Prism Coder

Give your AI agent memory that lasts. Persistent sessions, knowledge graphs, and offline tool-routing — fully local and free.

npm MCP Registry License: Apache-2.0 Models on HuggingFace

Prism Coder is an MCP server that gives Claude, Cursor, and other AI tools long-term memory that survives across sessions. It ships with the open-weight prism-coder model fleet (2B–27B) for fast, offline tool-routing — no cloud required.

No account needed. No API keys. Runs on your machine.
A paid subscription adds cloud sync, higher model tiers, and team features through the Synalux portal.

What Prism gives you

  • Session memory that survives restarts — resume projects with handoff notes, recent work, open TODOs, and configurable quick, standard, or deep context.
  • Local-first inference — bounded work is routed through local Ollama models first, with automatic 2B/4B/9B/27B selection based on installed models, available RAM, context fit, and subscription entitlements.
  • Route-output enforcement — route mode returns only well-formed calls to tools the host actually advertised. Standard and higher plans can add authenticated deterministic correction; route_guard: "local" keeps the prompt and draft entirely on-device.
  • One setup for every agent — prism connect configures Claude Code, Claude Desktop, Cursor, Gemini CLI, and Codex while preserving unrelated settings.
  • Subscription-aware skills — entitled skills are synchronized before the host launches, with safe upgrades, downgrades, conflict preservation, and offline last-good recovery.
  • Hook-free startup — MCP metadata and native instructions request Prism's startup context without requiring lifecycle hooks or a Prism-owned launcher.
  • Safe escalation and observability — inference outcomes are explicit, reserved content remains fail-closed, and local/cloud usage is recorded for review.

Get started

npm install -g prism-mcp-server
prism connect

Use prism connect --dry-run to preview changes, prism connect --all to configure every detected host, or prism connect --refresh to reconcile Prism-managed entries after an upgrade. Restart the host after connecting.

Prism works locally without an account, API key, or cloud subscription. Add a Synalux subscription when you want cloud memory, paid-tier skills, or team features.

Install as a plugin

Prism also ships as a plugin, which registers the MCP server and the startup skill for you.

Claude Code — from the community marketplace:

/plugin marketplace add anthropics/claude-plugins-community
/plugin install synalux-prism@claude-community

Codex — this repository is itself a plugin marketplace:

codex plugin marketplace add dcostenco/prism-coder
codex plugin add synalux-prism@prism

The plugin registers prism-mcp via npx -y prism-mcp-server. If you already configured Prism by hand — prism connect writes an mcp_servers.prism-mcp entry — you have that server twice under one key. Install the plugin or run prism connect, not both.

What prism connect changes about host subagents

connect steers bounded work to prism_infer on your machine rather than to host-spawned agents. What it writes differs per host, and it does not disable subagents everywhere — Claude Code keeps them and is pointed at an economy model instead. Prism's local workers stay available over MCP in every case.

Host Setting written Effect
Claude Code env.CLAUDE_CODE_SUBAGENT_MODEL = "sonnet" in ~/.claude/settings.json Subagents stay enabled, pinned to an economy model. Fan-out is discouraged by policy text, not by config
Gemini CLI experimental.enableAgents = false in ~/.gemini/settings.json Subagents off. Gemini exposes one boolean, so that is all there is to set
Codex features.multi_agent = false in $CODEX_HOME/config.toml (default ~/.codex), plus a bounded fallback: 2 threads, depth 1, cheap subagent model, 900s cap Subagents off, with a bounded profile underneath so a deliberate re-enable lands somewhere sane

Two things worth knowing:

  • experimental is Gemini's namespace, not ours. Prism is not enabling anything experimental — it writes false to a flag Gemini already defines at that path. Writing anywhere else would have no effect.
  • That namespace is by definition temporary. If Gemini promotes enableAgents out of experimental, Prism keeps writing the old path, Gemini reads the new one, and host subagents quietly turn back on. Nothing errors and the settings file still looks correct. If you see host subagents running while enableAgents reads false, check whether the key has moved before assuming connect failed to write it.

Both writes are idempotent in the sense that a host already configured this way is left untouched — but they are re-applied on every prism connect run, not only on --refresh. If you deliberately re-enable host subagents, the next connect will turn them off again. Keep them on by not re-running connect, or by re-enabling after each run.

What's New in v20.12.1

  • prism connect --refresh now converges every registration it owns, not just the top-level one — directory-scoped entries could otherwise keep launching an old build indefinitely.
  • prism update checks the installed package, not the CLI that happens to be running, so it can no longer report "current" while the install is stale.
  • The opt-in scheduled updater can actually start — the LaunchAgent now carries a PATH that includes node and npm.

What's New in v20.12.0

  • Prism now tells you when it's out of date. Session startup shows a one-line update notice when a newer release exists — cache-backed, at most one registry check per day, silent offline. PRISM_NO_UPDATE_CHECK=1 opts out.
  • Hands-free updates, if you want them. prism autoupdate enable sets up a daily prism update --if-idle: it updates only the global npm package, defers while any Prism server is running, and never touches host configuration — that stays behind a visible prism connect.

What's New in v20.11.1

  • Saving memory never gets refused. The save path used to reject session_save_ledger/save_handoff calls when its path-to-project heuristic disagreed with the project you declared — and the registry the heuristic trusted could contain junk from earlier auto-registration, so legitimate sessions ended unsaved. Your declaration now always wins; the disagreement is returned as an advisory warning, and auto-registration only accepts real repository roots.
  • Screenshots are evidence again. prism browser captures on macOS were silently upscaled to the size cap, so a screenshot no longer showed what actually rendered. Only genuinely oversized captures are resized now, and the cap no longer clips a standard 1920-wide viewport.

What's New in v20.10.0 – v20.11.0

  • Skill routing now works mid-session. New prompts are matched on-device as the conversation moves — not just on turn one — and injected within each host's real context limits (Claude Code caps hook output at 10k chars; Codex truncates by default), with pointer-first delivery when a payload can't fit inline.
  • prism connect is a converge command. It self-updates first, re-execs, then reconciles MCP registration, skills, and hooks — no more "fresh config, stale code" machines.
  • Scoped skills route on prompts too, and startup output survives hosts that discard structured tool content.

What's New in v20.9.0 – v20.9.3

  • Your skills follow your account. skill_save stores a skill at the scope you choose: this machine only (local, works offline and signed out), your account (user — every machine you sign into receives it), or a workspace (team — shared with members, admin-managed, optionally targeted to specific people).

  • Trim the catalog you don't use. skill_manage can release platform skills you never touch — freeing host skill-catalog budget — and restore them any time, losslessly. Deleting a scoped skill archives its final content locally first, so nothing is ever silently unrecoverable.

  • Delivery that queues instead of failing. Concurrent sessions no longer starve skill sync on the local config store (WAL + busy-timeout) — a failure that previously reported only "partial" where nobody could see it.

  • Withheld rules still bind. When the context budget can't inline a skill's text, the manifest of withheld names now states that those skills still govern the work and names every way to load them before completion claims.

  • The budget the floor never spent. A long-standing accounting bug meant no unprotected skill ever inlined at any normal context level — the always-inlined protected floor was debiting the budget meant for everything else. Task-matched skills (like the completion-evidence checklist) now actually arrive.

What's New in v20.8.2

  • Skill delivery now admits failure instead of hiding it. A filesystem permission edge case (a umask stripping the owner-execute bit) could leave skill sync writing nothing while reporting itself current — measured at nine days on a real machine. Broken managed directories are repaired in place, every directory is created umask-proof, and the repair path refuses symlinks via an O_NOFOLLOW descriptor.
  • A stale install tells you at startup. Prism now tracks the generation that actually reached disk separately from the one the database accepted; if they diverge, the startup banner says so in a warning placed where display truncation cannot cut it. A successful sync clears it automatically.

What's New in v20.7 – v20.8.0

  • First run proves the memory instead of describing it — session_bootstrap seeds one demo memory and shows it recalled from disk, so the save→recall loop is felt in session 1. One-shot, contained in its own prism-demo project, removable with one call.
  • Dashboard fixed — a quoting typo (shipped 2026-05-29) killed the inline script at parse time, so every dashboard since rendered "Loading projects..." forever. Fixed, and the ES5 lint now node --checks the built inline script so an unparseable dashboard can never ship again.
  • Trusted Publishing — npm releases authenticate via GitHub OIDC. No stored token to expire or leak, and every release carries a signed provenance attestation — you can verify the tarball you install was built from this repo by CI (npm audit signatures).
  • TLS enforced for cloud sync — a remote http:// storage URL is upgraded to https:// instead of silently sending session content in the clear.
  • Codex plugin collision + enabled-state detection — prism connect skips its own registration only when a plugin actually provides prism-mcp (cache present and enabled), preventing both duplicate and missing servers.
  • Windows CI stabilized; registry/npm listings realigned and deduplicated.

What's New in v20.6.0

Delivery Is Not a Suggestion

An audit of a real incident (an agent wiped demo data after announcing the wipe — with the ask-first rule committed, bundled, and absent from what any agent actually received) found the protected floor had outgrown every delivery budget: "unprotected" had quietly come to mean "never delivered".

  • ask-first and feature-preservation join the protected floor (14 → 16). Protected skills are always inlined; these two now reach every session.
  • Sync conflicts are loud and named. Startup used to say "· 2 local conflicts preserved" while safety skills sat months stale; it now names each frozen skill and states how to resume updates.
  • --storage accepts auto and synalux — the CLI rejected its own documented default and the production backend.
  • Disclosure: skill delivery informs; it does not gate. A live probe showed a host agent still edit unverified source with the rule loaded. If your threat model includes an agent acting against a loaded rule under task pressure, pair this package with mechanical gates (hooks, permissions, least-privilege roles). True of every prior release; stated from this one.

What's New in v20.5.3

Grounding Evidence Carries Its Age

Memory-grounded answers labelled their sources but never dated them, so a two-year-old note and yesterday's reached the model identically. Nothing in the evidence let it discount the stale one. Prompted by an external review naming the right risk for local-first memory: the data stays local, but bad grounding becomes permanent — storing everything on your machine removes the outside pressure that would otherwise surface a stale note.

Evidence now reads:

[SOURCE 1: ledger:8286581d (recorded 2025-05-29, 431 days ago)]

The date already existed in storage and was being dropped at the snippet layer, so this is plumbing rather than new data collection. Zone-less SQLite timestamps are normalised to UTC — read as local, a ten-minute-old record parsed hours into the future and its age was suppressed entirely, meaning the feature silently did nothing on the freshest memories. An absent or unparseable date renders as nothing rather than defaulting to now; defaulting would make the oldest memories, the ones most likely to be stale, appear freshest.

tests/integration/grounding-staleness.test.ts runs the reviewer's own probe — seed a deliberately outdated note beside a contradicting fresh one and assert the model receives both, visibly dated. Anyone can run it.

Not solved, and not claimed: retrieval does not weight recency. A stale note shown beside a fresh one is the easy case — the model sees both dates and can weigh them. The hard case is a stale note retrieved alone, because ranking is by keyword match and an old store returns old results; then the age label is the only defence and there is no fresher record to compare against. Tracked as TECH_DEBT.md #4.

What's New in v20.5.0 – v20.5.2

The First Message Never Leaves Your Machine

Symptom-triggered skills — the rules that fire on "can't see X", "no rows", "the list is empty" — are meant to load on the turn an incident report arrives. They never did: every host template called session_bootstrap with {}, so there was no prompt to match against.

Fixing that raised the question of where matching happens. It now happens locally. The 28 keyword rules are already public, so there was nothing a local match could not compute, and callPortal() has no prompt parameter at all — the guarantee is structural, not a promise. The portal request carries the project and role only.

A matched rule now arrives as content, not as a name. Native hosts outside the skill-file mirror had no way to read a rule they were only told about, so the rule body is inlined into the startup display, bounded and sized against the real per-project budget.

What's New in v20.4.0

An Explicitly Named Cloud Backend Fails Loud

Setting PRISM_STORAGE=synalux or =supabase with incomplete credentials used to downgrade silently to local SQLite. The switch was logged to stderr, which MCP hosts discard, so nothing surfaced it: sessions kept serving stale local context while the cloud held newer history, and context_source read local rather than any kind of warning. A session could run that way for weeks.

Naming a backend outright is a strong statement of intent, so it now throws — naming the missing variables and the PRISM_STORAGE=local opt-out — instead of quietly splitting your session history. auto is unchanged: it keeps its documented synalux > supabase > local degradation, pinned by a test.

Upgrade note: if you explicitly set PRISM_STORAGE=synalux|supabase and your credentials are incomplete, startup now fails with a named error instead of silently using local data. That error is the fix — set the missing variable, or choose PRISM_STORAGE=local deliberately. Default (auto) configs are unaffected.

The throw is deliberately not treated as a recoverable startup fault: that path exists for transient errors (rate limits, 5xx, DNS), which may degrade behind a visible notice. A missing credential is a configuration fault and must not be papered over.

Also: the skill block is now budgeted by default rather than only on request, so a large skill payload cannot crowd out briefing and history.

What's New in v20.3.2

Web Scholar: SSRF Hardening

Security release. Web Scholar scrapes article URLs that come from search-engine output, so the target is attacker-influenceable through SEO poisoning — and because what it scrapes is written into the memory corpus and passed to the configured LLM, a redirection to a local address meant reading an internal service and sending the result onward.

The host guard matched string prefixes instead of parsing the address, and six spellings of a local address got through: [::1] (URL.hostname keeps the brackets), 127.0.0.2 (only .1 was enumerated, not all of 127.0.0.0/8), 0.0.0.0, [::ffff:127.0.0.1], localhost. (a trailing dot defeated every suffix check at once), and [64:ff9b::7f00:1] (NAT64 embeds IPv4 in its low bits). Host classification now parses addresses and also covers CGNAT, benchmarking, multicast, reserved, and IPv6 unique-local and link-local ranges.

DNS rebinding is closed too. Every check read the URL string, so a hostname the attacker controls passed all of them and could still resolve to 127.0.0.1. Targets are now resolved first, every returned address is validated, and the connection is pinned to those addresses so the name is never resolved a second time — which also shuts the window between the check and the connect.

Scrape failures no longer vanish into a bare catch {}, a run is bounded by PRISM_SCHOLAR_SCRAPE_BUDGET_MS (default 60s) instead of stalling on a raised article count, and responses are capped at 8 MiB.

This is reachable only when scholar actually runs — scholar_research, or the background loop under PRISM_SCHOLAR_ENABLED=true — and when the attacker also controls DNS or a search result. Upgrade if you use Web Scholar.

What's New in v20.3.1

Prism Browser Reports Real Failures

prism browser could not fail a test. eval 1 === 2 returned status: ok with exit code 0, a page serving HTTP 500 reported status: ok, and console errors and uncaught page exceptions were discarded entirely. This release adds assertions — assert-text, assert-visible, assert-count, assert-url, assert-title, assert-eval, assert-no-page-errors — that return status: failed and a non-zero exit. open now reports http_status and fails on 400 or higher, screenshots are validated rather than assumed, and eval returns native JSON with its type instead of a Python repr.

The fingerprint layer had never been applied: a wrong keyword argument made the stealth library throw on every launch — 1,139 failures and 0 successes since April — while the runner reported it as active. It is fixed, and a layer that cannot be applied now fails loudly. The headless build no longer advertises itself through navigator.userAgentData or the Sec-CH-UA header, and a patch that corrupted Object.getOwnPropertyDescriptor on every page under test has been removed. These remain best-effort test aids, not a guarantee against bot detection.

--local-only now actually isolates: WebSocket, EventSource, WebRTC and sendBeacon egress bypass request routing and were never blocked, and service workers were allowed through. --cleanup was a no-op in the two modes agents use. Site isolation, phishing detection and popup blocking are no longer disabled by default, since these profiles hold live authenticated cookies.

New for test runs: --ephemeral-profile and --storage-state for hermetic authenticated flows, pages/switch-page so OAuth popups are reachable, --fail-fast, --fast, --trace/--video/--har, and profiles --prune-older-than for profile maintenance.

What's New in v20.2.7

Session Saves Survive Agent Restarts

Prism now remembers that a conversation successfully loaded its project context when the MCP server restarts or another Prism process handles the next request. session_save_ledger and session_save_handoff no longer fail with a false context_not_loaded error in that flow.

The recovery remains fail-closed: authorization is limited to the exact project and conversation, expires with the existing context window, and stores no plaintext conversation identifier. Cross-project, forged, malformed, expired, or future-dated receipts are still rejected. The release also updates PostCSS to the patched 8.5.23 release.

What's New in v20.3.0

Hybrid Memory Search (Portal Tier)

session_search_memory on the portal tier (Synalux-backed installs) now fuses semantic similarity with exact-term lexical matching via weighted reciprocal-rank fusion. Measured on blind probes against a real 8.5k-entry corpus: fused retrieval was never worse than semantic alone at top-5, and exact identifiers — TPNs, function names, error strings — now rescue queries that embedding similarity blurs. Results say how they were found — hybrid retrieval headers, per-hit sem#/lex# arms — and a lexical-only rescue is labelled exact-term match instead of pretending to a similarity score. Local SQLite installs keep pure vector search; hybrid needs the portal's lexical index.

What's New in v20.2.6

Safer Configuration Updates Across Every Agent

prism connect now reads Claude, Cursor, Gemini, and Codex configuration through a single verified file snapshot, preventing another process from swapping a file between Prism's safety check and its read. Supported symlinked dotfiles still work, while dangling or planted symlinks fail loudly instead of being followed or overwritten. This release also carries the patched dependencies and cross-platform release checks introduced in v20.2.5.

Cloud fallback is now documented consistently as Gemini 3.6 Flash. Plan ceilings govern automatic prism_infer routing; direct use of any downloaded model through local Ollama remains free on every tier.

What's New in v20.2.4

Reliable Session Memory That Shows Work, Not Greetings

Greeting-only assistant replies are skipped before ledger writes. Existing greeting rows are filtered at read time across native startup, MCP context, and prism load --json, while entries containing decisions, TODOs, changed files, or non-session events remain visible. Historical rows are not destructively deleted. If Synalux has a transient startup failure, Prism displays one bounded local last-good snapshot and clearly labels it; permanent authorization or validation failures still fail loud, and later writes remain cloud-routed.

What's New in v20.2.2

One Local-First Workflow Across Every Agent

prism connect now installs one orchestration contract for Claude Code, Claude Desktop, Cursor, Gemini CLI, and Codex. Bounded delegated work goes to session_task_route and the local prism_infer worker first; routine work must not create background host agents. Local workers can receive the active project's dashboard-configured quick, standard, or deep memory and select a RAM-safe 2B/4B/9B/27B model at call time. The router forwards complexity but does not choose the model; prism_infer owns the final decision using memory and context fit, installed models, live RAM, entitlements, and explicit caller overrides.

Codex and Gemini native agent fan-out are disabled during connect. Codex keeps a two-thread, one-level Terra/low fallback profile if the developer explicitly re-enables native agents later. Claude Code keeps native agents as a last-resort path but pins their model to Sonnet. Cursor and Claude Desktop do not expose a supported global subagent-policy file, so they receive the identical workflow through Prism's MCP server instructions. prism_infer safety boundaries and the host's final verification responsibility are unchanged.

Subscription-Tier Skills Arrive Before the First Host Launch

prism connect now downloads the authoritative Synalux skill manifest and materializes entitled packages in the native ~/.agents/skills directory before the command exits. Codex therefore sees the current skillset on its first launch instead of requiring a second restart. Prism rechecks the same snapshot at MCP startup, session load, and every five minutes—without host lifecycle hooks.

On the first user turn, Prism's native skill, MCP metadata, and managed host instructions request one session_bootstrap({}) call. Prism then uses the dashboard's developer name, Auto-Load Projects, and quick, standard, or deep setting. The response stays focused on greeting and session state because tier skills are already present in the host's native skill directory.

Hook-free MCP can provide and prioritize that ready-to-display block, but the host model still owns the final assistant message and may summarize it. Prism does not claim a deterministic verbatim greeting on third-party chat surfaces; that would require a host lifecycle hook, launcher, extension, or Prism-owned panel. Context loading itself remains complete even when a host shortens the visible reply.

Free accounts receive only the public hook-free prism-startup package; the MCP server still supplies a compact, non-proprietary safety and evidence contract. Authenticated paid accounts receive the protected behavioral and engineering packages plus the current subscribed routing set. The paid evidence-first-protocol keeps ordinary coding lightweight: one correlated reproduction is enough to begin an edit, while strict acceptance starts only before a completion claim, push, or release and inspects only the exact artifacts used as proof. Upgrades install newly entitled packages; verified downgrades remove only Prism-owned packages while preserving local skills and locally modified conflicts.

When upgrading an older Claude Code installation, prism connect removes only the exact Prism-owned startup, skill-sync, handoff, and drift hook actions from the legacy bootstrap. It also removes the recognized legacy Prism startup sections from ~/CLAUDE.md, preserves every other instruction, and installs a small ownership-marked native block that selects session_bootstrap({}) on the first turn. User hooks, custom instruction sections, and near matches remain untouched; native skills and server-side reminders preserve those Prism features without host lifecycle hooks. Because hosts expose no native session-end callback, handoff at shutdown is instruction-driven rather than a guaranteed lifecycle event.

After Claude Code's native user registration succeeds, the same default or --refresh command checks the nearest .mcp.json from the current directory through the home directory. It removes only the exact legacy prism-mcp entry { "command": "npx", "args": ["-y", "prism-mcp-server"] } that would otherwise shadow the user registration. Custom Prism entries and their additional fields, plus unrelated servers, are preserved; malformed files fail loud without changes. --dry-run reports the recognized migration without changing the file.

What's New in v20.2.1

Subscription-Aware Memory Storage

prism connect now carries an explicit PRISM_STORAGE=auto|local|synalux|supabase into every managed host registration and rejects invalid values before changing a config file. In auto, a portal-confirmed free tier uses local SQLite, while Standard, Advanced, and Enterprise use Synalux cloud memory. If entitlement resolution is unavailable, Prism fails closed instead of splitting history across backends. Storage remains independent of local-first model routing.

What's New in v20.2.0

One Command Connects Every Supported Host

Install Prism globally and run prism connect. It detects Claude Code, Claude Desktop on macOS, Windows, and Linux (beta), Cursor, Gemini CLI, and Codex, then safely registers the server from the installed package. Existing custom entries are untouched; --dry-run previews changes and --refresh updates only Prism-managed entries.

What's New in v20.1.0

Every Inference Outcome Is Now Observable

prism_infer gains a failure contract: pass escalation: "report" and every call returns a structured gate_outcome — success, degraded (gate-failed output served anyway, explicitly flagged), or refused (typed, with reason, instead of a thrown error). Degraded output can no longer serve silently.

Big Prompts Work Locally

Prompts over 4000 chars were blanket-refused when cloud was off. Now the full text gets a deterministic reserved-keyword scan plus a head+middle+tail excerpt classification — clean oversize prompts serve locally with a distinct UNCERTAIN_LENGTH audit marker. Clinical/reserved handling is unchanged (and its keyword floor got stronger).

No More Silent Truncation

Tier context limits now match the live Modelfiles (27b/9b are 4096-token models; 4b/2b are 32768 — the old table had it backwards). Tiers that can't hold your prompt are skipped with a visible ctx_insufficient reason; if nothing fits, you get the full prompt on cloud or a loud error — never an answer computed from a silently-clipped prompt.

Know Which Plan You're Actually Running Under

Entitlements carry a source field: portal (real), unconfigured (free by design), or fallback_free (portal unreachable — free limits ASSUMED). Pass strict_entitlements: true to fail loud instead of running degraded.

What's New in v20.0.8

verify_behavior Works Again

The verify_behavior tool crashed on every call (-32602 expected object, received string) — the handler returned a bare string instead of an MCP CallToolResult object. Fixed, with contract + fail-closed regression tests so the safety gate can never silently break again. If you're on 20.0.6/20.0.7, update.

From v20.0.7: Reserved-Content Safety, Skills Auth, Delegation Metrics

Reserved clinical content is now Claude-or-refuse (never served by a smaller model than the one that refused it), skill delivery gained a JWT auth fallback (paid-tier skills now reach machines using only PRISM_SYNALUX_API_KEY), and every prism_infer call is recorded in a persistent infer_metrics ledger. Full details in CHANGELOG.md.

What's New in v20.0.5

Local-First Delegation — 15 Categories, Measured Rate

The local-inference-first skill covers 15 hard-trigger categories (code gen, regex, format conversion, summarization, documentation, factual lookup, classification, shell commands, config gen, and more). Pasted code blocks now trigger delegation regardless of question phrasing. Measured delegation rate: 30-35% on engineering sessions, 40-60% on transform/content sessions. Rate depends on prompt mix, not the skill — the instruments now self-validate with nonDelegatedCount to prevent curated-set tautologies.

Think-Only Retry (v20.0.4)

Qwen 3.5 models (9B/27B) with thinking enabled could burn all tokens on <think> blocks and return empty content, causing a cascade to 4B. Now detects think-only responses and retries the same tier with thinking disabled — preserving model quality instead of falling to a smaller model.

What's New in v20.0.3

Layer 1 Cold-Model Resilience

The reserved-category classifier now retries once with a longer timeout on cold-model failure, then falls back to a deterministic keyword backstop before refusing. Over-length prompts (>4K chars) are classified as UNCERTAIN before reaching the classifier — prompt padding can no longer force the ERROR branch. This eliminates the cold-start refusal problem without weakening the safety gate.

Keyword Backstop for Reserved Content

When the LLM classifier fails (timeout, injection, resource pressure), a deterministic regex floor catches reserved vocabulary (restraint, seclusion, self-harm, suicide, overdose, crisis de-escalation, etc.) including inflected and verb forms. Blocks prompt-padding and classifier-injection attacks on the ERROR path.

Single-Source Safety Text

The safety statement in the MCP server instructions field now imports from boundaries.ts — one source of truth instead of two hand-maintained copies. Boundaries version bumped to v3 with an explicit delivery decision documented in code.

Reserved-Category Safety Gate — All Tiers (v20.0.2)

The Layer 1 semantic classifier now runs for every user, not just paid tiers. Reserved clinical content is refused on free tier when cloud is unavailable — fail-closed.

Ledger Dedup (v20.0.2)

session_save_ledger deduplicates identical entries within a 5-minute window.

Evidence Script (v20.0.2)

scripts/generate-evidence.sh regenerates all 5 evidence files with built-in assertions. Run bash scripts/generate-evidence.sh to verify the full pipeline.

What's New in v20.0.0

License: AGPL-3.0 → Apache-2.0

Prism MCP is now Apache-2.0. The thin-client architecture means all proprietary value (skill resolution, tier gating, billing, cloud inference) lives server-side — the open client carries no moat to protect. Apache-2.0 removes the enterprise adoption friction that AGPL caused.

Thin Client Architecture

Skill routing, budget management, and content resolution have moved server-side to the Synalux portal. The MCP client is now a thin API caller — simpler, smaller, and portable across any host (Claude Code, Gemini, Cursor, autonomous scripts). Offline fallback reads the last successful response from local SQLite.

Clean-Room Voyage AI Adapter

The Voyage AI embedding adapter was independently reimplemented from the Voyage API docs to ensure 100% project-owned copyright. Default model updated to voyage-3.5. See PROVENANCE.md for details.

Server-Side Drift Detection

Session drift detection (GATE 5) no longer requires Claude Code hooks. The timer runs server-side per conversation, piggybacked on every MCP tool response. Works for any host.

CLA Requirement

External contributions now require signing the Individual CLA. The CLA check is merge-blocking on the main branch.

Quickstart

The free tier needs no account, no API key, and no cloud. Install Prism, then register it with every supported MCP host already installed on your machine:

npm install --global prism-mcp-server
prism connect

prism connect detects Claude Code, Claude Desktop (macOS/Windows/Linux), Cursor, Gemini CLI, and Codex. Use prism connect --all to target all five, --host <name> for one host, or --dry-run to preview the files that would change. Existing prism and prism-mcp entries are never overwritten by default. --refresh updates only an entry previously created by Prism; custom entries remain untouched. For Claude Code, both the default command and --refresh also remove the exact legacy project-scoped npx -y prism-mcp-server entry from the effective ancestor .mcp.json after the native user registration succeeds. No custom or near-match project entry is changed. Close the target MCP hosts before a non-dry-run registration so they cannot edit their configuration at the same time.

The same connection installs the local-first orchestration contract:

Host Managed containment
Codex features.multi_agent=false; a 2-thread, depth-1 Terra/low fallback profile is retained for explicit re-enable
Gemini CLI experimental.enableAgents=false
Claude Code CLAUDE_CODE_SUBAGENT_MODEL=sonnet; managed instructions reserve it for last-resort fallback
Cursor Canonical policy delivered through MCP initialize instructions
Claude Desktop Canonical policy delivered through MCP initialize instructions

All five receive PRISM_AGENT_POLICY=local-first in their managed Prism MCP entry. Routine tasks use the RAM-aware local worker; native/background fan-out is not the default workflow. session_task_route supplies a complexity hint; prism_infer remains the single owner of model and thinking selection and can choose 27B when its viability gates support it.

Set PRISM_STORAGE before running prism connect to preserve an explicit storage choice in the generated host entries. This does not change local-model routing; Synalux cloud storage separately requires an active cloud-memory entitlement.

Codex registration preserves unrelated ~/.codex/config.toml content, appends only the marked Prism MCP block, and updates only the documented local-first feature/agent keys. CODEX_HOME is respected when set and must already exist, matching Codex's own contract. Restart Codex CLI, the IDE extension, or the ChatGPT desktop app after connecting.

Restart the connected host and your agent now has memory backed by a local SQLite database (~/.prism-mcp/data.db). See IDE setup for manual configuration and host-specific paths.

Optional — local model fleet for offline tool-routing. Pull whichever fits your hardware:

ollama pull dcostenco/prism-coder:2b    # 3.3 GB  · on-device / lowest RAM · sees images (100% on our routing suite)
ollama pull dcostenco/prism-coder:4b    # 3.5 GB  · verifier · sees images (100%)
ollama pull dcostenco/prism-coder:9b    # 6.7 GB  · default router · sees images (95.7%, reasons before answering)
ollama pull dcostenco/prism-coder:27b   # 16.8 GB · complex code / quality · text only (100%)

Prism detects both the namespaced (dcostenco/prism-coder:9b) and bare (prism-coder:9b) Ollama tags automatically.

The 2b/4b/9b tiers carry a vision tower and accept screenshots through prism_infer({ images: [...] }) — pass absolute paths or base64. Image requests are refused rather than answered blind when no tier (or the Layer 1 safety classifier) can actually see the image, so a text-only model is never handed a prompt about a screenshot it never received. The 27b is text only.

What it does

Your AI agent forgets everything between sessions. Prism fixes that — and adds verification, drift detection, and multi-agent coordination on top.

Mind Palace — persistent memory that survives across sessions

Every conversation feeds a persistent store. The next session loads the right context automatically — no re-explaining.

The dashboard shows your current project state, pending TODOs, intent