Run AI on your own PC.

Start with typed conversation

The established Windows CUDA route requires a compatible NVIDIA GPU. The guided Recommended setup uses Gemma 4 12B QAT Q4_0, a default 128K conversation context and the pinned llama.cpp b10964 CUDA runtime. Model, runtime, optional vision projector and speech packs are separate from the PCS download. Keep them outside the application folder so a fresh PCS installation can reuse them.

1. Choose Set up PCS on the welcome screen, then select Local on this PC → Open Local model & files in its Connection step. For an existing setup, use Settings → Connections.
2. In Local model & files, choose Download to folder… or Use existing files… for the model and matching llama.cpp CUDA runtime. Wait for verification.
3. Continue through Sharing, optional Voice and Screen, then Review and Finish setup. When changing an existing connection in Settings instead, review Memory & sharing and choose Save settings.
4. Use microphone starts off. Type a short message and press Enter or Start & send ↗, or choose Start PCS before typing.

Typed conversation needs no cloud AI key or speech pack. Hardware needs depend on the model, context window, vision and any GPU speech components. CPU semantic search is a separate feature; it does not enable CPU conversation. The Radeon, Linux and Mac previews have their own narrower limits.

Recommended and compatible Custom GGUF templates receive PCS’s thinking choice: off for direct ordinary answers, with thinking available for tool rounds and selected helper work. A template without that switch ignores it. This does not enlarge the reply allowance; complex replies can still reach their limits.

How long conversations are handled

Under Settings → Connections → Local model & files, choose a conversation context of 8K, 16K, 32K, 64K, 128K or 256K for Recommended on CUDA or Custom GGUF. Choose FP16 or Q8_0 KV cache. FP16 is the default; Q8_0 can reduce cache memory use, with a possible speed tradeoff. The model, available GPU memory and enabled features determine what fits.

Long Local sessions automatically compact their working context at about 80,000 prompt tokens, or earlier when another limit requires it. Recent complete exchanges remain in the prompt. Older exact wording stays searchable for that session, including conversation awaiting memory processing; saved originals remain unchanged. This temporary archive ends when the session closes. A notice confirms compaction, and the first reply afterward may take longer while the model rebuilds its prompt cache.

Unchanged prompt prefixes can be reused on CUDA. Active replies take priority over eligible background memory formation, which can retry later. These changes manage repeated work; actual speed and long-session answer quality depend on your setup. A large context window does not guarantee that every detail will be recalled.

How Local chooses memories and screen input

Local checks whether your question needs saved context before composing its reply. A question such as “What did we decide about my project?” can initiate a lookup without a separate instruction to use memory. It asks for focused evidence, with activity scope, sharing permissions, current sources and required corrections still checked.

When Vision is enabled, Local includes an available current frame only for a relevant visual question or follow-up. A personal identity question should use permitted saved evidence; a question about the character you are playing can need the screen instead. A question can need both. This controls which image reaches the reply; screen capture has separate controls.

The optional memory hints setting is separate and is not required for this Local check. The decision adds some processing time, and models can still misjudge relevance or answer incorrectly. A blocked or failed lookup does not establish that nothing was saved; inspect the PCS lookup notice and important source material.

Custom GGUF and saved setups

Saved setups make it easier to return to a compatible Local model without rebuilding its configuration. They use the same PCS vault; changing the model does not replace your saved memories and notes. Review the connection-switching steps.

Experimental model support · Custom GGUF lets you select model and runtime files, choose a context window up to 256K, and run Check compatibility. A selected vision projector also needs checking. Checks cover text, tools, structured output and selected vision; they do not guarantee every architecture, long-session stability or model quality. Choose a window supported by your model and hardware; a smaller context may not fit every enabled tool.

Custom GGUF now uses llama.cpp’s automatic fitting: weights that do not fit available GPU memory can remain in system RAM, with headroom reserved for other work and a selected projector. RAM offload can slow prompt processing. The Recommended model keeps its tested GPU placement; the Vulkan preview does not support Custom GGUF. The tested Gemma 4 26B-A4B Custom model still crashed after long prompts, despite passing short compatibility checks. This is not a generally qualified replacement for the Recommended 12B.

Under Saved custom setups, choose New custom profile, name it, and Save current setup. Keep up to eight named setups, including model, runtime, projector, context, listening and voice choices. Enable Show in Conversation picker to switch while stopped.

Existing Custom GGUF setups keep the window you chose. A Recommended setup saved before 0.9.107 runs at 128K; until Settings are saved again, its automatic cache choice remains Q8_0 beside GPU listening or voice, otherwise FP16. Helper jobs such as Prepared Notes retain their own context limits.

Use profile selects a saved setup for the next Start. Save unfinished Settings edits before switching. Profiles leave API keys and sharing controls in Settings. Review which provider will receive original excerpts if personal memory sharing is enabled. Remove profile keeps its files and current settings.

Control how much the AI speaks on its own

Before Start, choose Follow my lead, Think with me or Take the lead in Conversation pace. These choices guide how PCS develops your thread. Quiet for now pauses optional suggestions, and Stay here keeps attention on the current direction.

The factory Companion direction now asks for calmer replies, proportionate praise and fewer routine follow-up questions. Recognized factory values can update; custom directions are preserved. This does not reset your initiative, pace or PCS Perspective settings.

For Local only, Allow brief quiet-moment openings in Local is a separate checkbox that is off by default and is not restored from saved Settings. Select it before Start and choose Take the lead to invite a brief contribution during a pause. Other conversation connections do not offer this Local-only opening control. Their ordinary pace controls remain available; separately enabled PCS Perspective can explore in the background on supported cloud connections too.

An opening needs a current idea, at least 45 seconds of quiet, the visible and focused Conversation page, an empty composer, quiet playback and no other work. Attempts stay at least three minutes apart, with at most one per complete human input and two within a fifteen-minute invitation. Typing, leaving or unfocusing the page, Quiet, Stay here and Stop cancel a pending opening.

Preparing an idea can add one Local request before your ordinary reply, making that reply slower. The model may choose no idea. An opening uses no tools and is not saved as a conversation source or learned memory. Only an opening confirmed displayed by the browser joins the Local session context for your next reply; that confirmation does not establish that you read or heard it.

Quiet openings remain experimental. Behavior and usefulness can differ by model, and voice or long-session quality is not established by this interface illustration. The app’s Help and docs/CONVERSATION_CARRYING.md describe the full limits.

PCS Perspective is separate from these Local quiet openings. Independent Perspective exploration keeps microphone input on, and anything you say ends the running step. Only a requested focused pause or an Explore now step with that permission turns input off; use Return to conversation because spoken interruption is unavailable then. A pause PCS requested returns its result to the conversation. It uses the displayed Preparation model, can make extra requests and does not independently browse or read Materials. Review Perspective controls and sharing.

Search by meaning, on CPU

The optional Nomic search model finds related wording in memories and permitted conversation passages without cloud embeddings. Open Settings → Optional tools → Search by meaning, select both the model and CPU search runtime, enable Use local semantic memory search, save and Start.

The temporary index stays in RAM and clears at Lock. Ordinary word search remains available while it builds or if the search model fails. This search is English-focused and can miss relevant passages; it complements inspection and exact wording searches.

Only Google and OpenAI supply PCS’s cloud embedding-index maintenance. Meta, Claude, ChatGPT and OpenRouter can use word/navigation recall and this separately configured local index. For cloud search indexes, Search capacity… shows space use and offers a separate review to retire an old provider’s index. Removing search vectors preserves saved memory wording and does not start a rebuild.

Set up speech recognition and spoken replies

Open Settings → Optional tools → Local listening & voice. For microphone input, select one Offline English listening pack: CPU or NVIDIA GPU. Local, Meta, Claude, ChatGPT and OpenRouter microphone conversations use these local speech choices. The cloud text connections receive recognized conversation text, with listening and reply speech handled locally. Then choose a reply engine:

• Windows: an installed desktop voice.
• Kokoro · CPU: the matching CPU voice pack.
• Kokoro · NVIDIA GPU: the matching GPU voice pack and available GPU resources.

Kokoro remains the recommended local voice for everyday use. Typed Local, Meta, Claude, ChatGPT or OpenRouter conversation needs no listening or voice pack. Choosing a cloud text connection does not download packs or require a Local conversation model. Read along uses an available local browser/OS voice, separately from this conversation voice. Reading alone does not send a source to an AI.

Save, enable Use microphone, Start, and allow browser microphone access. Local recordings allow up to 60 seconds, warn at 50 seconds and end after 1.5 seconds of continuous silence. A recording that reaches the limit is processed once as partial speech; Calendar, Notebook, Remember and Practice changes still require a complete turn. Meta, Claude, ChatGPT and OpenRouter use 30-second local recordings with a warning near 22 seconds and a 0.8-second silence pause.

Listening pauses during recognition, processing and replies. GPU listening stays loaded during the session. If speech fails, text replies can continue; use a fresh Start after changing the voice setup.

Start prepares the selected engines; GPU voice can take around a minute on its first preparation. PCS records separate startup and recognition timings in Bug report without retaining audio or recognized words. These measurements help diagnose delays; they do not establish faster startup or more accurate recognition on your hardware.

Share your screen

The setup guide’s Screen step helps you connect OBS and select a scene. For Local Vision, choose a scene and matching Local projector under Settings → Optional tools → Vision. Preview the scene before sharing. Local can attach an available permitted frame when the question or its follow-up needs visual context; it does not send continuous video to the reply model. See how Local chooses reply context.

The default maximum Vision image width is 1920 pixels, preserving its proportions. Existing saved width choices stay in place. A larger permitted image can use more processing and context; review the scene itself for private content.

Optional online lookup and preparation

Local conversation and memory have no cloud AI fallback. For public web search, enable Local online lookup under Settings → Optional tools → Web & study research and add its separate Brave Search API key. Queries and public-page requests leave the PC; the selected model interprets excerpts locally. Keep private details out of queries.

Enabled public lookup has its own permission and remains available during manual activity scopes without an extra activity approval. Automatically follow my current activity and its classification tools are removed; ordinary saved context no longer waits for model classification. Manual details and known progress can still guide relevant story answers, with spoiler avoidance best effort. Saved-memory permissions remain separate.

With saving enabled, Local Study fetches Wikipedia with your online-sharing permission and writes retained Study notes with the active Local model; it does not need a Brave key. Saving a Local research note means the vault needs PCS 0.9.126 or newer. Materials preparation and notebook outlines can use a supported Local or explicitly selected cloud helper. Review the provider and source scope before approving a portion or bounded Prepare book sequence; a Local conversation does not make a separately selected cloud helper local. Local preparation can use up to 80,000 source characters per portion when the complete request fits, with a checkpoint after each portion and a new review on Resume preparation. The Recommended preparation helper keeps a 32K context independently of the conversation setting. See Prepared Notes and storage and sharing.

You can also review a UTF-8 .txt or .md attachment on the welcome screen and choose Save reviewed text in Materials while stopped. This copies exact text to an ordinary, unencrypted local file outside the vault, without an AI request. Back up Materials separately. Attachment choices and limits

Jev is an optional cloud service

Settings → Optional tools → Jev · help choose the next direction can use TypeSafe to choose among proposed directions when carrying is enabled. It starts off, needs its own API key and review, and requires a separate cloud permission with Local. The ordinary conversation model proposes ideas and writes the reply; Jev selects one idea or none. It does not write memories, perform research or start a conversation during silence. Jev does not enable conversation carrying or quiet openings.

The reviewed request can contain your latest complete statement, the current carrying mode and proposed directions. A proposal can reflect context already shared with the conversation model. An optional recent-conversation setting adds up to four eligible whole messages, bounded to 6,000 characters and ten minutes; unconfirmed or no-longer-permitted messages are omitted. Review these inputs before allowing them to leave the PC.

If you choose to remember credentials and save settings, PCS can restore Jev only for that exact reviewed connection and sharing choice. A changed key or unreviewed setup needs another review; turning Jev off removes that setup’s approval. Without remembered credentials, its key and review last only on the open page.

Library exploration is a separate Jev trial. Select a saved record in Continuity → Library, then Explore connections. Review the exact saved excerpts and required corrections before allowing a TypeSafe request. The conversation Jev switch can stay off; its approval does not authorize this saved-record sharing. Results are provisional suggestions, with no automatic memory or Map changes. Review the Library workflow.

Jev requests may add latency and provider cost. These optional trials do not establish better replies or cheaper conversations. Keep conversation Jev and online lookup off, and avoid Wikipedia Study, Library exploration and cloud preparation requests, when you want the Local workflow to stay offline.

Radeon, Linux and Mac are separate previews

The Windows Radeon Vulkan text preview uses the recommended pinned model with a 16K context, CPU speech and CPU search. It requires an explicitly selected compatible discrete AMD adapter; integrated/APU devices are outside this preview. Custom models, vision, larger contexts and GPU speech are not qualified on this route. Real card/driver inference and performance still need testing; this is not an established Radeon support guarantee.

The current 0.9.154 download is Windows x64. Earlier experimental builds are available on the release page tagged 0.9.114: Linux x64 0.9.115 and macOS Apple silicon 0.9.114 Experimental 1. Matching sources are on the same release page. These are archived experiments, not 0.9.154 ports. Follow each build’s included platform instructions; their features and verification differ from the current Windows package. Windows checks do not establish Linux or Mac device acceptance.

Stop, skip and check the result

Stop thinking cancels a Local answer, pending lookup and optional Jev selection after cancellation settles. Skip speech clears playback and unfinished speech synthesis. These controls do not undo already-saved actions; a save already committing must settle with its receipt. Stop PCS ends the session and unloads Local components.

The Local reply stopped… notice fades after six seconds, leaving your draft and conversation open. With reduced motion enabled, it dismisses without animation.

If startup fails, read the stage and check verified model/runtime paths and GPU support first. Speech downloads will not fix a model error. If a save is claimed, check the separate PCS save result and the record. The memory guide explains saving health, corrections and recovery.

Before upgrading, keep an encrypted backup and the old application folder. Restore into a fresh installation before creating a vault there; original Materials and Local model packs need their own copies. Follow the backup, restore and reader-version steps.