M
MCP Directory CURATED DIRECTORY
911 MCP servers indexed
中文
← Back to directory
P

Prism Coder

Community
Persistent memory + local AI for coding agents.
How to install ↓ Compare GitHub source repository ↗
Category
Dev Tools #121 of 438
Stars
★ 157 Very popular
Transport
stdio (local process)
Runtime
Node.js · Python
Credentials
No credential needed
License
Apache-2.0
Last commit
Tools
11
60FMRS · C

Prism Coder is a feature-rich MCP server offering persistent memory, local inference, and advanced verification. Ideal for local-first workflows, but setup may require managing multiple components.

Strongest · Documentation 16/20 Weakest · Reliability 6/20

Reliability
6/20
Security and permissions
8/20
Maintenance
16/20
Documentation
16/20
Setup experience
14/20
Why each score
Reliability 6/20
No source code, CI workflows, or committed tests were provided; the manifest lists env vars but not a tool catalog, so declared tools cannot be cross-checked against handlers. The release notes document several past functional failures (verify_behavior crash, browser assertions not failing, fingerprint layer never applied), which shows edge-case history. Happy-path setup via npm/stdio is plausible, but without execution evidence or test files, confidence is limited; deducted for missing verification artifacts.
Security and permissions 8/20
No real tokens appear in install examples; secrets are declared via env vars and marked isSecret. README discloses data flow (local SQLite vs Supabase/Synalux cloud), PHI redaction claims, fail-closed route guard, and an SSRF hardening release. However, these are self-reported and no source or audit is supplied to verify least privilege, confirmation workflows, or network boundaries. No red-line violations are evident from the supplied material, but unverifiable claims cap the score.
Maintenance 16/20
Repository is not archived, has an Apache-2.0 license, npm package version 20.7.0, and a detailed release history showing frequent updates and dependency patches. Only one open issue is visible. There is no explicit security-response channel or public roadmap, and commit/issue-response evidence was not supplied, so a small deduction applies.
Documentation 16/20
README is extensive: quickstart, CLI, host-specific behaviors, configuration reference, plan limits, and limitations (self-published benchmarks, hook-free caveats, experimental namespace risk). It links to deeper docs and includes upgrade notes. However, tool-by-tool parameter schemas are referenced but not shown, and troubleshooting is minimal; some claims (e.g., HIPAA-ready, accuracy numbers) are not backed by independent evidence.
Setup experience 14/20
Setup is clearly described: global npm install, `prism connect`, dry-run/refresh flags, host support and env vars, plus manual IDE setup links. A mainstream client can register the stdio server. However, no CI execution or committed test evidence was supplied to confirm the install path, and Claude Desktop on Windows/Linux is marked beta; static calibration caps setup at 15, and missing runtime verification justifies 14.

Static review · not runListed 2026-08-07

Read the FMRS scoring method →

Fit and risk

What it can accessRuns commands or codeUses the networkControls a browserConnects to a database

Best for

  • Developers who value privacy and want a local-first workflow
  • AI agent users who need reliable tool routing and verification
  • Teams that need cross-session memory and knowledge graphs

Not for

  • Users who need full cloud features without a Synalux subscription
  • Those who need to run large models on modest hardware (27B requires ~16GB)
  • Scenarios requiring official vendor support

Required permissions

  • Filesystem access to read/write local SQLite database and config files
  • Network access for Brave Search, Gemini AI, Supabase cloud sync, and Synalux portal
  • Execution of local Ollama models for inference

Risks and side effects

  • Local storage may contain sensitive data; encryption recommended
  • Cloud features require API keys which could leak if mishandled
  • Model outputs may hallucinate; verification is essential
  • Local model routing may lead to stale information if context not managed properly

Setup

Before you start

Runtime：Node.js · Python

BRAVE_API_KEY optionalsecret Brave Search API key for web search tools; get it at brave.com/search/api — not needed for local free-tier use.
GEMINI_API_KEY optionalsecret Google Gemini API key for AI-powered summarization and embeddings; obtain from Google AI Studio.
SUPABASE_KEY optionalsecret Supabase service role key for database access; treat as a secret.
Other optional settings (5)
SUPABASE_URL optional Supabase project URL for session memory and knowledge persistence.
PRISM_USER_ID optional User ID for multi-tenant row-level security isolation; defaults to 'default'.
GCP_PROJECT_ID optional Google Cloud project ID for Vertex AI Discovery Engine integration.
VERTEX_LOCATION optional Vertex AI location/region; defaults to 'global'.
VERTEX_DATA_STORE_ID optional Vertex AI Discovery Engine data store ID for enterprise search.
  1. Install the npm package globally: npm install -g prism-mcp-server.
  2. Run prism connect to auto-detect and register Claude Code, Claude Desktop, Cursor, Gemini CLI, and Codex. Use --host <name> to target a specific host, --dry-run to preview changes.
  3. Restart your MCP hosts.
  4. (Optional) Pull local models: ollama pull dcostenco/prism-coder:2b, :4b, :9b, or :27b.
  5. (Optional) Set environment variables like BRAVE_API_KEY, GEMINI_API_KEY, SUPABASE_URL, etc.

Check that it works

Install prism-mcp-server globally and run prism connect, then restart the host; session_bootstrap and the other Prism tools should appear in the client's tool list, and the first turn should show dashboard-configured project context.

Troubleshooting

  1. Ensure all required environment variables are set (e.g., `BRAVE_API_KEY`, `GEMINI_API_KEY`)
  2. Run `prism connect --dry-run` to check changes
  3. Restart MCP hosts to apply config
  4. Check that local Ollama service is running and accessible
  5. Consult docs and issue tracker for known issues

Things to try

Once connected, you can ask your AI assistant things like:

  • Call session_bootstrap to load context for my current project
  • Search my memory for what I decided about the auth flow last month
  • Use knowledge_ingest to teach Prism this codebase
  • Use prism_infer in code mode to generate a binary search in Python locally

Tools 11

session_bootstrap read-only
Hook-free first-turn greeting and dashboard-configured context.
session_load_context read-only
Explicit project reload or older-server startup fallback.
session_save_ledger writes
Append an immutable session log entry.
session_save_handoff writes
Save live state for the next session.
knowledge_search read-only
Semantic + keyword search over all memories.
query_memory_natural read-only
Memory-first Q&A with a grounded live-source fallback on paid tiers.
session_detect_drift read-only
Detect when a session has drifted from its goal.
verify_behavior read-only
Pre-edit scenario challenge — catch bad changes before they happen.
Show 3 more tools
knowledge_ingest writes
Teach Prism a codebase or document.
prism_infer read-only
Local-first inference (route/chat/code modes, thinking, cloud escalation).
inference_metrics read-only
Session delegation or persisted MCP + VS Code panel local/cloud stats.

Use cases

Coding assistants that retain project context and decisions across sessions
Offline-first AI tool routing with local models
Multi-agent coordination with shared project memory
Reliable AI edits with verification and drift detection

Supported clients

Claude Desktop
Claude Code
Cursor
Gemini CLI
Codex

Listed from the project's documentation, not tested by this site.

Overview

Prism Coder is an MCP server that provides long-term memory across sessions for AI tools like Claude, Cursor, Gemini CLI, and Codex. It ships with an open-weight model fleet (2B–27B) for fast, offline tool routing, no cloud required. The free local tier offers persistent session memory (knowledge graph, session ledger, time travel), cognitive routing, multi-agent hivemind, behavioral verification, and drift detection. Paid subscriptions unlock cloud sync, higher model tiers, and team features. The server also includes a CLI for connecting multiple hosts, model management, and local browser testing.

Similar servers

Context7 80 · B

Upstash's official server providing up-to-date third-party library docs for AI coding assistants

★ 63.3k · Tools 2 Compare with this →
Compare with other MCP servers

Source revision a6e09a1100b8 Data synced 2026-10-19 Read the FMRS scoring method