Show HN: Prismag – Per-block model routing for the terminal and any IDE

Developer Rufus SD released Prismag, an open-source CLI tool that routes different parts of a single prompt to different AI models using @@tags, enabling per-block model selection in terminals and IDEs like Cursor and Claude Code. The tool addresses the limitation of current AI coding tools that force a single model per conversation or require manual context switching across multiple chats.

One prompt enters. Each block routes to the right model. Tag any block with @@model and PRISMAG sends it to the model you chose — planning to Opus, implementation to Composer, summaries to a fast model — without switching the IDE picker or juggling chats. prismag @@opus: design the auth flow @@composer: implement the middleware ── @@opus → claude-4.6-opus-high-thinking ─────────────────────────── Use short-lived access tokens with rotating refresh tokens because… ── @@composer → composer-2.5-fast ─────────────────────────────────── // middleware/auth.go func RequireAuth next http.Handler http.Handler { … } routed 2 blocks · chained · 1.8s Today's AI coding tools force a binary choice: - Pick one model for the whole conversation, or - Open multiple chats and split the work by hand. Neither matches how you actually work. Planning wants depth Opus . Implementation wants speed Composer . Review wants a different lens entirely. | Without PRISMAG | With PRISMAG | |---|---| | One model per chat | A model per block, in one prompt | | Switch the picker between tasks | @@opus: … @@composer: … and go | | Manual context copy-paste between chats | Output of block N chains into block N+1 | | Auto-routing by cost/latency OpenRouter | You choose the model per block | | YAML/Python pipelines LangGraph/CrewAI | Chat-native @@ syntax, zero config | Prompt with @@tags ──▶ parser ──▶ orchestrator ──▶ model backends ──▶ sectioned result ▲ └── ContextStore in-memory · or maind - The trigger is , not @@ @ — a bare @ collides with the IDE's mention menu. @@ travels as plain text through every chat surface. - Routing is deterministic and owned by the CLI + registry.yaml . - Blocks run serial + chained by default output N → context N+1 , or --parallel for independent blocks. - Context flows through a pluggable store — in-memory by default, or maind https://github.com/rufus-SD/maind for encrypted, cross-session memory. Go 1.26+ go install github.com/rufus-SD/prismag@latest or clone and build git clone https://github.com/rufus-SD/prismag.git cd prismag && make install 1. Guided onboarding — environment, optional API keys, model discovery, registry prismag setup 2. Wire routing into your editor auto-detects the tool prismag init 3. Route a prompt prismag run "@@opus: plan the cache layer" "@@composer: implement it" Or just run prismag with no args to drop into the interactive prismag session. PRISMAG works in two ways , from the same global config: CLI / REPL — runs in any terminal, on any OS. Executes each block via provider APIs using your keys. Universal, deterministic. In your IDE — prismag init writes a rule that teaches the agent to route @@ blocks through PRISMAG. Where the IDE supports per-task subagents, each block is dispatched to its own subagent + model using your subscription no API keys needed . | Editor | Rule file | Dispatch | |---|---|---| Cursor | .cursor/rules/prismag-routing.mdc + .cursor/agents/ | subagents any model | Claude Code | CLAUDE.md + .claude/agents/ | subagents Claude + API fallback | Windsurf | .windsurf/rules/prismag-routing.md | runs via prismag run | GitHub Copilot | .github/copilot-instructions.md | runs via prismag run | Cline | .clinerules/prismag-routing.md | runs via prismag run | Roo Code | .roo/rules/prismag-routing.md | runs via prismag run | Aider | CONVENTIONS.md | runs via prismag run | generic | .prismag/rules.md | runs via prismag run | prismag connect cursor or: claude, windsurf, copilot, cline, roo, aider, generic Subagent dispatch gives true per-block model switching where the editor exposes it Cursor, Claude Code . Everywhere else, the agent runs prismag run and shows the sectioned output verbatim — same routing, same result. @@<alias : <task context shared with every block goes here, before the first tag @@opus: review the security implications of this auth module @@composer: write the unit tests for AuthService @@fast: summarize the diff in 3 bullets @@alias is case-insensitive and maps to a model via registry.yaml .- Text before the first @@ is shared context for all blocks. Serial + chained by default; --parallel for independent blocks.- Chained runs fail fast; parallel runs tolerate partial failure. aliases: opus: model: claude-opus-4-6 concrete id + offline fallback match: claude-opus-4-6 family resolved against the live model list provider: anthropic agent: opus-planner subagent used when routing in-IDE description: Deep reasoning, architecture, security review composer: model: composer-2.5-fast provider: cursor agent: composer-implementer description: Fast implementation, multi-file edits fast: model: gpt-5.3-codex provider: openai description: Cheap, quick summaries and simple transforms Two optional top-level keys remove friction for everyday use: default: opus4.8 untagged prompts route here, so prismag "do X" needs no @@tag exec: CLI tool-loop defaults — set permissions once, no flags per run enabled: true let blocks act on this machine write files, … shell: true also allow run shell approve: ask ask = confirm each action y/N default · auto = no prompt root: ~/Desktop optional: confine file actions to one tree The same model has a different id in every context — claude-opus-4-8 on the Anthropic API, claude-opus-4-8-thinking-high in Cursor, a local tag in Ollama. Pinning one string breaks the moment a provider renames or bumps a model. So PRISMAG treats an alias as a family and resolves it to a currently-valid id from the live model list for the active context queried with your keys in the CLI, cached 12h; the agent-maintained cache in the IDE . It picks the best match deterministically, self-heals across renames, and falls back to the pinned model when offline. Set match: to make the family explicit; otherwise model doubles as it. Inspect what's available any time with prismag models . | Command | What it does | |---|---| prismag | Interactive prismag session or onboarding on first run | prismag setup | First-time setup: keys, model discovery, starter registry | prismag init tool | Wire routing into this project auto-detects the editor | prismag connect <tool | Write the integration rule + subagents where supported | prismag run "@@..." | Route and execute a tagged prompt untagged → default: alias; --exec / exec: lets blocks act | prismag route "@@..." | Show the delegation plan without executing --json too | prismag list | List @@aliases with availability marks | prismag models | Show models available right now | prismag doctor | Diagnose keys, registry, and environment | prismag sessions | List saved REPL session transcripts | prismag resume id | Reopen a past session with its context | PRISMAG calls provider APIs directly — keys go straight to the vendor, never to a gateway. Keys are read from the environment, a ~/.config/prismag/.env , or stored encrypted in maind https://github.com/rufus-SD/maind when present. ANTHROPIC API KEY only: + OPENAI API KEY: @@opus ✓ ready @@opus ✓ ready @@fast ✗ needs OPENAI API KEY @@fast ✓ ready Inside an IDE that dispatches subagents, blocks route via your subscription — no API keys required. Route any block to a model running on your own machine — no API key, no cloud, $0 per token. Both Ollama https://ollama.com and vLLM https://github.com/vllm-project/vllm expose an OpenAI-compatible API, so PRISMAG talks to them natively streaming included . ollama pull qwen2.5-coder:7b serves on http://localhost:11434 aliases: local: model: qwen2.5-coder:7b provider: ollama or: vllm base url: http://localhost:11434/v1 optional override description: Local model — private, free, offline prismag run "@@local: refactor this function" runs entirely on your box Endpoints default to http://localhost:11434/v1 Ollama and http://localhost:8000/v1 vLLM ; override per-alias with base url or globally with OLLAMA BASE URL / VLLM BASE URL . Mix freely — plan locally, implement in the cloud: @@local: draft then @@opus: review . By default a CLI block returns text — PRISMAG is a router, not an agent. Turn on exec and a block can take real actions through a small, permission-gated tool loop: it asks before every step, so you grant rights action-by-action. Set it once in registry.yaml exec.enabled: true plus a default: alias, and the everyday flow needs no tag and no flags — like an agent that asks first: prismag "create a folder on my desktop named poems" ⚠ allow run shell: mkdir -p ~/Desktop/poems ? y/N y ✓ run shell: mkdir -p ~/Desktop/poems Prefer per-run control instead? Skip the config and pass --exec flags always override config : prismag run --exec "@@opus4.8: create ~/Desktop/poem.txt with a short flower poem" - Tools: write file , read file , and run shell exec.shell: true / --exec-shell . - Every action needs approval; approve: auto or --yes skips the prompt use with care , and a non-interactive shell denies by default. root: confines file actions to one tree. Destructive commands are refused by default — rm -rf / , mkfs , dd of=/dev/… , fork bombs, shutdown , etc. are blocked even if approved , so a careless y or approve: auto can't wreck your machine. Ordinary deletes still work via the normal prompt. Override only with exec.allow destructive: true .- The protocol is provider-agnostic a fenced prismag JSON action , so it works on Anthropic, OpenAI, OpenRouter, and local Ollama/vLLM models alike. CLI-only by design : inside an IDE the agent already has its own tools, so PRISMAG just emits a delegation plan there. In the prismag REPL, toggle it with :exec :exec shell , :exec yes , :exec off . PRISMAG already is the router, so it calls provider REST APIs directly with no self-hosted proxy, DB, or admin UI to trust and patch. That keeps the dependency/supply-chain surface tiny — direct APIs, a single static binary. PRISMAG is a routing protocol any agent can speak — no SDK required. Shell out to prismag route --json to get a deterministic plan which model runs which block , then dispatch with your own model access; or prismag run --api to have PRISMAG execute and return the result. See INTEGRATIONS.md /rufus-SD/prismag/blob/main/INTEGRATIONS.md . maind https://github.com/rufus-SD/maind is the optional memory backend: an encrypted, local-first store the CLI and your IDE agent share. With both wired in, context survives across blocks, sessions, and editors. See CONTRIBUTING.md /rufus-SD/prismag/blob/main/CONTRIBUTING.md . See SECURITY.md /rufus-SD/prismag/blob/main/SECURITY.md for credential handling and vulnerability reporting.