nicobailon/pi-mcp-adapter v2.18.0

on GitHub

Added

• Added settings.freezeDirectTools to keep direct MCP tool registration stable after initial sync while preserving explicit reconnect refreshes. Thanks @ddfourtwo for PR #254.
• Added best-effort Linux OAuth credential recovery when Pi inherits a revoked session keyring, allowing explicit re-authentication through a fresh keyctl session helper. Thanks @anthod0 for issue #248 and the validation prototype.
• Ranked, paginated MCP tool search: best matches come first in a short page of 12 instead of an unranked dump of every match with full schemas, so the model stops guessing and each search costs a fraction of the tokens. Misses on describe/call now return top-5 "Did you mean" suggestions, letting the model self-correct a typo or missing prefix in the same turn instead of burning a round trip.
• Optional approveTools patterns (global and per-server) add the missing middle tier between "tool runs instantly" and "tool hidden entirely": flag risky tools and Pi asks before running them — Allow once / Allow for session / Deny — across proxy, direct, resource, and iframe-originated calls. Safe tools keep full speed; a deny is a normal result the model adapts to, not a crash.
• Opt-in mcp_script trusted JavaScript MCP scripting turns N-step jobs into one call: loop, filter, and chain tools for tool-restricted subagents where every round trip costs child context. Scripts discover tools with await tools.search({ query }), inspect exact shapes with await tools.describe({ path }), and call them with tools.call(path, args) — no more guessing names from outside the script. A runaway script can never freeze Pi itself: scripts run isolated from the main process and are force-stopped at their time limit, even if stuck in an infinite loop. Result details include a calls trace of every invoked path and outcome, and a bundled mcp-scripting skill teaches the full workflow on demand. Every scripted call still goes through auth, output guarding, and the approval gate.
• HTTP connection failures are now probe-classified into a plain-language diagnosis (for example "endpoint returned HTML (200) — this URL does not appear to speak MCP") instead of an opaque "fetch failed", so setup mistakes are fixed in seconds. Healthy connections are never probed.
• Tool parameters render as compact TypeScript shapes ({ query: string; limit?: number }) in describe and search, replacing multi-line schema dumps — the model reads less and acts sooner, with the previous formatting kept as a fallback for exotic schemas.
• /mcp setup gained curated one-click presets (DeepWiki, Context7, Notion, GitHub, Chrome DevTools): pick, preview the exact config write, confirm — new servers in under a minute with no hand-typed setup.

The ranked search scoring, did-you-mean suggestions, approval patterns, endpoint shape probe, TypeScript-shaped schemas, and codemode design in this release are adapted from Executor by Rhys Sullivan (@RhysSullivan). Thanks Rhys.

Fixed

• Brought MCP Apps UI hosting in line with the current spec: provider HTML now runs in a real sandbox, gets a restrictive default CSP even when the resource omits one, and _meta.ui.visibility is honored so app-only tools stay out of the model tool list while model-only tools cannot be called from the UI.
• MCP Apps UI sessions are now easier to open from Moshi and remote terminals: the local UI server uses Moshi-discoverable low ports, answers preview discovery probes, serves a loopback-only landing shell, prints Moshi/SSH access hints for remote sessions, and fits the host shell better in narrow in-app browsers. UI-submitted model context is now captured as a bounded handoff, wakes the agent like prompts and intents, and remains available through mcp({ action: "ui-messages" }) after the UI closes.