metis

Competitive Landscape Synthesis

Compiled 2026-05-09. Entry point for the per-stream reports in this folder: coding agents, local-first platforms, routing layers, skills and memory.

Headline

The “TUI/CLI multi-provider coding agent” lane is crowded — OpenCode (157k stars), Claude Code (122k), Gemini CLI (104k), Codex (81k), Cline (62k), Goose (45k), Aider (45k), Roo Code (24k), Crush (24k). Server-client architecture, BYO key, Ollama support, and multi-provider routing are table stakes, not differentiators.

metis’s defensible wedge is the trio of:

1. Bounded, agent-curated memory with hard byte budgets (~2KB MEMORY.md, ~1.5KB USER.md)
2. Lossless canonical message format that round-trips Anthropic’s cache_control, thinking, citations, and tool_use blocks across providers
3. Task-fingerprint pattern learning → cold-start routing recommendations and auto-derived skills

No competitor combines all three. The architecture pitch alone won’t differentiate post-OpenCode — the learning/memory mechanics have to be the headline.

The closest competitors, ranked

Rank	Product	Why it’s close	Where metis differs
1	OpenCode (sst) — 157k★, MIT	Already a server/client + TUI architecture, multi-provider via Models.dev, mid-session swap	No bounded memory, no fingerprint learning, no canonical-IR claim
2	Goose (Block) — 45k★, Apache-2.0	True multi-provider (incl. Ollama), Recipes ≈ skills, lead/worker model split, Tauri desktop	Recipes are author-written, not auto-derived; no bounded memory
3	Roo Code — 24k★, Apache-2.0	Per-mode model assignment + orchestrator delegation = exactly metis’s “agent-decided delegation across tiers”	VS Code-only, no canonical IR, unbounded .roorules
4	Aider — 45k★, Apache-2.0	architect+editor two-tier ≈ metis’s planner/worker; CONVENTIONS.md ≈ skills; /model swap	CLI-only (no server), two tiers not three, no skills standard
5	Letta (memory side) — 23k★, Apache-2.0 + cloud, Series A	Hierarchical bounded memory (core/archival/recall) with agent self-edit tools — closest prior art for metis’s memory model	Not a coding agent, server/cloud-leaning
6	gptme — 4k★, MIT	CLI-first local agent, markdown-logged sessions, multi-provider, MCP — closest spiritual match on the local side	No skills standard, no bounded memory, no server/client split

Verified facts that matter

agentskills.io is real and broadly adopted. Anthropic-originated open standard. Spec at agentskills.io/specification: SKILL.md with YAML frontmatter, optional scripts/ / references/ / assets/ dirs, progressive disclosure (~100 token metadata at startup, full body on activation). ~35+ implementers verified May 2026 including Anthropic, OpenAI Codex, Google Gemini CLI, GitHub Copilot, Cursor, JetBrains Junie, OpenCode, Goose, Roo Code, Letta. Betting on this format is the right call — it’s the de facto interop layer.

No successful agent-skills marketplace exists. GPT Store hit ~3M GPTs but is a graveyard (broken discovery, no quality bar, abandoned wrappers). Cursor Directory is healthy but unmonetized. Anthropic’s anthropics/skills is a small curated repo, not a marketplace. Open lane, but lessons: curation > volume; sandboxing scripts/ trust is the hard part.

Bounded-curated memory is genuinely contrarian. The dominant pattern is “vector-store-everything-and-RAG-it-back” (mem0, Zep, Cognee, most LangChain memory) — they treat eviction as a bug. Only Letta (bounded core blocks with agent self-edit tools) and Anthropic’s memory tool (file-based, no enforced cap) treat eviction as a feature. metis’s hard byte budgets are tighter than both. “Eviction is a feature” is a real wedge.

LiteLLM does NOT solve the canonical-format problem. Open issues from the last few weeks prove it:

• #27512 (2026-05-09) — Anthropic Messages retry drops thinking blocks
• #27469 (2026-05-08) — tool_call.function.arguments lost in OpenAI→Anthropic conversion (regression in v1.83.7)
• #15601 — thinking blocks missing on requests with tool calls
• #26625 / #20418 / #20485 — Bedrock + Vertex cache_control placement broken
• #26937 — citations on Bedrock Converse not supported

Recommendation: keep metis’s typed adapters per provider; treat Anthropic blocks as authoritative internal shape. Use LiteLLM (optionally) only as transport below adapters for users who want a single key endpoint.

Vercel AI SDK is the highest “lunch-eat” risk. Cleanest typed message abstraction in TS (24k★), but TS-only, no built-in agent loop yet, and has its own thinking-block bugs (#13430, #13703 still open Mar–Apr 2026). If they ship a typed Agent with delegation primitives, they compete on the SDK side.

Standalone routing-as-a-service is fading. Not Diamond’s Python SDK was archived Dec 2025; RouteLLM (LMSYS) hasn’t pushed since Aug 2024. The router-without-an-agent thesis is weakening — which means metis’s delegate(tier, task, context) as an in-loop primitive is novel and underexploited.

The genuine gaps metis can own

1. Bounded byte-budgeted memory — every competitor uses unbounded markdown (CLAUDE.md, AGENTS.md, .clinerules) that silently bloats context. Hard budgets + agent curation is novel.
2. Auto-derived skills from past task patterns — Goose Recipes and Claude Skills are author-written; nobody auto-generates skills from successful runs.
3. Three-tier routing (manual → YAML → agent-delegated) with role-attributed cost (planner vs worker) — Aider has two tiers without role attribution; nobody exposes delegation as a first-class agent tool.
4. Lossless canonical IR with all Anthropic-native blocks round-tripping — every router has bug-of-the-week problems on this; only Vercel AI SDK is close on shape (and it’s TS-only with its own bugs).
5. Replays surviving provider changes — depends on (4); no router product offers this because their IR is lossy.
6. Plain-git-remote sync of memory + skills — Cursor uses proprietary cloud, Continue Hub is closed, most tools don’t sync at all.

Where metis risks being a me-too

• “TUI multi-provider agent with markdown rules” is saturated — at least 8 credible options. Yet-another TUI is not a wedge.
• Server/client split stopped being differentiating after OpenCode hit 157k stars.
• Provider catalog (Anthropic + OpenAI + Ollama + OpenRouter) is behind the field on day one — Cline, Roo, Goose, OpenCode all already ship Ollama and 5+ providers.
• MCP support is implicit table stakes by 2026 — not in metis’s pitch (Phase 3), which is a gap to close, not a feature to claim.

Strategic implications

• Lead the public narrative with memory + skill-learning + cost-attributed delegation, not architecture.
• Adopt agentskills.io from day one — verified standard, 35+ implementers, free distribution surface.
• Don’t depend on LiteLLM for canonical IR — write per-vendor adapters and treat Anthropic’s content blocks as the internal shape.
• Pull Ollama and MCP support forward from Phase 3 to Phase 1 if possible — they’re table stakes.
• Watch Letta’s memory architecture closely (core blocks + self-edit tools); it’s the validated prior art.
• Watch Vercel AI SDK — if they ship an Agent abstraction, they’re the most credible “ate metis’s lunch” candidate.

Methodology note

Four parallel research streams ran on 2026-05-09. Each agent verified GitHub stars, last-push dates, license, and (where applicable) live open-issue text via the GitHub API (gh) — those numbers are authoritative as of the date above. SaaS pricing, funding rounds, and feature claims drawn from each project’s public docs and may be stale. Re-verify before citing externally.

This site is open source. Improve this page.