✦ Podium

A local-model terminal coding agent — Claude Code's powers, 100% on your machine.

Agentic loop · tools · skills · plan mode · subagents · auto-compaction — tuned for small context windows and modest Macs.

---

╭──────────────────────────────────────────────────────────────╮
│ ✦ Podium  ·  local-model coding agent                        │
│ qwen2.5-coder:14b  ·  ~/projects/acme-api                    │
│ type / for commands · /help for the list                     │
╰──────────────────────────────────────────────────────────────╯
› refactor the auth middleware to use async/await
  ⚙ Read(src/middleware/auth.js)
  ⚙ Edit(src/middleware/auth.js)
Done — converted the three callback chains to async/await and kept the
error semantics identical. Want me to run the tests?

⠹ Pondering…
▓▓▓░░░░░░░ 34% · 4.9k/14.4k
──────────────────────────────────────────────────────────────────
› /m
  /model      ← run, download, or delete a model
  /models
 ↑/↓ select · Tab complete · Enter run
──────────────────────────────────────────────────────────────────

Why Podium?

Cloud coding agents are wonderful — until you're offline, on a private codebase, rate-limited, or just don't want your source leaving the building. Podium gives you the same agentic experience with a model running entirely on your Mac.

The hard part of "local" is that the models are smaller and the context windows are tight. Podium is built around that constraint:

• 🧠 Honest hardware fitting — it only ever offers you models that will actually run, with 🟢/🟡/🔴 verdicts computed for your machine. No more OOM surprises.
• 🪶 Tiny footprint by design — a sub-1k-token system prompt, progressively-disclosed skills, capped tool output, and aggressive auto-compaction keep the window lean.
• 🔧 Tool-calling that actually works — many local models emit tool calls as plain text instead of using the native API. Podium's dual-path parser catches both, with bounded auto-repair, so the agent loop doesn't stall.
• ⚡ Fast after the first token — the model is pre-warmed on launch and kept resident, and responses stream out as they're generated.

Install

Available now on npm and via the Haplo Homebrew tap:

# npm
npm install -g podium-cli

# Homebrew
brew install haplollc/tap/podium

Update anytime — it detects how you installed it:

podium update

Requirements

• macOS (Apple Silicon), Node ≥ 20
• A local-model backend (any one):
  • Ollama — recommended · brew install ollama && ollama serve
  • LM Studio — start its local server (localhost:1234)
  • MLX — mlx_lm.server (localhost:8080)

First run

podium

Podium detects your Mac's memory, shows only the models that will run, downloads your pick with a progress bar, and drops you into a REPL with a live context meter.

╭──────────────────────────────────────────────────────────────╮
│ ✦ Podium setup  ·  pick, download, or delete a model         │
│ Machine  Apple M2 · 24 GB RAM  (≈16.8 GB usable for a model) │
│ Backend  Ollama ✓       runs 100% on your machine            │
│ 🟢 runs comfortably      tight   (8 too big for this Mac)    │
│ 🟢 Qwen2.5-Coder 7B       4.7 GB · ✓ installed               │
│ ❯ 🟢 Qwen2.5-Coder 14B      9 GB ★ recommended · ✓ installed │
│ 🟡 gpt-oss 20B             14 GB · ⤓ download                │
│ 🔴 Qwen3-Coder 30B         19 GB · ⤓ download                │
╰──────────────────────────────────────────────────────────────╯

Features

🛠 Tools	Read · Write · Edit · Bash · Grep · Glob · TodoWrite — all output-capped for small contexts
🧩 Skills	Claude Code-compatible SKILL.md with progressive disclosure. Ships with commit, review, explain, test — and reads your ~/.claude/skills too
🤖 Subagents	The Task tool spawns an isolated-context agent that returns one concise report — keeping exploration out of the main window
📋 Plan mode	/plan flips to read-only; the agent investigates and proposes a plan before touching anything
🎚 Permission modes	default · acceptEdits · plan · yolo, with interactive y/n approval prompts
🧠 Memory	Hierarchical PODIUM.md / CLAUDE.md (user → project)
✨ SOUL.md	Give Podium a personality/voice — per-project or global. It also learns durable preferences ("always be concise") and asks before saving them
↩️ Rewind	/rewind jumps back to an earlier point in the conversation and undoes the file changes made since — pick a message, press Enter
🪝 Hooks	SessionStart · UserPromptSubmit · PreToolUse · PreCompact from ~/.podium/settings.json
🌐 Web	WebSearch + WebFetch (website scanning) — and it tells you when you're offline instead of failing
🔌 Multi-backend	Ollama · LM Studio · MLX behind one interface, auto-detected

In-session commands

Type / and a letter for an autocomplete dropdown.

Command	What it does
/setup	Re-run the setup wizard
/model	Pick, download, or delete a model
/models	List installed models
/pull <name>	Download a model
/skills	List available skills
/soul	Show Podium's personality — /soul <preference> to add one, /soul reset to clear learned ones
/plan	Toggle plan mode (read-only)
/context	Show the context meter + token breakdown
/compact	Summarize + shrink the conversation now
/rewind	Step back to an earlier message and undo file changes since
/clear	Reset the conversation
/<skill>	Run a skill (e.g. /commit, /review)

Models

Podium ships a curated catalog spanning every RAM tier — and shows you exactly what fits. A few highlights (full list in models/catalog.json):

Tier	Picks
8 GB	granite4:micro-h · qwen2.5-coder:3b · smollm2:1.7b
16 GB	qwen2.5-coder:7b · granite4:tiny-h · qwen3:8b · phi4-mini
24 GB	qwen2.5-coder:14b ★ · gpt-oss:20b
32 GB	qwen3-coder:30b · glm-4.7-flash · devstral:24b · codestral:22b
64 GB	qwen3-coder-next · gpt-oss:120b · llama3.3:70b

How it works

Podium is a pnpm/TypeScript monorepo of small, focused packages:

packages/
  hardware/   Mac memory detection + model-fit calculator (🟢/🟡/🔴)
  providers/  Ollama · LM Studio · MLX behind one Provider interface
  core/       agentic loop · context manager · compaction · tool-call parser
  tools/      Read/Write/Edit/Bash/Grep/Glob/TodoWrite + Skill/Task/ExitPlanMode
  skills/     SKILL.md parse + discovery + progressive-disclosure registry
  tui/        Ink/React — wizard, REPL, context meter, autocomplete
  cli/        the `podium` binary, config, slash commands, hooks

The context manager tracks a token budget per turn and auto-compacts (retained prefix + summarize-the-tail) before the window fills. The agentic loop prefers native function-calling but falls back to parsing text-emitted tool calls — the trick that makes small local models usable as agents.

Develop

git clone https://github.com/haplollc/podium
cd podium
pnpm install
pnpm -r build
pnpm test                       # 130+ unit tests
node packages/cli/bin/podium.js

# Live tests against a real model (needs Ollama + a pulled model):
PODIUM_LIVE=1 pnpm vitest run packages/cli/test/live.test.ts

License

MIT © Haplo LLC

_{Built for people who want their coding agent to stay on their own machine.}