: Use when reviewing code for correctness, style, or security.
---

# Code Review

When invoked:
1. Read the changed files.
2. Look for off-by-one errors, race conditions, missing input validation.
3. Suggest improvements with concrete examples.

Skills written in this format for other agentskills.io-format tools load unchanged. Two ways to invoke:

• The model calls it as a tool. Every skill is auto-wrapped as skill_<normalised-name> (e.g. write-a-prd → skill_write_a_prd). The model invokes via native tool-calling.
• The user types /skill:<name> [message]. Same effect; the slash form keeps the original skill name with hyphens preserved.

Full reference in docs/user-guide/skills.md.

Slash commands

Type at the prompt during an interactive session. Full reference in docs/user-guide/slash-commands.md. The most common:

• /help — list every registered command.
• /skill:<name> [message] — prepend a skill's body to the next user message.
• /tree · /fork [n] · /clone — navigate / branch the session.
• /compact [bias] — collapse older history into a summary; optional text biases the summarisation prompt.
• /reload — re-read TOML config + reload skills mid-session.
• /verbose — toggle full vs. truncated tool-result panels.
• /plugins — list loaded plugins (built-in + entry-point) with version.

Configuration

User-global at ~/.letscode/config.toml, project-local at .letscode/config.toml:

model = "accounts/fireworks/models/minimax-m2p5"
base_url = "https://api.fireworks.ai/inference/v1"
frontend = "basic"
timeout = 300

[memory]                 # plugin-owned section, passed through to letscode-memory
relevance = false

letscode --init writes a commented starter file (it never overwrites an existing one).

Resolution order (lowest → highest priority):

1. Built-in defaults
2. User config (~/.letscode/config.toml)
3. Project config (./.letscode/config.toml)
4. Environment variables (LETSCODE_MODEL, OPENAI_BASE_URL, OPENAI_API_KEY)
5. CLI flags

API keys are deliberately not loaded from the file. Full reference in docs/user-guide/configuration.md.

AGENTS.md and CLAUDE.md files in cwd and its ancestors, plus ~/.letscode/AGENTS.md, are loaded automatically and appended to the system prompt. Use --no-context-files to skip.

Writing a plugin

A plugin is a Python package with at least one @hookimpl declaration and a letscode entry-point. Minimal example:

import pluggy
from pydantic import BaseModel

from letscode.agent.tools import ToolContext, ToolResult, tool
from letscode.llm.types import TextPart

hookimpl = pluggy.HookimplMarker("letscode")


class _GreetParams(BaseModel):
    name: str


@tool(name="greet", description="Say hello to someone.")
async def _greet(params: _GreetParams, ctx: ToolContext) -> ToolResult:
    del ctx
    return ToolResult(content=[TextPart(text=f"Hello, {params.name}!")])


@hookimpl
def letscode_register_tools(registry):
    registry.add(_greet)

[project.entry-points.letscode]
my_plugin = "mypkg.plugin"

pip install -e . and the LLM can now call the greet tool.

For the full surface — every hook, the stable-import contract, the versioning rule, packaging notes, ExecutionEnv overrides, the trust model — see docs/plugins/authoring.md. For the design rationale behind the extension model itself, see docs/architecture/extension-model.md.

A worked reference plugin lives in plugins/letscode-memory/ — cross-session memory via /remember, /memories, and /forget, implemented in a handful of hooks against the same stable surface external plugins target.

Project layout

src/letscode/       # the app: agent loop, CLI, frontends, tools, skills, plugin system
plugins/            # in-repo reference plugins (letscode-memory, letscode-goal, letscode-textual)
docs/               # public documentation (Zensical)
tests/              # a_unit / b_integration / c_e2e (the latter gated by LETSCODE_E2E=1)

A monorepo split is on the roadmap once external plugin contracts are exercised. For now everything ships as a single package.

Development

uv sync                           # install
make check                        # lint + type-check
make test                         # pytest + coverage gate
make                              # check + test
make docs                         # build docs (Zensical, --strict)
make docs-serve                   # live-reload docs preview
make verify-dist                  # build the wheel and drive it from a clean venv
make verify-plugin                # discover letscode-memory from a built install

For the end-to-end smoke suite (real LLM, costs money or runs locally):

LETSCODE_E2E=1 \
LETSCODE_E2E_BASE_URL=https://api.fireworks.ai/inference/v1 \
LETSCODE_E2E_API_KEY=$FIREWORKS_API_KEY \
LETSCODE_E2E_MODEL=accounts/fireworks/models/minimax-m2p5 \
uv run pytest tests/c_e2e/

Documentation

The full documentation lives in docs/ and is built with Zensical. Top-level sections:

• Getting started — install, first run, switching providers.
• User guide — configuration, slash commands, skills, sessions.
• Plugins — authoring guide, reference plugin.
• Architecture — extension model, agent loop.
• About — changelog, roadmap, lessons learned, related projects.

License

MIT.