attune-ai MCP Server

Attune AI

Spec-driven development for Claude Code — turn requirements into reliable software.

🌐 Docs & guides: attune-ai.dev

---

Attune AI gives Claude Code persistent memory. Your agent stops starting from zero: a stash → recall → promote loop carries decisions, bugs, and hard-won lessons from one session into the next, and a retrievable lessons corpus surfaces the right lesson at the exact moment a prompt needs it — local-first, working from a plain pip install attune-ai. The economics are measured, not promised: recall loads a few hundred exactly-relevant tokens instead of your whole corpus — 67× fewer tokens on our own 800+ lesson store, retrieved at P@3 96% on a frozen trap-moment benchmark (details in the memory suite below).

Around that memory core, the same package also ships a spec-driven, multi-agent toolkit: 20 workflows and 47 MCP tools dispatching 2–6 domain-specific subagents behind Socratic quality gates, RAG grounding with a citation-per-claim contract (mean per-claim faithfulness CI-gated at ≥ 0.97; 0.98 currently measured, N=20 runs on the 40-query golden set), and generation fact-checking — one install, one MCP server.

We run our own knowledge base on it: the docs, help templates, and 800+ engineering lessons at attune-ai.dev are authored, grounded, and maintained entirely by Attune's own stack.

Managing and creating help-content, docs, or knowledge-bases? That's attune-gui — a dedicated Living Docs dashboard wrapping attune-rag, attune-help, and attune-author in a single UI. attune-ai is the developer workflow hub; attune-gui is the docs hub.

---

New in 10.0.0 — one memory architecture, no legacy layer

Major version, one breaking change: the legacy memory-graph API (attune.memory.MemoryGraph and its node/edge types) is removed. Curated memory has been plain .md files served from Redis since 9.6.0 — the graph was the layer nothing living called, confirmed by a usage audit before deletion (receipts in docs/specs/archive/memorygraph-value-gate/). If you never imported MemoryGraph — telemetry says that's everyone — nothing changes: same install, same memory loop, same measured economics below. Accessing a removed name raises an error naming the successor, and there is no data migration because the graph store was already retired.

The memory suite — out of the box, measured

Your agent stops starting from zero — with a plain pip install attune-ai. The memory suite matured across recent releases (curated promotion in 9.5, files-canonical unification in 9.6, Redis / Agent Memory Server client as a core dependency in 9.7). The loop:

• Stash on stop — a Stop hook extracts decisions, bugs, and references from the session (local LLM when available, heuristic fallback) and writes them to the memory store: a local file by default, Redis Agent Memory Server when one is reachable.
• Recall at the door — a SessionStart hook surfaces the most recent findings for your project, and warns when the memory backend is unreachable instead of degrading silently.
• Promote what endures — a reviewed stash→curated path: the agent drafts a node per candidate, you verdict each one (the 30-day test), and promotion lands a git-tracked .md file in the curated corpus. Files are the store; Redis serves them — recall pulls a few hundred exactly-relevant tokens on demand instead of re-reading whole files into context.
• Lessons at the trap moment — a UserPromptSubmit hook retrieves your project's engineering lessons (from .claude/lessons.md or CLAUDE.md) when a prompt hits a known trap; a PreToolUse hook surfaces curated rules at the exact tool call they govern.
• On demand — /recall <topic> searches both stores with results labeled by source; /remember captures and manages facts explicitly; a full MCP tool surface (memory, personal memory, Redis/AMS) gives agents the same access.

Memory is local-first — nothing leaves your machine, and without a Redis server everything degrades to the file backend with clear guidance. Your memory, your corpus: we dogfood the loop on our own 800+ engineering lessons, retrieved via attune-rag at P@3 96% (100% on the high-severity subset) on a frozen trap-moment benchmark.

The economics, measured (2026-07-05 snapshot; corpus has grown since — the ratios below only improve as it does). Durable memory was 303,205 tokens across 752 docs at measurement time; a session recalls only the relevant slice:

The lessons injection stays budget-capped no matter how large the corpus grows, and recall is a single warm Redis call — so both wins widen as your memory does. Numbers from benchmarks/memory_savings.py on our dogfood store (cl100k_base).

---

Ecosystem

attune-rag and attune-verify both ship as core dependencies of attune-ai — retrieval grounding and generation fact-checking are on by default, no extra needed. attune-help is standalone — not pulled in by a standard attune-ai install, but available as an optional corpus for attune-rag via pip install 'attune-rag[attune-help]'.

---

How It Works

1. Skills trigger automatically

Say what you need in Claude Code and the right skill activates:

"review my code"        → code-quality skill
"scan for vulns"        → security-audit skill
"generate tests"        → smart-test skill
"plan this feature"     → planning skill

No command to remember. Claude reads your intent and picks the skill. Each skill runs a specialist multi-agent team, not a single prompt.

2. Multi-agent teams, not single prompts

Every workflow dispatches 2–6 subagents in parallel. Each reads your code with Read, Glob, and Grep. An orchestrator synthesizes their findings into a unified result:

security-audit → vuln-scanner + secret-detector + auth-reviewer + remediation-planner
code-review    → security + quality + perf + architect
test-gen       → identifier + designer + writer

Subagents are assigned models by task complexity — Opus for deep reasoning, Sonnet for analysis, Haiku for fast scanning — keeping cost proportional to value.

A head start on /agents, too. Beyond the workflow teams above, attune-ai ships a curated set of Claude Code subagents — security-reviewer, spec-author, refactor-planner, release-prep-auditor, and more — that appear in your /agents list the moment the plugin installs. Claude Code gives you the mechanism to build subagents; attune-ai gives you a running start — use them as-is, or fork one as the scaffold for your own.

3. Socratic before execution

Workflows ask questions before executing, not after. The spec workflow brainstorms, then plans, then executes. planning clarifies scope before writing a line of code. This eliminates the most common failure mode: confidently solving the wrong problem.

Dynamic communication — the agent adapts how it talks to you. The headline of this release: instead of a fixed wall of prose, Attune now dynamically shapes each exchange to fit the moment — rendering an interactive form in response to your prompt whenever a structured turn communicates better than text. A multi-part question becomes one form you answer with a click; a recommendation arrives as weighable cards; a disagreement is shown side-by-side so you can overrule it in one tap. This is a deliberate effort to improve human/AI communication — making the back-and-forth faster, clearer, and less ambiguous. Three of the four constructs fire at a fork — a point where the conversation can't move forward without your choice; the fourth (progress) is a status report, not a fork. The agent picks the right construct for the moment:

• intake (fork) — gathers several independent decisions as a single (multi-select-capable) form, instead of N back-and-forth questions.
• decision (fork) — offers a recommended option with a rationale and per-option tradeoffs, rendered as cards (consumed by the /spec approval gate).
• pushback (fork) — when the agent disagrees with your stated approach, it shows "your approach" beside "I'd suggest instead" with a "why", and you overrule or switch with one pick (/spec plan review).
• progress (report) — a done / in-progress / blocked status board whose blocked items are a picker for what to fix next (/spec execute).

All constructs share one declarative form model and validator, render richly on widget-capable surfaces (e.g. claude.ai / Cowork) and degrade gracefully to a recommendation-first menu elsewhere. The terse reply vocab (y / go / 1) answers any of them.

4. RAG-grounded generation

attune-rag (core dep) grounds LLM generation in retrieved corpus passages and enforces citation-per-claim, delivering 0.98 mean per-claim faithfulness on the current CI-gated benchmark (40 queries, N=20 runs, floor ≥0.97) — the large majority of generated claims are grounded in their cited passages. The citation-per-claim design itself was chosen via an A/B comparison (2026-04-19): the conservative per-query bucket rate (a single ungrounded claim disqualifies the whole response) dropped from 46.7% without the contract to 6.7% with it. Retrieved passages are wrapped in sentinel tags to prevent prompt injection. The Claude provider automatically caches the stable RAG context prefix, eliminating repeated token costs across calls.

5. Memory that compounds across sessions

Most AI coding sessions start from zero. Attune ships a cross-session memory loop — stash on stop, recall at the door, reviewed promotion into a git-tracked curated corpus, and lessons retrieved at the exact trap moment they guard against. Covered in depth in "The memory suite" section at the top of this README; the hook-by-hook mechanics and tunables live in the "Session continuity & cross-session memory" section below.

---

Dynamic forms — structured human/AI communication

Attune improves how you and the AI communicate by dynamically using interactive forms (shipped in 9.3.0). Instead of a fixed wall of prose, Attune renders the right form whenever a structured turn communicates better than text — a multi-part question becomes one form you answer with a click, a recommendation arrives as weighable cards, and a disagreement is shown side-by-side so you can overrule it in one tap. Three constructs fire at a fork — a point where the conversation needs your choice to continue; the fourth (progress) is a status report, not a fork:

• intake (fork) — gather several independent decisions in one form
• decision (fork) — a recommended option with rationale + per-option tradeoffs (/spec approval gate)
• pushback (fork) — agent dissent shown as "your approach" vs "I'd suggest instead"; overrule or switch in one pick (/spec plan review)
• progress (report) — a done / in-progress / blocked board whose blocked items are a fix-next picker (/spec execute)

All constructs share one declarative form model and validator, render richly on widget-capable surfaces (e.g. claude.ai / Cowork) and degrade gracefully to a recommendation-first menu elsewhere. The terse reply vocab (y / go / 1) answers any of them.

---

Get Started in 60 Seconds

Plugin (works standalone)

claude plugin marketplace add Smart-AI-Memory/attune-ai
claude plugin install attune-ai@attune-ai

Then say "what can attune do?" in Claude Code.

Add Python Package (unlocks CLI + MCP)

pip install attune-ai
attune            # shows your next steps

Then check your setup with attune validate and run your first workflow: attune workflow run code-review --path src/.

Setup fight you? Tell me where — I'm actively fixing this.

The core install includes the CLI, all workflows, and the MCP server. See Installation Options for per-surface extras (API-mode agents, ops dashboard, Redis memory).

What Each Layer Adds

Note: Skills use your Claude subscription at no extra cost. CLI and MCP tools make direct Anthropic API calls — API key required. See API Mode.

---

Cheat Sheet

---

Workflows

---

MCP Tools

47 tools organized into 6 categories:

Workflow (22)

security_audit code_review bug_predict discovery_sweep performance_audit refactor_plan simplify_code deep_review test_generation test_audit test_gen_parallel doc_gen doc_audit doc_orchestrator release_notes health_check dependency_check secure_release research_synthesis analyze_batch analyze_image rag_knowledge_query

Help (5)

help_lookup help_init help_status help_update help_maintain

Memory (4)

memory_store memory_retrieve memory_search memory_forget

Personal Memory (4)

personal_memory_capture personal_memory_recall personal_memory_topics personal_memory_forget

Utility (8)

auth_status auth_recommend telemetry_stats context_get context_set attune_get_level attune_set_level list_capabilities

Elicitation (4)

elicitation_ask elicitation_render_form elicitation_collect_response elicitation_render_widget

---

Accuracy & Faithfulness

RAG grounding — 0.996 per-claim faithfulness (over 99%)

Measured on a 15-query golden set with retrieval held constant. The per-claim faithfulness score (how much of what the model says is grounded in cited passages) is the headline metric. The conservative per-query bucket rate (a single ungrounded claim disqualifies the whole response) is shown alongside for completeness — they measure related-but-different things, and the per-claim number is the right "how trustworthy is each statement" answer:

The gain comes from the prompting contract (citation-per-claim), not from retrieval. Full methodology:

• docs/rag/faithfulness-decision-2026-04-19.md
• docs/rag/ab-report-2026-04-19.json

Help resolver — 48/48 benchmark queries pass at P@1

• tests/unit/help/fixtures/golden_queries.yaml

---

Why Attune?

---

Installation Options

pip install attune-ai works out of the box — the CLI, all workflows, the MCP server, RAG, cross-session memory (the Redis / Agent Memory Server client is a core dependency as of 9.7.0), and the Agent SDK. Memory features activate when a Redis Stack server is reachable and degrade with guidance when not. Add extras only for the surfaces you use:

([redis] remains as an empty backward-compat alias.) Extras combine — for example pip install 'attune-ai[developer,ops]'. Keep the quotes: zsh and bash treat square brackets as glob characters.

Contributing? Clone and install the dev toolchain instead:

git clone https://github.com/Smart-AI-Memory/attune-ai.git
cd attune-ai && pip install -e '.[dev]'

The [rag] extra is a no-op alias kept for backward compatibility — attune-rag is now a core dependency included in every install.

---

Platform Support

Notes for native Windows:

• Claude Code supports native Windows (10 1809+). Installing Git for Windows enables the Bash tool; without it, the PowerShell tool is used (opt-in via CLAUDE_CODE_USE_POWERSHELL_TOOL=1).
• Under PowerShell, the security-validation hook applies a strict command allowlist and fails closed: commands it does not recognize are blocked rather than silently passed. Use Git Bash or WSL2 for the full experience.
• OS-level sandboxing is available on macOS/Linux/WSL2 only, not native Windows.

Redis on Windows

Redis has no native Windows build. Docker is the recommended path:

docker run -d -p 6379:6379 redis:7-alpine

Without a reachable Redis, cross-session memory degrades gracefully to the local file backend — attune.memory.session_stash.backend_status() reports fallback: true so the degradation is visible, and no errors are spammed to the session.

---

API Mode

export ANTHROPIC_API_KEY="sk-ant-..."     # Required
export REDIS_URL="redis://localhost:6379"  # Optional

Model Routing

export ATTUNE_AGENT_MODEL_SECURITY=sonnet  # Save cost
export ATTUNE_AGENT_MODEL_DEFAULT=opus     # Max quality

Budget Controls

export ATTUNE_MAX_BUDGET_USD=10.0  # Override

One-flag cheap mode for pattern-matching workflows (forces every inherit-default subagent onto Haiku; security/architect/plan/quality keywords still get their pinned model):

attune workflow run bug-predict --cheap     # Haiku-default subagents
attune workflow run refactor-plan --cheap

See your spend live on the dashboard (attune ops → home) — today / 7-day / MTD / 30-day tiles fed from the Anthropic admin cost-report API.

---

Security

• Path traversal protection on all file operations (CWE-22)
• Memory ownership checks (created_by validation)
• MCP rate limiting (60 calls/min per tool)
• Hook import restriction (attune.* modules only)
• PreToolUse security guard (blocks eval/exec, path traversal)
• Prompt input sanitization (backticks, control chars, truncation)
• PII scrubbing in telemetry
• Automated security scanning (CodeQL, bandit, detect-secrets)

See SECURITY.md for vulnerability reporting and full security details.

---

Privacy & Telemetry

Attune AI keeps usage data local-first. An opt-in, anonymous usage ping is available to help the project understand which workflows people actually use — it is OFF by default and sends nothing unless you explicitly turn it on.

When enabled, each ping carries exactly this, and nothing more:

• the package (attune-ai) and its version
• the workflow name you ran (e.g. workflow.security_audit)
• your OS (darwin / linux / windows) and Python version (e.g. 3.12)
• a rotating, anonymous install id (a random UUID you can reset)
• a timestamp

It never sends paths, code, prompts, arguments, filenames, project names, cost, tokens, or model data — the payload is frozen in source and guarded by a regression test. Transport is fire-and-forget with a short timeout, so it can never block, slow, or crash the CLI, and the collection endpoint stores no IP address and no request headers.

attune telemetry status     # show exactly what would be sent
attune telemetry enable     # opt in (mints an anonymous install id)
attune telemetry disable    # opt out

DO_NOT_TRACK=1 and ATTUNE_USAGE_PING=0 force it off regardless of config; ATTUNE_USAGE_PING=1 forces it on. Full payload disclosure is in SECURITY.md.

---

Session continuity & cross-session memory

Lightweight hook surfaces keep long Claude Code sessions oriented and recoverable — and carry what you learned into the next one. All are opt-in via plugin install and silent until they have something to say.

Tunable defaults

• ATTUNE_AI_COMPACT_WARNING_THRESHOLD (default 0.70) — fraction of context window before the warning fires.
• ATTUNE_AI_CHARS_PER_TOKEN (default 4.0) — utilization estimator's chars-to-tokens factor.
• ATTUNE_AI_CONTEXT_WINDOW_TOKENS (default 200000) — context window assumed by the estimator.
• ATTUNE_AI_WORKSPACE_ROOTS (os.pathsep-separated paths: : on POSIX, ; on Windows) — override the workspace roots scanned for specs/.
• ATTUNE_AI_SENTINEL_DIR (default ~/.attune) — directory for the once-per-session warning sentinel.
• ATTUNE_LESSON_RECALL / ATTUNE_JIT_RECALL (set 0 to disable) — off-switches for the prompt-time and tool-call recall hooks.
• ATTUNE_LESSON_RECALL_FLOOR (default 8.0) — minimum retrieval score before a lesson surfaces at prompt time.

The transcript-size proxy is crude but monotonic: the warning fires when the user's total content characters cross the threshold once. If your real auto-compact triggers consistently earlier or later than the warning, drop the threshold to 0.65 or raise it to 0.75.

---

Migration

attune-help and attune-author now ship from the Smart-AI-Memory/attune-ai marketplace alongside attune-ai itself. The separate Smart-AI-Memory/attune-docs marketplace is retired.

New users:

/plugin marketplace add Smart-AI-Memory/attune-ai
/plugin install attune-help@attune-ai
/plugin install attune-author@attune-ai

If you previously installed either from attune-docs:

1. /plugin uninstall attune-help@attune-docs
   /plugin uninstall attune-author@attune-docs
   

2. /plugin marketplace add Smart-AI-Memory/attune-ai
   /plugin install attune-help@attune-ai
   /plugin install attune-author@attune-ai
   

---

Links

• Full Documentation
• Plugin Setup
• attune-gui — Living Docs dashboard
• GitHub Repository

Apache License 2.0 — Free and open source.

If you find Attune useful, give it a star — it helps others discover the project.

Acknowledgments

• Anthropic — For Claude AI, the Model Context Protocol, and the Agent SDK patterns behind the multi-agent orchestration layer
• Boris Cherny — Creator of Claude Code, whose workflow posts validated Attune's plan-first, multi-agent approach
• Affaan Mustafa — For battle-tested Claude Code configurations that inspired the hook system

View Full Acknowledgements

---

Built by Patrick Roebuck using Claude Code.