metis

Coding Agents

Verified 2026-05-09. GitHub stars, licenses, and last-push dates verified via gh API; pricing/feature claims from each project’s public docs.

Competitor table

Product	Surface	Providers / Local	Memory & Skills	Mid-session swap / routing	License & Price	Stars	URL
Claude Code	CLI/TUI	Anthropic only (+ Bedrock/Vertex passthrough). No local.	CLAUDE.md, slash-commands, hooks, Skills (filesystem-loaded MD bundles)	/model swap; no auto-routing	Proprietary, requires Anthropic login; usage-billed	121,976	github.com/anthropics/claude-code
OpenAI Codex CLI	CLI/TUI	OpenAI primarily; some BYO	AGENTS.md	Model flag; no auto-routing	Apache-2.0 wrapper, OpenAI key required	81,340	github.com/openai/codex
Gemini CLI	CLI/TUI	Google Gemini; OpenAI-compatible endpoints	GEMINI.md	Manual flag	Apache-2.0; free tier via Google login	103,520	github.com/google-gemini/gemini-cli
Cline	VS Code ext	Anthropic, OpenAI, Bedrock, Vertex, Ollama, OpenRouter, LM Studio	.clinerules/, Memory Bank pattern (community)	Per-mode model picker; “Plan/Act” two-model split	Apache-2.0; BYO key	61,558	github.com/cline/cline
Roo Code	VS Code ext (Cline fork)	Same as Cline + custom modes	.roorules, Custom Modes, Boomerang/orchestrator delegation	Yes — per-mode model assignment, orchestrator delegates subtasks	Apache-2.0; BYO key	23,959	github.com/RooCodeInc/Roo-Code
Aider	CLI	LiteLLM-backed → ~all providers; Ollama yes	CONVENTIONS.md, repo map, chat history	/model swap; “architect+editor” two-tier (planner + cheaper applier)	Apache-2.0; BYO key	44,575	github.com/Aider-AI/aider
Goose (Block)	CLI + desktop (Tauri)	Anthropic, OpenAI, Google, Ollama, Databricks, Bedrock, OpenRouter	.goosehints, Recipes, MCP extensions	“Lead/worker” model split; manual swap	Apache-2.0; BYO key	44,829	github.com/block/goose
OpenCode (sst)	CLI/TUI + client/server	75+ providers via Models.dev; Ollama/LM Studio	AGENTS.md, agents folder	Yes — explicit per-agent model, server architecture	MIT; BYO key	157,462	github.com/sst/opencode
Crush (Charmbracelet)	TUI	Anthropic, OpenAI, Groq, Ollama, OpenRouter	CRUSH.md	Manual switch	FSL-1.1-MIT; BYO key	24,056	github.com/charmbracelet/crush
Continue.dev	VS Code/JetBrains ext + CLI	Broad incl. Ollama; “Hub” for sharing configs	rules/, blocks, config.yaml	Multi-model (chat/edit/autocomplete roles)	Apache-2.0; free OSS, paid Hub/Enterprise	33,057	github.com/continuedev/continue
Plandex	CLI/TUI + server	OpenAI, Anthropic, OpenRouter; Ollama via OR	Plan branches, context layers	Built-in “model packs” mapping tiers → roles	MIT; OSS + cloud tier (last push Oct 2025 — slowing)	15,344	github.com/plandex-ai/plandex
gptme	CLI/TUI	OpenAI, Anthropic, OpenRouter, local (llama.cpp/Ollama)	Persistent log, tools as Python	Manual	MIT; BYO key	4,294	github.com/gptme/gptme
Open Interpreter	CLI/REPL	Multi via LiteLLM; local OK	Profiles	Manual	AGPL-3.0; BYO key (last push 2025-05-27 — stalled)	63,441	github.com/OpenInterpreter/open-interpreter
Tabby	Self-hosted server + IDE plugins	Local-first GPU inference; cloud optional	Repo indexing	N/A (completion-centric)	Source-available; free self-host + paid tiers	33,491	github.com/TabbyML/tabby
Zed (Zed AI)	Standalone IDE	Anthropic, OpenAI, Ollama, OpenRouter	Slash-commands, rules	Per-thread model	GPL/AGPL mix; freemium hosted models	82,308	github.com/zed-industries/zed
Cursor	Standalone IDE (VSCode fork)	Anthropic, OpenAI, Gemini, xAI; no Ollama	.cursorrules, Memories, AGENTS.md	Yes (Auto-route + manual)	Proprietary; freemium ($20/mo Pro)	n/a (closed)	cursor.com
Windsurf (Codeium)	Standalone IDE	Hosted models; limited BYO	Memories, rules	Cascade auto-routes	Proprietary; freemium (Cognition acquired 2024)	n/a	windsurf.com
GitHub Copilot Workspace	Web + IDE	OpenAI/Anthropic via GH	Repo-bound spec/plan/impl flow	Model picker (added 2024+)	Proprietary; $10–39/mo	n/a	github.com/features/copilot
JetBrains AI Assistant / Junie	JetBrains IDEs	OpenAI, Anthropic, Google, local via LM Studio/Ollama	Project rules	Yes	Proprietary; subscription	n/a	jetbrains.com/ai
Amazon Q Developer	IDE + CLI	Bedrock (Claude/Nova)	Q rules	Limited	Proprietary; freemium	n/a	aws.amazon.com/q/developer
Tabnine	IDE plugins	Hosted + on-prem; local “Protected” model	Personalization on team code	Limited	Proprietary; freemium + Enterprise	n/a	tabnine.com

Marketing claims that don’t survive scrutiny

• Cursor markets “use any model” but does not support Ollama / local models.
• Windsurf is fully cloud after the 2024 Cognition deal, despite “local indexing” marketing.
• Goose and OpenCode are the most legitimately provider-agnostic OSS agents.
• “Multi-provider” in many smaller tools (gptme, Plandex) is just LiteLLM/OpenRouter passthrough.

Closest 3 competitors to metis

1. OpenCode (sst) — 157k stars, MIT, already a server/client architecture with TUI clients, multi-provider via Models.dev, mid-session swap. The most direct architectural twin and the biggest threat to metis’s “local server + thin clients” pitch.
2. Goose (Block) — Tauri desktop + CLI, true multi-provider including Ollama, Recipes (skill-like), lead/worker model split. Block’s funding gives distribution muscle.
3. Aider — CLI not server/client, but its architect+editor two-tier routing, CONVENTIONS.md, and /model swap are the closest philosophical match to metis’s layered routing + skills story.

Roo Code is a near-fourth: per-mode model assignment + orchestrator delegation is exactly metis’s “agent-decided delegation across tiers.”

Genuine gap metis can own

• Bounded, structured memory with explicit byte budgets (~2KB MEMORY.md, ~1.5KB USER.md). Every competitor uses unbounded markdown (CLAUDE.md, AGENTS.md, .clinerules) that silently bloats context. A budgeted memory contract is novel.
• Pattern learning from past tasks → accumulated skills as first-class artifacts. Goose Recipes and Claude Skills are author-written; metis claims auto-derived skills, which nobody ships well.
• Three-tier routing (manual → YAML rules → agent-delegated) is more structured than Aider’s two-tier or Roo’s mode picker.
• Per-turn cost tracking as a core UX surface (some have it as logs; few make it primary).
• Sync via plain git remote for memory/skills — most tools either don’t sync or use a proprietary cloud (Cursor, Continue Hub).

Already crowded / me-too risk

• “TUI multi-provider agent with markdown rules” is saturated: Crush, OpenCode, Aider, gptme, Goose CLI, Codex, Gemini CLI, Claude Code all overlap. Yet-another TUI alone is not a wedge.
• Server/client split is no longer differentiating after OpenCode’s 157k-star traction.
• Provider-agnostic adapters is table stakes; “Anthropic + OpenAI first, Ollama later” is behind the field on day one — Cline/Roo/Goose/OpenCode already ship Ollama.
• MCP support is implicit table stakes by 2026; not mentioned in metis’s pitch — a gap to close, not a differentiator to claim.

Bottom line

metis’s defensible wedge is the memory discipline + auto-skill accretion + cost-aware tiered routing triangle, delivered through a server that any client (TUI today, Tauri/web later) can attach to. The architecture pitch alone won’t differentiate post-OpenCode; the learning/memory mechanics have to be the headline.

This site is open source. Improve this page.