Prism MCP

What dcostenco/prism-mcp MCP server does

The dcostenco/prism-mcp MCP server adds durable memory to AI agent workflows. Session information can survive restarts, allowing an agent to resume work with handoff notes, recent activity, pending TODOs, and a selected context depth. The project also provides knowledge-graph features, local tool routing, a web-based Mind Palace dashboard, time-travel history for sessions, cross-client memory sharing, code-mode templates, and morning briefings.

The package exposes 25 tools, 6 resources, and 4 prompts. Its local savings reporting records calls handled on the user's machine and reports token volume, local share, and model-level details rather than converting usage into an estimated dollar value.

How it works

Prism runs locally and uses SQLite for persistent state. Its bundled open-weight prism-coder models support offline routing through Ollama, with model selection based on available installed models, memory, context requirements, and subscription entitlements. Local routing is intended for bounded work; route mode returns calls only for tools advertised by the host.

The session_bootstrap capability supplies startup context, while prism_infer makes local workers available over MCP. The local_savings tool and the prism savings command provide usage reporting. Hosts can request startup context through MCP metadata and native instructions without requiring a Prism-owned launcher or lifecycle hooks.

Setup and configuration

Install the npm package globally and run prism connect to configure supported agent hosts. The command preserves unrelated settings. --dry-run previews changes, --all configures every detected host, and --refresh reconciles Prism-managed entries after an upgrade. Hosts should be restarted after connecting.

The dcostenco/prism-mcp MCP server does not require an account, API key, or cloud subscription for local operation. An optional Synalux subscription adds cloud synchronization, higher model tiers, paid skills, and team features. Prism can also be installed as a Claude Code or Codex plugin, but the plugin and a manual prism connect setup should not both register the server.

Tools and capabilities

• Preserve session memory across restarts.
• Load quick, standard, or deep startup context.
• Route bounded tasks to local Ollama models.
• Share memory between supported clients.
• Inspect project state and session history through the Mind Palace dashboard.
• Rewind and replay session history.
• Generate savings reports for locally served work.
• Synchronize subscription-aware skills when entitled.

Limitations and notes

Local routing depends on compatible models being installed through Ollama and on the machine having enough memory and context capacity. The savings report measures tokens and documents its assumptions; it is not a direct accounting of money saved and may undercount some usage.

Cloud memory, higher model tiers, and team features require the optional Synalux subscription. Host configuration behavior differs by client. For example, Claude Code retains subagents but points them toward an economy model, while Gemini CLI and Codex receive settings that disable host subagents. These settings are reapplied whenever prism connect runs.

Hosts without lifecycle hooks, including Cursor, receive the startup digest during bootstrap but do not use Prism's hook-based post-compaction reinjection. The README also marks Codex compaction behavior as unverified end to end.