dcostenco/prism-mcp

Install

npx -y prism-mcp-server

Capabilities

• tools
• logging
• prompts
• resources

Server instructions

Prism MCP — The Mind Palace for AI Agents. On the first user turn of every conversation, including greetings, call session_bootstrap exactly once BEFORE any user-facing response, passing the user's verbatim first message as {prompt: "<first user message>"}. The prompt is matched against prompt_keywords ON-DEVICE to load symptom-triggered skills on turn one; it is used for routing only and never leaves the machine. Pass {} only when there is no user message. When the startup display lists "Symptom-triggered skills", those skills matched the first message's trigger rules — read each before proposing any change. Surfacing a name is not loading it: hosts that do not auto-load skill files must fetch each body with knowledge_search, passing the skill name exactly as listed. Emit no preamble. Print the complete tool result verbatim as the entire first-turn startup display, before any optional answer. Do not summarize, paraphrase, rename headings, reformat, or omit any returned section. Preserve its order and line content. For a greeting-only prompt, stop after the verbatim startup display. Do not substitute session_load_context while session_bootstrap is available; use session_load_context only for an explicit project reload or as an older-server fallback. Use session_save_ledger to log completed work and session_save_handoff to preserve state for the next session. Reuse the conversation_id returned by session_bootstrap in structuredContent for those saves and for session_detect_drift, the 60-minute goal-alignment drift check. Do not add the id to the visible greeting. ## Prism local-first orchestration For bounded, verifiable delegated work, use Prism's local worker before any host-native or background subagent. When session_task_route is available, call it first. If it returns target `claw`, call `prism_infer` with the recommended arguments, `cloud_fallback: false`, and the `project` and `conversation_id` from this session when known. Prism loads that project's dashboard-configured quick, standard, or deep memory and selects a RAM-safe local model. Never create host-native or background subagents for routine work, never fan out, and never nest agents. If local inference is unavailable, refused, degraded, or the task requires host tools or reserved judgment, continue in the current host thread. A host-native subagent is a last resort: at most one, no nesting, using the configured economy model. The current host remains responsible for verifying local output before using it. ## Prism evidence workflow During diagnosis and editing, one trustworthy correlated reproduction is enough; do not block coding on inspecting unrelated diagnostic screenshots, trace frames, or abandoned attempts. Before a completion claim, push, or release, exercise the corrected path with fresh evidence from the current build and bind stateful proof to the exact run and entity. Inspect every artifact used to support the claim. When a screenshot is requested or used as proof, the active agent must open it and compare its visible state with the issue's expected and forbidden states. Complete that review yourself; do not ask the user to verify it. A screenshot is an observation, not absolute truth. Reject stale, wrong-run, wrong-entity, or visibly failing evidence even when its metadata says passed. Architecture: session_save_ledger and session_save_handoff require context loaded by session_bootstrap or session_load_context when a conversation_id is supplied. Safety boundaries are enforced in code — shown so hosts avoid wasted round-trips. - **Crisis/self-harm** inputs are intercepted before reaching any model. - **BCBA reserved categories** (restraint, seclusion, physical management, dosing): Layer 1 classifies; RESERVED/UNCERTAIN escalate to cloud or are refused. A keyword backstop covers classifier failure; oversize prompts get a full-text keyword scan + excerpt classification. - **Dangerous output** (restraint instructions, overdose methods, self-harm guidance) is blocked regardless of host. - AAC access is never restricted as a consequence. All cloud inference routes through the Synalux portal for billing, tier-gating, and audit.

Links

• Repository ↗
• JSON record (API)
• Tool schemas (API)