Choosing recipes — the catalog, by function

This guide is curation, not mechanics — it groups the tool-providing recipes by what they actually do, flags where two recipes solve the same problem (so you don’t stack redundant or conflicting tools into one profile), and suggests starting combinations. For the stack manifest schema and build/run lifecycle, see the stacks guide.

Availability (as of 2026-07-05)

Not everything below is landed. Before building a stack, know what’s actually buildable:

On main (buildable today): agent-carnet, beads, beads-stealth, caveman, codebase-memory-mcp, context-mode, floating-recipe, greet, gstack, mikes-universal-setup, openbrain-example, ping, repowise, rtk, serena, time. beads-stealth is a sibling of beads — the same bd task tracker in fully invisible mode (zero git footprint); conflicts: [beads], so pick one.
Unmerged feature branches (NOT build/capability-verified): agentmemory, gbrain, headroom, hindsight, hyperpowers, solidspec, superpowers, tokensave. “Exists in a branch” ≠ “works” — treat these as in-development.
Capability-confirmed live (dogfooded in the claude_review-harness stack): beads, caveman, rtk, codebase-memory-mcp, mikes-universal-setup. Caveat: beads needs server-mode Dolt for its dependency graph (see ROADMAP.md).

Code intelligence — pick one pairing, not all three

Recipe	Mechanism	Scope
`codebase-memory-mcp`	tree-sitter → graph (158 languages)	Read-only structural queries: `search_graph`, `trace_path`, `query_graph` (Cypher), `get_architecture`, `detect_changes`. The cross-language map.
`serena`	real LSP, per-language servers	Precise symbol operations: retrieval, rename, refactor, find-references. Explicitly disables its own basic file/search tools (`--context ide`) to avoid overlapping the harness’s. The scalpel.
`tokensave`	tree-sitter → libSQL graph (50+ languages)	Same core mechanism as `codebase-memory-mcp` (near-duplicate there) plus dedicated token-savings accounting (`tokensave gain`/`cost` — per-call token deltas, session/lifetime counters, cost-by-model) and graph-based edit primitives (`str_replace`, `ast_grep_rewrite`, rename preview) that graze `serena`’s territory without a live language server.

Recommended: codebase-memory-mcp + serena (complementary — map + scalpel). Swap in tokensave instead of codebase-memory-mcp only if you specifically want its token-cost telemetry; don’t run all three (codebase-memory-mcp and tokensave overlap directly on mechanism).

Token / output compression

Recipe	Compresses
`caveman`	The agent’s own generated speech (terse “caveman” mode, ~75% fewer output tokens)
`rtk`	Output of explicit dev-command wrapping (`rtk git status`, `rtk read file.rs`, 60–90% smaller)
`tokensave`	Substitutes graph queries for raw file reads/greps, with instrumented savings
`headroom`	On-demand compression of large tool outputs/logs/JSON via MCP tools (`headroom_compress`/`retrieve`/`stats`, 60–95% reduction) — candidate for this group; not yet confirmed as a default pick

These four address different points in the pipeline (agent speech / explicit CLI wrapping / code-graph substitution / on-demand blob compression) — non-overlapping, safe to combine as needed.

Not in this category: context-mode looks like it should be (it’s about “context”), but its actual skill (catalog/recipes/context-mode/skills/context-mode/SKILL.md) is explicitly “a tracer, not a full indexing engine” — it just appends notes to a persist-mounted file to prove cross-launch persistence works. Zero compression mechanism. It belongs in the memory group below.

Lightweight per-project memory — compatible together

Recipe	Storage	Job
`agent-carnet`	`.carnet/` markdown, auto-expiring	Notebook — free-form notes
`beads`	`.beads/` embedded Dolt graph	Task/issue tracker
`context-mode`	`~/.context-mode/notes.md`, persist-mounted	Proof-of-concept persistence tracer, thinnest of the three

Different jobs, no overlap — safe to run together. context-mode is mostly of interest as the persist-spike reference implementation; agent-carnet is the fuller notebook for the same idea.

Shared cross-project memory — pick one, or none

Recipe	Backend	Note
`agentmemory`	Local REST store (harnessed service) + stdio MCP shim	Simplest to stand up yourself
`gbrain`	Long-lived host service, referenced by URL	Needs the `gbrain` service already deployed
`hindsight`	Host docker-compose stack (AlloyDB Omni)	Heaviest; lowest roadmap priority; needs the compose stack already running

All three are “one shared brain across instances/projects” — pick the one whose backend you’re actually willing to operate; don’t run more than one.

Dev methodology / workflow — pick one, and check for task-tracker overlap

Recipe	Delivers
`gstack`	Garry Tan’s ~50-skill suite — browser automation, design, PDF, and the planning/review workflow skills (`plan-ceo-review`, `office-hours`, `qa`, `review`) referenced throughout this repo’s own ROADMAP process
`hyperpowers`	22 skills / 10 commands / 5 subagents — brainstorming → planning → TDD → review → verification, with its own flat-markdown task tracker at `plans/active/<slug>/{plan,context,tasks}.md` (explicitly local-only, deleted on completion)
`superpowers` (see naming note below)	14-skill composable methodology suite (TDD, systematic debugging, brainstorming, planning, subagent-driven dev, code review, git worktrees)
`solidspec`	Spec-driven CLI methodology: spec → plan → tasks → implement → ship

hyperpowers ships its own task-tracking — it does not use or install beads. They’re two independent answers to “where does the agent track tasks” (flat markdown files vs. a graph DB). Running hyperpowers (or any methodology suite with its own task-doc convention) alongside beads means two uncoordinated tracking systems, not a complementary pairing — pick the methodology suite’s own tracker, or beads, not both, unless you’ve deliberately decided the agent should use one for X and the other for Y.

Naming: the recipe is now superpowers (lowercase, plural) on the feat/recipe-Superpowers branch — the rename off the original capitalized Superpowers landed there (commit rename(recipe): superpower -> superpowers), but the branch itself is still unmerged (see Availability above).

harnessed self-test / example stacks

Not workflow tools — these exist to exercise harnessed’s own composition mechanics (skill-only / stdio-MCP / service-sidecar / mount-widening / dual-harness patterns) and are why catalog/stacks/claude_gstack_ping_time_greet and friends exist.

Recipe	Proves
`greet`	Skill-only recipe (no MCP)
`time`	stdio MCP server + skill
`ping`	Network-native MCP via a service sidecar
`floating-recipe`	Mount-widening (`--mount-folder`)
`openbrain-example`	URL-based remote-MCP template, `url_env` wiring
`gstack`	Multi-recipe composition at scale (~50 skills) — also a real methodology suite, see above

Suggested starting stacks

# Everyday coding
name: claude_dev-core
harness: claude
recipes: [context-mode, codebase-memory-mcp, serena, agent-carnet, beads, rtk]

# + one methodology suite (pick one)
name: claude_dev-methodology
harness: claude
recipes: [context-mode, codebase-memory-mcp, serena, agent-carnet, beads, rtk, gstack]
# — or (once merged to main) swap `gstack` for the planned `hyperpowers` / `superpowers` /
#   `solidspec`; drop `beads` if the methodology suite you pick ships its own task tracker.

# Heavier compression focus
name: claude_dev-compressed
harness: claude
recipes: [context-mode, codebase-memory-mcp, serena, caveman, rtk]