data-ai MCP Server
Persistent session memory for AI coding agents — local-first, with on-device inference, associative recall, and drift detection. Works with Claude Code, Cursor, and Codex.
Discovered via smithery and last synced Today.
1. Install the package
npx prism-mcp-server
2. Add to claude_desktop_config.json
{
"mcpServers": {
"prism-mcp-server": {
"command": "npx",
"args": [
"prism-mcp-server"
]
}
}
}Config file location: ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) / %APPDATA%\Claude\claude_desktop_config.json (Windows)
Package
Contents
Prune outdated or incorrect memories (4 modes + dry_run)
Record corrections, successes, failures, learnings
Location-based POI discovery
Announce yourself to the team
Link
8
HDC compositional state resolution with policy-gated routing
No
No
53.96%
Academic paper analysis via Gemini
Real-time internet search
`local` / `synalux` / `supabase` / `auto`
What
0.009
+Regex, Symbol
Semantic + keyword search over all memories
Auto-compact old entries via Gemini summarization
Sync graduated insights to `.cursorrules` / `.clauderules`
Retrieve active plan state and current verification gating position
No
No
No
No
15
Bitbucket
50.7%
Basic
No
~58%
50
Explicit project reload or older-server startup fallback
Set per-project TTL retention policy
Increase entry importance (+1)
Decrease entry importance (-1)
JS extraction over local search results
JS extraction over web search results
Pulse every ~5 min to stay visible
Decompose natural language goals into an execution plan that references verification requirements
No
No
No
No
No
~0.8s
What it uses from Prism
Routing accuracy
≥90%
2,000 req
0.1855
Tests
Cost
Reclaim ~90% vector storage (v5.1)
What It Prevents
`) significantly boosts the memory's ranking score. Extreme failures and extreme successes surface to the top. * **UX Warnings:** If an agent retrieves memories that are historically negative, Prism intercepts the prompt injection: `⚠️ Caution: This topic is strongly correlated with historical failures. Review past decisions before proceeding.` Your AI now has a "gut feeling" about bad code. ### The Paradigm Shift
8
Local 27B
What it means
4
56.73%
Single
Value
Full export (JSON, Markdown, or Obsidian vault `.zip` with `[[Wikilinks]]`)
Universal post-processing with 8 built-in templates
See all active teammates
No
Revert to any previous version (non-destructive)
AI-grounded answers from Brave
Scores task complexity and recommends host vs. local Claw delegation (`claw_run_task` when delegable; host fallback when executor/tooling is unavailable)
No
No
Headline
15
No
No
No
GDPR-compliant deletion (soft/hard + Art. 17 reason)
No
No
No
No
Emergency kill switch to halt a running background pipeline
If cloud
No
No
No
Required
No
72.51%
~2–3s
Brain integrity scan + auto-repair (`fsck`)
Browse all historical versions of a project's handoff state
Atomically update step status/result with verification context
No
No
Feature
—
Details
Create and enqueue a background autonomous pipeline
No
No
No
> format · disambiguation │ │ rules for similar tool pairs │ ├──────────────────────────────────────────────────────────────┤ │ Layer 2: SFT TRAINING (244 examples, 4 rounds × 500 iters) │ │ 142 tool examples + 102 reasoning/abstention examples │ │ 21 keyword-aware Chain-of-Thought templates │ ├──────────────────────────────────────────────────────────────┤ │ Layer 3: INFERENCE-TIME VALIDATION │ │ Post-inference regex filter rejects false positive tool │ │ calls when prompt matches general programming patterns │ │ (context manager, LSTM, Express.js) without Prism intent │ └──────────────────────────────────────────────────────────────┘ ``` <details> <summary><strong>Source: Layer 3 Inference-Time Validator</strong></summary> ```python # General programming patterns — NOT Prism tools GENERAL_PROGRAMMING_PATTERNS = [ r'\bcontext\s+manager\b', r'\bcontextlib\b', r'\b__enter__\b', r'\bforget\s+gate\b', r'\blstm\b', r'\bcatastrophic\s+forgetting\b', r'\bexpress\.js\b', r'\bdjango\b', r'\bflask\b', r'\bgarbage\s+collection\b', r'\bload\s+balanc', ] # Prism-specific intent (overrides rejection) PRISM_INTENT_PATTERNS = [ r'\bprism\b', r'\bsession\s*ledger\b', r'\bhandoff\b', r'\bknowledge\s+base\b', r'\bproject\b', r'\bledger\b', r'\bsave.*(?:session
Vector similarity search across all sessions
Poll the current step, iteration, and status of a pipeline
No
If cloud
No
No
100%
Pull people / projects / decisions from text
Know srch
OS Web Speech API
Cloud model
**No**
Append an immutable session log entry
What it measures
Resend
Requires BAA
Save live state for the next session
**TTS (Text-to-Speech)**
LiveKit
Latency
> **Methodology**: 102-case pool across 12 categories. Scores are 3-seed mean (seeds 2027/2028/2029, zero variance across all seeds). All fine-tuned models use the Qwen3 nothink template with keyword-trigger routing prompts and `-> respond directly (no tool)` for the no-tool class. Full runner: [`tests/benchmarks/prism-routing-100/benchmark.py`](tests/benchmarks/prism-routing-100/benchmark.py) · Cascade runner: [`tests/benchmarks/cascade-14b-32b-opus/cascade_eval.py`](tests/benchmarks/cascade-14b-32b-opus/cascade_eval.py). > > **These are NOT general-purpose LLM benchmarks.** This eval measures routing precision on 6 specific MCP tools. The prism-coder models are specialists trained on this exact task — they match or exceed Claude on routing while Claude dominates on general reasoning, coding, and open-domain QA. The value is **offline reliability at zero cost**, not replacing cloud AI. **iOS deployment:** On-device inference via **llama.cpp Swift SPM**. Auto-selects by device RAM: 14B on iPad Pro 16GB (100% routing), 8B on iPhone/iPad 8GB (100%, OOM fallback to 1.7B at 100%). CoreML not viable — coremltools doesn't support Qwen3 attention ops. Integration: `LLMEngine.swift` → `prismNativeBridge.askAI()` → token stream. WiFi fallback: Mac Ollama (`OLLAMA_HOST=0.0.0.0`). ### Benchmarks — run them yourself All benchmarks are open-source. Reproduce every number in this README: ```bash git clone https://github.com/dcostenco/prism-coder cd prism-coder pip install anthropic requests # Per-model solo eval (102 cases, 3 seeds) python3 tests/benchmarks/prism-routing-100/benchmark.py --models 14b 8b 32b 1b7 opus # Cascade eval — 14B → 32B → Opus (Claude Opus as etalon) export ANTHROPIC_API_KEY=sk-ant-... ollama pull dcostenco/prism-coder:14b dcostenco/prism-coder:32b python3 tests/benchmarks/cascade-14b-32b-opus/cascade_eval.py ``` **Not a general function-calling benchmark.** This measures routing precision on 6 specific MCP tools. We don't claim to beat Claude on general capabilities. We match or exceed Claude on the ONE task that matters for offline AAC: correct tool routing, every time, under 2 seconds, with zero cloud.
Avg lat
Provider
No-tool
Primary
WASM espeak-ng
+ HIPAA BAA + custom fine-tuning
Offline
Twilio
What happened
Memory-first Q&A with a grounded live-source fallback on paid tiers
Srch mem
Inworld TTS-2 (cloud)
Offline dictionary (1,261 × 20 langs)
—
Stripe
What it does
Auto-link related memories into a graph
unlimited
Think
21s
Local Ollama models
Automatic routing tier
Free
used_cloud
$0
55s
--- ## Plans All on-device models are free to run locally via Ollama on every tier. A subscription gates **cloud** features, higher model ceilings, and increased limits. Local model ceilings are advisory — on-device models run on your Ollama regardless of plan; the ceiling gates cloud inference and `prism_infer` routing.
2,000
Detect when a session has drifted from its goal
What it checks
License
Pre-edit scenario challenge — catch bad changes before they happen
Dead code, orphan services, scaffold code, missing auth
Local-first inference (route/chat/code modes, thinking, cloud escalation)
Default
Session delegation or persisted MCP + VS Code panel local/cloud stats
Off (fast) — except a tier that reasons better, e.g. 9B
Ollama endpoint
Per tier: on for 9B, off for 4B/2B
Force local SQLite regardless of credentials
Per tier: on for 9B, off for 4B/2B
Portal analytics token (optional — metrics display works without it)
Paid-tier portal key (`synalux_sk_...`)
Managed containment
synalux
All skills available locally (run `sync-skills.sh` to populate)
1
`features.multi_agent=false`; a 2-thread, depth-1 Terra/low fallback profile is retained for explicit re-enable
Amazon Q Developer
Network latency
Memory text, file paths, search queries, and inference prompts/drafts when their cloud feature is used, sent to the portal over TLS. Cloud memory writes are PHI-redacted; inference and route requests are transient.
Routing accuracy¹
Prism Coder
Zep
1,024
Hook-free first-turn greeting and dashboard-configured context
Teach Prism a codebase or document
Token volume local serving kept off your cloud model, all time / 30 days / session
week
Default mode is `advisory` (warn but allow). Set `*_MODE=block` for hard enforcement. Hooks look for full audit scripts in the repo first (`hooks/lib/`), then `~/.claude/hooks/` fallback, then minimal inline checks. --- ## Self-hosting (Enterprise) Run the full model stack on your own hardware — no cloud, full data sovereignty. **Requirements:** Mac M2 Pro+ (48 GB recommended) or Linux + NVIDIA GPU, plus [Ollama](https://ollama.com). ```bash ollama pull dcostenco/prism-coder:9b # default router export LOCAL_LLM_URL=http://localhost:11434 ``` Self-hosted routing stays local: `9b → 4b` on desktop/server and `2b` on mobile/iPhone, with 27B available when installed and RAM-safe. Synalux-hosted paid tiers can use Gemini 3.6 Flash as the cloud fallback. For iOS or another machine on the same network, run `OLLAMA_HOST=0.0.0.0 ollama serve` and point `LOCAL_LLM_URL` at the host's IP. --- ## Configuration reference