From a question to a decision you can defend.
ChatCatharsis is intentionally more deliberate than ordinary chat. Every control exists to preserve a difference, declare an authority boundary, or produce a receipt.
One complete governed run
- Open an archive.Your research body is saved to the device vault first; export important work whenever you want a portable copy.
- Declare the frame.Name the archive, review the directive, and confirm the three council roles, providers, and reasoning depth.
- Ask one bounded question.Click Prepare Council to review the question, member instructions, compiler, and mining limits. Authorize the prepared run to begin. Changing the setup requires a fresh preparation.
- Watch the Council work.First consideration, Disagreement Delta, reconsideration, separate member mining, then one synthesis run automatically. A failed or cancelled stage stops the flow and retains completed evidence.
- Decide only when you mean to.Read the final synthesis and Remaining Disagreements. Each member mine is saved separately on this device and can be exported. Synthesis remains a proposal until a separate human ratification action.
Automatic run readiness
The automatic flow is the default setup. Live recursive mining supports the configured OpenAI, Anthropic, Google Gemini and Groq models with active workspace BYOK connections. The compiler also requires BYOK. Published standard prices were reviewed October 7, 2026 and expire for mining admission November 6, 2026 unless reviewed again. Unsupported providers, expired prices, or unavailable credentials stop readiness checks before first-answer calls. The mixed OpenAI/Claude/Gemini Council of Nine is supported. Ollama, Mistral, DeepSeek, xAI, Perplexity and Cohere mining remain unavailable until their bounded adapters and prices are reviewed.
The mining reservation ceiling is $1 per member, including all child calls; response and synthesis usage is additional. The preparation shows the selected Council ceiling. These reservations are not PayMAXAIOS charges or a guaranteed invoice.
Simulation demonstrates the whole flow without paid calls. Manual review mode remains available for separate stage actions. PromptCatharsis prompt generation and KBaios retrieval are pending integrations; preparation currently previews the actual instructions assembled here.
Mined results are proposals with textual citations. Separate model artifacts are not independent corroboration, and mining quality improvement has not been established by a live comparison. Export important results because browser data can be cleared.
Device vault, optional backup, export, and deletion
An archive holds one research thread and its messages, preserved evidence, decisions, local ledger cache, receipts, and accepted sources. The research body is written to ChatCatharsis's browser-managed device vault first rather than requiring a cloud-storage subscription.
Device-local custody is a core Free Forever capability. It does not require an account. Browser storage can still be removed when site data is cleared, so use Protect device storage when offered and export important work. A free account adds shared live-model access, teams, BYOK connections, and an optional cloud-backup switch for cross-device recovery.
Turning cloud backup off stops new automatic full-workspace and source-file backups; it does not delete an older backup. If this device vault is empty, a signed-in account backup can be restored into it. A small digest-oriented account ledger may still be retained to support receipts, usage, and governed account operations without duplicating the full research body.
Delete selected appears only after at least one archive is selected. Shift-click another archive to select the entire range, review the named selection, and use the export-first option before typed confirmation. Deletion removes the selected device-vault research. If cloud backup is enabled, the matching backed-up research is also removed; billing and provider-usage authority records remain separate.
Write a prompt that exposes a decision
Ask for a comparison, boundary, failure path, or next action. Include the outcome you need, the constraints that matter, and what would count as useful evidence.
Prefer: “Which option should we test first, and why?” Over: “Tell me about these options.” The first creates a decision surface; the second invites a general essay.
Simulation, shared free inference, and BYOK
Simulation teaches the complete workflow without making provider calls. It can be used with the device vault without an account, and its outputs remain plainly labeled as simulated.
Shared live mode uses only providers ChatCatharsis has explicitly approved for shared free-model inference. A signed-in Free or Complete workspace may use that shared capacity without a ChatCatharsis per-run charge, subject to provider availability, rate limits, and fair-use protections. The service does not silently fall back to operator-paid providers.
BYOK lets a workspace use a specific supported provider account. Free workspaces may connect three providers; Complete workspaces may connect all supported providers. The connected key is verified, encrypted in server-side custody, never returned to the browser, and never written into archives or receipts. Any provider charges and quotas belong to the connected provider account.
Current shared-provider policy: Groq is the default shared provider when configured. Google Gemini shared use is separately gated because its data-use posture depends on how the operator account is configured. Other paid-capable operator credentials are not shared fallbacks.
OpenAI BYOK cost routing and receipts
OpenAI is not an automatic shared fallback. When a workspace deliberately connects OpenAI, reasoning depth is cost-routed instead of sending every request to the most expensive model. Low uses GPT-5.6 Luna, medium uses GPT-5.6 Terra, and high uses GPT-5.6 Sol. Output ceilings also rise with the selected depth.
For live OpenAI responses, the provider resolution records reported input tokens, cached-input tokens, output tokens, the applicable published rates, and an estimated USD cost. That estimate is preserved in the ledger and decision receipt; it is an audit aid, not a replacement for the provider invoice.
Default safety gate: the server rejects an OpenAI request when its conservative preflight maximum exceeds $0.50. The operator may set a lower ceiling with OPENAI_MAX_ESTIMATED_REQUEST_USD.
The master directive
The master directive defines the rules shared by every candidate. It freezes when you send the prompt so the receipt can show the exact instruction used. Declare method, constraints, and authority boundaries—do not hide the desired answer inside it.
Compiler and Expert Council
The Compiler compares the three independent Round 1 candidates and creates the Disagreement Map. In the authorized automatic flow, the compiler prepares one synthesis after all member mines have been saved. Remaining Disagreements identifies surviving differences and what could resolve them. It never makes the human decision.
Distinct roles beat duplicate voices. Use logic testing, counterpoint, and an operational or domain perspective so the DMap has meaningful differences to expose.
Read three candidate answers independently
Round 1 candidates answer before seeing one another or the DMap. They remain separately visible so a smooth synthesis cannot silently erase their differences.
Use the Disagreement Map, then observe movement
The DMap runs automatically only after all three valid Round 1 responses exist. It separates consensus, conflict, outliers, and risk without selecting a winner.
During automatic reconsideration, every model receives the original question, its own Round 1 answer, and the same DMap—never the peers' full answers. Each valid Round 2 response must declare whether it held, revised, resolved, remained disputed, or became uncertain, and explain why.
Keep evidence choices separate from human authority
Saving Round 1 or Round 2 evidence is not ratification. Authorizing the automatic flow includes synthesis after complete Round 2 and all saved member mines. In manual review mode, synthesis requires a separate request. The synthesis remains non-authoritative until you separately ratify it; you may also defer or reject it.
Verify, export, and preserve receipts
The local research record binds the prompt, directive, candidate readings, DMap, Round 2 movement evidence when present, provider resolution, human decision, lineage, and digests. The account ledger can preserve compact digest evidence without needing the full research body. Verify recomputes a digest; Export creates a portable user-owned copy. A valid receipt proves what the system recorded—not that every underlying claim is true.
Teams keep roles and custody explicit
Create a team workspace from Account, invite a verified email address, and assign editor or viewer access. Owners manage provider connections; members can use an available connection without seeing its key. Device-local research does not become shared merely because a user joins a team; cloud-backed team records remain an explicit account capability.
ChatVaultAI desktop vault contents remain local. Team membership and shared ChatCatharsis records do not upload a member's private desktop vault.
Choose your model access
Start with shared cloud routing where included access is available, or connect your provider API keys. Ollama Cloud uses a key and needs no local installation. Provider limits and charges still apply. Automatic Council member mining currently requires workspace API keys for cloud seats; included access alone does not authorize that stage.
Local Ollama is an optional advanced connection for people with installed models and suitable hardware. It supplies only that person’s browser session. Local execution does not establish better quality or faster answers.
Connect Ollama on your computer
Ollama Local is a separate opt-in browser provider. It is never added to shared free routing, and your computer does not supply other users. Ollama Cloud continues to use its cloud key. Local execution is available in the main research workspace's live automatic Council flow.
- Install a completion model in Ollama yourself. ChatCatharsis does not download or install models.
- Quit Ollama. Set OLLAMA_NO_CLOUD=1, retain the default loopback binding (127.0.0.1:11434), and set OLLAMA_ORIGINS to the exact origin you use for ChatCatharsis, including scheme and port. Do not use a wildcard, LAN binding or public tunnel. Restart Ollama. On Windows, set these in your account environment variables and restart the app.
- In the research workspace's Ollama Local panel, authorize access and click Discover local models. Allow a browser local-network prompt if you choose. Select the installed model and supported context size, then run the connection self-test.
- Select Ollama Local for the member or compiler seats you want, enable live automatic Council, and review the prepared run before authorizing. All local seats use the tested model. Reloading requires consent and testing again.
The self-test runs a small real inference; it checks connectivity and reported identity, not answer quality or Council suitability. Cloud stubs and models without local completion metadata are refused. Names and digests are supplied by your local service, not independently attested. Keep cloud features disabled to prevent local-service cloud forwarding.
If discovery or testing fails, check Ollama is running, the exact allowed origin, installed model capabilities, supported context size, and browser permissions. Browser policies vary; direct HTTPS-to-loopback access must be tested in your installed browser. A blocked browser stops the flow. There is no server localhost proxy, tunnel, extension bridge, or automatic cloud fallback.
Local prompts go directly from your browser to Ollama. Local answers can be sent to cloud seats in a mixed Council and follow normal record-storage settings. Context admission and output limits can stop a long run; evidence is never shortened to fit. Saved member mines and checkpoints remain proposals and can be exported through the existing recovery controls.