Helmdeck MCP server

helmdeck

Today's helmdeck install ran a full 6-step code-edit loop (clone, read, patch,

test, commit, push) on gpt-oss-120b for $0.07. The same loop on Cursor

or Claude Code direct via Sonnet would have cost $0.30+. Same outcome,

~5× cheaper — and the "expensive" stack isn't even the most expensive option.

Most browser agents require GPT-4o or Claude Sonnet to work reliably.

Helmdeck is built for the other 99% of deployments — **local 7B models,

air-gapped environments, and teams that can't send credentials to a

cloud API.** It wraps every browser, desktop, git, and code action

into a single typed JSON call that even a small model can fill in correctly.

The numbers above are the consequence: when packs absorb the work the

LLM would otherwise burn tokens rediscovering, cheap or local models do

agentic work that frontier-model APIs charge 10× more for.

A self-hosted, containerized platform for AI agents, exposed as Capability Packs — schema-validated, one-shot JSON tools — and native MCP. The defining metric is ≥90% pack success on 7B–30B-class open-weight models, something no frontier-targeting competitor is optimizing for.

📊 Full per-task comparison with reproduction recipe at . These are one maintainer's findings; we welcome community reproductions.

Why this exists

Smart models thrive on bash and a README. Weak models stall on open-ended interfaces. Helmdeck closes that gap by hiding browser sessions, desktop actions, credentials, and multi-step workflows behind single typed REST / MCP calls.

Three audiences specifically:

• Self-hosted AI teams who can't leave their VPC and need MCP-native infra that doesn't phone home.
• The LocalLLaMA / Ollama crowd running 7B–30B models — pack contracts keep small models reliable where open-ended tool surfaces fail.
• Security-sensitive orgs who need agents to log into SaaS apps without the model ever seeing a credential (vault-backed placeholder tokens + MCP-level audit).

Status

v0.22.0 shipped — agents that work on free models, with memory. The release closes ADRs 047–050:

• Pipeline routing + routing memory (ADR 047) — the helmdeck.route meta-pack recommends the best pack/pipeline for an intent (with structured gap warnings when nothing fits), backed by per-caller learned defaults surfaced through the helmdeck://routing-guide and helmdeck://my-defaults MCP resources and a Routing Memory management UI.
• Memory write surface + OpenClaw bridge (ADR 048) — helmdeck.memorystore persists durable user facts (read back via helmdeck://my-memory), an optional embedding sidecar powers OpenClaw's memorysearch, and a QMD corpus bridge exposes helmdeck memory to OpenClaw.
• Intent decomposition (ADR 049) — helmdeck.plan turns a multi-action prompt into an ordered, pipeline-aware step plan plus a rewritten_prompt.
• LLM context manager (ADR 050) — internal/llmcontext compacts catalog-heavy prompts to fit small-model context budgets (tiered per-model budgets, cascading select + lexical rank, optional two-pass filter), surfaced through helmdeck://context-budgets and helmdeck://my-plans.

Shortened. The full README is on GitHub.