Claira Private
Other languages
Claira Private runs a case entirely on your own model, either a shared connection your organization operates or a local model on your own computer, for one flat monthly price with no tokens.
Claira Private
Claira's standard plans run on Claira's own models, hosted inside your deployment's data residency boundary. For most teams that is all they need. Claira Private is for the cases that need something else: every AI feature in the case runs on your own model, and never on one of Claira's.
Claira Private is a plan for the whole case, chosen when the case is activated, as the alternative to the standard plans. It has one flat monthly price per case, with no tokens and no document limits. See Pricing and billing.
A Claira Private case can reach your model in two ways, and both are included:
| Shared connection | Local model | |
|---|---|---|
| Where the model runs | On infrastructure your organization operates, reachable from the internet | On your own computer, or another machine on your network |
| Who sends the documents to it | Claira's servers | Your browser, directly |
| Who uses it | Everyone on the case | Only you, in this case |
| Who sets it up | Anyone on the case, once for the case, under Team settings | Each user, in their own browser, under Your settings |
| Address it uses | A public https:// address | http://localhost:11434 and similar |
| Best for | A model your firm hosts for the whole review team | A model on a laptop or workstation, where document text must not leave the machine |
Both are for models you run. Neither requires a model trained or tuned for Claira: an off-the-shelf open-weights model served over the OpenAI chat-completions API is exactly what they expect.
Choosing Claira Private
You choose Claira Private when you activate Claira on a case. The activation screen opens on the standard plans, from Starter to Enterprise, with Pro selected. Below them, click the narrow Claira Private row: its details replace the list, with Continue to go on and Back to standard plans to return. Then choose a payment method, as for any other plan. See Case Activation.
Once the case is active, open Settings and set up your model: a shared connection for the whole case, in the Shared connection section under Team settings; a local model for yourself, in the Your model section under Your settings; or both. Your model is also where each person chooses which of the two they use, and it marks the one you use In use only when it can run scans, or Not set up when there is nothing to run on yet. Until one of them is set up, the case has no model to run on, and AI features tell you so rather than running. For a walkthrough of each step, see Set up Claira Private.
Changing plans. A case cannot move between a standard plan and Claira Private while it is active, in either direction. To change, deactivate the case, then activate it again and choose the other option. See Plans and Billing.
What Claira Private includes
On a Claira Private case, every AI feature runs on your model:
- Scans: single reviews, bulk scans run in your browser, and Multi-Code.
- Agent Mode, Insights, the Prompt Generator, Revise with Agent and Case Context from documents. These use the shared connection, or your local model while you have one turned on.
A few things work differently from the standard plans:
- Background processing is not available in Claira Private. Bulk scans run in your browser; keep the window open until the task finishes.
- Scans are text only. Image, audio and video scans are not available, and Scan as stays on Text.
- Bulk tasks and chats get standard names. Claira names them with its standard descriptive naming rather than a model-written title.
- There is no model picker and no automatic model routing. The case's model is the one you connected. Claira never moves work onto one of its own models, and a feature that cannot run on your model says so instead of running somewhere else.
- There are no tokens. No balance, no free daily claims, no token packs and no overage. See Pricing and billing.
Shared connection
A model endpoint your organization hosts on the public internet, connected once for the case. Everyone on the case then works on it: text scans, Agent Mode, Insights and Claira's other AI features are sent to your endpoint. Everything else about your workflow stays the same: your prompts, your connected fields and your results.
Before you start
Your endpoint must:
- Accept requests in the OpenAI chat-completions format: the
/v1/chat/completionsrequest that most model servers and gateways speak, including vLLM, Ollama behind a gateway, Azure OpenAI, OpenAI-compatible proxies, and hosted inference providers. - Be reachable from the internet over HTTPS, with a certificate issued by a publicly trusted certificate authority.
- Authenticate with an API key that Claira sends as a bearer token.
If your model runs on-premises, place it behind a gateway or reverse proxy that exposes it publicly over HTTPS. Claira's servers cannot reach private or internal addresses. See Limits in this release for the full list, and if the model runs on your own machine rather than on shared infrastructure, use a local model instead, which is built for exactly that.
Where to set it up
- Open your case workspace and select the Claira pane.
- Open the Modes menu and choose Settings.
- Under Team settings, open the Shared connection section and click Set up the shared connection (or Edit the shared connection once one is saved).
The shared connection sits under Team settings because it applies to the whole case. In Your model, under Your settings, the Shared connection option shows its status and its Open in Team settings button takes you to this section.
The panel says so, and it is worth repeating: the connection is shared with everyone in the case. Anyone on the case can scan with it, edit it or revoke it.
While the section is closed, it names the connection, or reads Not set up when none is saved. The panel shows the connection status at a glance:
- Not connected: no endpoint is saved for this case yet.
- Connected: an endpoint is saved and responding normally.
- Needs attention: an endpoint is saved, but recent requests failed or your key was rejected. See Errors and recovery.
Every field has an info icon. Hover it for a short explanation and a Learn more link that opens the matching section of this page.
What you need
Display name
The name Claira shows for your model everywhere a model is named: the Shared connection and Your model sections in Settings, the bulk-scan confirmation dialog, and the Bulk Task History report. It is how the connection appears to people on this case.
Where to find it: you choose it. Pick a name your reviewers will recognize, such as Firm Llama 70B or Review model (EU cluster). Up to 80 characters.
Endpoint URL
The base address Claira sends requests to. Claira adds /chat/completions to it for each request, and /models when it looks up the model names your endpoint offers.
Where to find it: in your model server's or gateway's documentation, usually labelled base URL. It typically ends in /v1, for example https://models.example.com/v1. Paste it without a trailing slash and without /chat/completions on the end.
What a good value looks like:
- Starts with
https://: plainhttp://is refused. - Uses a public hostname. Private network addresses (
10.x,192.168.x,172.16.xto172.31.x),localhost, link-local and cloud-metadata addresses, and*.internalhostnames are refused. Those belong to a local model. - Contains no username, password, or
#fragmentinside the address. - Contains no query string, with one exception: Azure OpenAI. Azure addresses a deployment as
https://<resource>.openai.azure.com/openai/deployments/<deployment>and requires an?api-version=parameter, so a URL carrying exactly?api-version=<date>and nothing else is accepted and the version is sent with every request. Any other query string is refused.
Model ID
The exact model name your endpoint expects in each request. Claira sends it as the model value.
Where to find it: if your endpoint supports model discovery, Claira lists the available names as soon as you have entered the Endpoint URL and the API key, before any test. When your endpoint offers a single model, Claira chooses it for you; otherwise, pick one from the list. If your endpoint does not list its models, Claira says so. If the list is empty, or your endpoint offers more than 200 names, or you prefer to type the name, click Type it manually and enter the identifier your provider documents (for example meta-llama/Llama-3.3-70B-Instruct, or an Azure deployment name).
API key
The secret your endpoint requires. Claira sends it as a bearer token with every request.
Where to find it: generate it in your provider's or gateway's console. Use a key dedicated to Claira so you can rotate or revoke it without touching other systems.
How Claira handles it: the key is write-only. Once saved, the field reads •••• saved. Enter a new key to replace it., and nobody can read the key back out of Claira. It is stored encrypted, never written to logs, and only ever sent to the endpoint you configured. To rotate the key, paste the new one and click Test & save. The saved key is reused only for the endpoint host it was saved with: if you change the Endpoint URL to a different host, enter that endpoint's key as well before testing or saving.
Concurrent requests
How many documents Claira may send to your endpoint at the same time during a bulk scan.
What a good value looks like: start at 1, the default. After a successful Test connection, Claira measures how many parallel requests your endpoint sustains with stable latency and offers to apply that number. See Optimizing your connection. Allowed range: 1 to 64.
Context window
The largest request, in tokens, that your model accepts: prompt, document, and response together.
Where to find it: in your model's documentation or your server's launch settings, often called context length or max model length. Enter the real value: the long-context check in the connection test sends a request sized to roughly half of it, so an inflated value makes that check fail. Your endpoint decides at scan time whether a document fits; one it rejects as too long is reported as "content too long" in the scan results, whatever value you enter here. Default: 32,768 tokens.
Max output tokens
The longest response, in tokens, that Claira asks your model to produce.
What a good value looks like: the default of 8,192 suits summaries and coding prompts. Lower it if your endpoint rejects requests that ask for a large output; raise it only if your prompts genuinely need very long answers (a full translation of a long document, for example) and your model supports it.
Residency acknowledgment
A checkbox you must tick before saving. It confirms you understand that document text and prompts sent to your model leave Claira's deployment region and travel to the endpoint you configured, at your direction. Claira's data-residency commitments stop at that endpoint; what happens next is governed by your agreement with whoever hosts it. See Privacy and Security.
Testing the connection
Two buttons run the checks:
- Test connection runs every check, including calibration, without saving anything. Once the model is chosen, one run is enough for a first setup, and later you can use it to measure your endpoint again. Run without a model, it tests the only model your endpoint offers, or asks you to choose one in Model ID when there are several.
- Test & save runs the quick checks (everything except calibration) and saves the connection when they pass. It stays disabled until the required fields and the residency acknowledgment are filled in.
Each check is listed with a PASS, WARN, or FAIL tag, a one-line result, and a Learn more link.
You can run the connection test up to 10 times per hour for a case. Past that, the test is refused with a message asking you to try again later. The saved connection keeps working normally; this limit applies only to running the test itself.
The whole test runs within a time limit of about ten minutes. If your endpoint answers very slowly, the optional checks that could not start before the limit are marked WARN as skipped: the checks that did run are still reported, and a connection still saves when the required ones pass. The Multi-Code check is never skipped this way: it runs before the optional checks, so a slow endpoint cannot leave Multi-Code switched off without telling you why. Run Test connection again when the endpoint is less busy to complete the picture.
| Check | What it verifies | If it warns or fails |
|---|---|---|
| URL | The endpoint URL is well-formed, uses HTTPS, and points to a public address. | FAIL: fix the address as described under Endpoint URL. No other check runs until this one passes. |
| Reachability | Claira can open a secure connection and your endpoint accepts the API key on a minimal request. | FAIL with a key error: the key was rejected. Check for stray spaces or an expired key. FAIL with an unreachable error: check the hostname, the firewall, and the certificate (self-signed certificates are not accepted). |
| Models | Your endpoint answers the model-discovery request. | WARN: discovery is not available. Type the Model ID manually. Saving still works. |
| Completion | A representative review prompt of about 1,200 tokens returns a real answer. Measures latency and tokens per second. | FAIL: the model ID may be wrong, or the model returned nothing. Confirm the ID and try a smaller Max output tokens value. |
| Multi-Code | Three sample requests return well-formed JSON keyed by field aliases. | WARN: Multi-Code is switched off for this connection; single-field prompts still work. A more capable or instruction-tuned model usually fixes this. |
| Streaming | Your endpoint can stream partial answers. | WARN: single-review answers appear all at once instead of word by word. Scans work normally. |
| Long context | A request sized to about half your stated context window (up to about 48,000 tokens) succeeds. | WARN: lower the Context window to what the endpoint actually accepts, otherwise large documents will fail. |
| Calibration | Sends increasing numbers of parallel requests (1, 2, 4, 6, 8, and beyond where your deployment allows) and stops when latency rises to more than two and a half times the single-request latency or an error appears. | Informational: reports the highest parallelism your endpoint sustained. Runs only with Test connection. |
A connection saves when URL, Reachability, and Completion pass. Warnings are recorded with the connection so you can see later which capabilities were switched off and why.
Optimizing your connection
Start at 1. A new connection sends one document at a time. That is slow but safe: it cannot overload an endpoint, and it lets you confirm results before scaling up.
Apply the recommendation. When a Test connection run finds that your endpoint sustains more parallel requests than your current setting, a green banner reads Your endpoint sustained N concurrent requests with stable latency. Click Apply recommended (N) to set Concurrent requests to that value in one click.
The scan speed slider follows your setting. During a live bulk scan the Scan speed slider still runs from Slowest to Fastest, but its scale is now your connection's capacity: Fastest means the full number of concurrent requests you configured, and each step down sends proportionally fewer. Balanced, the default, keeps a little in reserve.
When to raise it: the failed counter stays at zero, results arrive cleanly, and you want the run to finish sooner. Run Test connection again first if your endpoint's hardware or quota changed.
When to lower it: documents fail with "endpoint not responding" or timeout errors, or your endpoint's own dashboard shows it saturating. Claira also protects your endpoint automatically: when errors cluster, it temporarily sends fewer parallel requests than you configured and returns to your setting once responses are healthy again. The panel shows Needs attention when the connection reports Degraded (most recent requests failed) or when your key was rejected; a milder error rate slows the run without changing the status. The state is shared across Claira's servers: a rejected key or a run of failures seen anywhere shows on every refresh until your endpoint answers successfully again (a passing Test connection with your saved key counts), and the health line names the last problem in plain words; the error code itself is in Diagnostics.
What routes to your model
On a Claira Private case, everything that uses a model goes to your endpoint, for everyone on the case who has not turned on a local model:
- Single reviews: the Scan Document button in the Prompt Lab, and single-document Multi-Code.
- Bulk scans: bulk tasks that run in your browser tab. The confirmation dialog shows a Model row naming your model, with a warning underneath if the connection currently needs attention.
- Multi-Code: every field in a Multi-Code pass, single or bulk, provided the connection test confirmed that your model returns well-formed JSON.
- Agent Mode conversations.
- Insights generation and revision.
- Case Context drafted from documents, and updates from Agent Mode. Your model receives each document's extracted text.
- Revise with Agent suggestions.
- The Prompt Generator.
How much each of these can do depends on your model. Agent Mode, for example, needs a model that supports function calling to start scans, propose fields or edit Case Context. Choose a capable instruction-tuned model for the best results.
Nothing is sent to one of Claira's models instead. If the shared connection is missing or revoked, these features tell you that no model is connected and do not run. Bulk tasks and chats are named with Claira's standard descriptive naming, without a model call.
Limits in this release
- Text scans only. Scan as stays on Text. Image, audio, video and Auto scans are not available in Claira Private.
- Live runs only. Bulk scans run in your browser tab; keep the window open until the task finishes. Background processing is not available in Claira Private.
- Publicly reachable HTTPS endpoint with a certificate from a public certificate authority. Self-signed and private-CA certificates are not accepted.
- No IP allow-listing. Claira does not publish fixed outbound addresses in this release, so your endpoint must accept authenticated requests from the internet. Control access with the API key, not a firewall rule.
- No private addresses. Loopback, private-network, link-local, and internal hostnames are refused. Use a local model for a model on your own machine.
- One connection per case. Saving a new endpoint replaces the previous one for that case.
- OpenAI chat-completions format only. Endpoints that speak another request format need a compatible gateway in front of them.
Errors and recovery
Errors from your endpoint are reported with the same codes as other scan errors. See Error Codes for the full reference, including the Claira Private codes. The situations specific to your own model:
Your endpoint is not responding. Documents fail with an "endpoint not responding" message, and during a bulk scan an amber banner appears above the live results once several documents in a row fail this way. If the outage stops the workers, the task is marked Incomplete: fix the endpoint or its network path, confirm with Test connection, then click Resume Task in Bulk Tasks: documents that already have results are not scanned again. See If a scan stops early.
Your API key was rejected. Claira stops sending requests immediately rather than failing every remaining document, the banner says the endpoint rejected the key, and the connection shows Needs attention. Paste a valid key into the API key field and click Test & save; the connection clears as soon as the new key is verified, and you can resume the task.
A document is too large for your model. The document fails with the "content too long" message. Check that Context window matches your model, or split the document.
A scan was refused before it started. A Multi-Code run on a connection whose Multi-Code check did not pass is refused with a message naming the reason. Run the fields as single-field prompts, or connect a model that passes the check.
Diagnostics. On the connected row of the panel, open Diagnostics to see the most recent errors your endpoint returned (time, error code, upstream status, and latency) alongside the current health state. Share that view with whoever runs your endpoint when you contact them.
Revoking the connection
Under Team settings, open the Shared connection section, click Edit the shared connection, and click Revoke on the connected row. Because revoking affects everyone in the case, Claira asks you to confirm first. Once you confirm, Claira stops using the endpoint and its key immediately. Until someone connects a model again, scans and other AI features on the case are refused with a message saying no model is connected, except for users working on their own local model. You can connect the same or a different endpoint again at any time; enter the key again when you do.
Local model
A local model runs Claira on a model on your own computer. Your browser sends each document straight to that model and reads the answer back. For the review itself, the document text never leaves your machine: it does not travel to Claira's servers, and it does not travel to anyone else's.
While your local model is on, Claira's servers make no model calls of any kind for you on that case: not to Claira's models, and not to the shared connection. Every answer you see is produced by your model, in your browser. And if Claira ever cannot confirm whether your local model is on (a momentary service hiccup, say), it refuses the request and tells you so, rather than guessing; see Local model errors.
A local model is yours, and per case. Turning it on affects only your own work in that case; a colleague in the same case keeps using the shared connection. That matters, because a model on a machine is personal: nobody else's review is interrupted by your choice.
What you need on your machine
- A model server on your computer or network: Ollama, LM Studio, vLLM, or llama.cpp. See Supported local servers.
- A model pulled and loaded in that server. A capable instruction-tuned model gives noticeably better review results than a very small one; see What works with a local model for what each feature needs.
- The server configured to accept requests from Claira: one setting per server, described below.
- Chrome's permission to reach your own network, granted once. See The browser permission.
Turning it on
- Start your model server and load the model you want to use.
- Open Settings in the Claira pane, open the Your model section under Your settings, and choose Local model.
- Under Supported local servers, make the setting your server needs: Claira lists it for each server, with the Claira address filled in. Restart the server afterwards. See Supported local servers.
- Click Detect local servers. Claira looks on the usual ports (11434, 1234, 8000 and 8080) and lists whatever answers; click Use this server on a server's row. If nothing answers, click Enter the endpoint manually and type the address in Endpoint URL yourself. Claira treats
http://localhost:11434,http://localhost:11434/andhttp://localhost:11434/v1as the same address. - Choose the Model from the list your server returned, and give it a Display name: what Claira will call it in Settings, the indicator, and task history. Add an API key only if your local server requires one.
- Click Confirm in this browser. If Claira reaches your model through the bridge window, that window shows your model ready to confirm: click Confirm and start there, and Claira carries on.
- Claira runs a short set of checks against the model and shows a PASS, WARN, or FAIL row for each: that the address is reachable, that the model completes a review prompt, whether it returns well-formed JSON for Multi-Code, whether it supports JSON output mode and function calling, whether it streams, and how much it can handle at once. When the checks pass, your local model is turned on.
Claira confirms that scans in this case now run on your model in this browser, and a Local model badge appears in the app chrome from then on. You can close the panel: your local model keeps running, and so does the bridge window if you use one.
If your model's capabilities change (you loaded a larger model, or raised its context window), click Re-run capability checks to measure it again without setting the connection up from scratch.
Supported local servers
Browsers refuse to let a web page talk to a server on your own machine unless that server explicitly says the page is welcome. That setting is called CORS, and each server spells it differently. In every case you are allowlisting the Claira address (the address Claira itself runs at, which looks like https://nuix-ca.claira.to), and not your model's address.
| Server | Usual port | What to configure |
|---|---|---|
| Ollama | 11434 | Set OLLAMA_ORIGINS to the Claira address before starting Ollama, for example OLLAMA_ORIGINS=https://nuix-ca.claira.to ollama serve. By default Ollama accepts browser requests only from 127.0.0.1 and 0.0.0.0, so without this it refuses Claira. On macOS, set it with launchctl setenv OLLAMA_ORIGINS <address> and restart Ollama; on Linux, add it to the systemd service override; on Windows, set it as a user environment variable and restart Ollama. |
| LM Studio | 1234 | Open Server Settings and turn on Enable CORS. If LM Studio runs on a different machine from your browser, also turn on Serve on Local Network. |
| vLLM | 8000 | Usually nothing. vLLM starts with --allowed-origins ['*'], which already accepts Claira. If your deployment narrowed that list, add the Claira address to it. |
| llama.cpp / LocalAI | 8080 | Start the server with CORS enabled and add the Claira address to its allowed origins. If you run it behind a reverse proxy, the proxy is what has to allow the address. |
Any other server that speaks the OpenAI chat-completions API works too: enter its address manually and configure its own CORS setting the same way.
The browser permission
Recent versions of Chrome ask before letting a website reach a server on your own computer or network. The first time Claira contacts your model, Chrome shows a local network permission prompt. Choose Allow. Claira cannot reach your model until you do, and if you dismiss or block the prompt, scans fail with a message saying the browser did not allow access.
To change your answer later, open Chrome's site settings for the Claira address and set the local network permission there.
For IT: pre-granting the permission. Administrators can grant it for everyone with Chrome's LocalNetworkAccessAllowedForUrls policy. Set it to the Claira frontend origin (the address Claira is served from, such as https://nuix-ca.claira.to), and not to the model endpoint. It is the page making the request that the policy names, not the destination. The companion LocalNetworkAccessBlockedForUrls policy will prevent a local model from working if the Claira address is listed there.
Firefox and Safari have no equivalent gate, so no prompt appears in those browsers.
The bridge window
Claira runs inside Nuix Discover as an embedded panel. Browsers do not let an embedded panel reach your own network, however you answer the permission prompt: the restriction is on being embedded, not on the permission. So when it needs to, Claira opens a small separate window at Claira's own address, which is not embedded and is therefore allowed to talk to your model. Every request to your model goes through that window.
- Claira tries the direct route first. In browsers and configurations where the panel itself can reach your model, no bridge window is needed and none opens. The Local model badge tells you which is in use.
- Open it from the Local model panel or the badge with Open bridge. If your browser blocks the pop-up, allow pop-ups for Claira, or use the link Claira offers instead.
- Keep it open while Claira is working. It is where your model runs. Closing it stops Claira from reaching your model: the badge turns amber, and Claira blocks new scans with a panel offering to reopen it rather than failing document after document.
- It is one window, reused. Clicking Open bridge again focuses the window you already have.
- You confirm your model once. When you click Confirm in this browser and Claira needs the bridge window, it fills the window's form with the model you chose, under Confirm your local model, and the Local model dialog asks you to confirm it there. Check it and click Confirm and start: Claira carries on in the dialog by itself. Opened on its own for the first time, the form is headed Confirm your local model too; Change endpoint appears only once an endpoint is confirmed in that window.
The bridge window shows the endpoint address and model it is using, whether the browser has granted the local network permission, whether your model is responding, how many requests it has handled, how many are in progress, and the last error if there was one. It is your evidence that your own model is doing the work.
Confirm and start contacts your model before saving. When you confirm an endpoint in the bridge window, Claira sends it the same model-list request a scan would, from that window. If the server does not answer, refuses the request, or rejects your key, the window says why (the same message a scan would show), records it under Last error, and keeps the previous endpoint (or none) until you fix it. The button shows a pending label while it waits, and a later attempt starts fresh, so an endpoint that was simply not running yet can be confirmed once it is.
What the bridge window will and will not do. It only ever contacts the endpoint you confirmed in that window. The Claira panel may fill in the form for you, but nothing is contacted or stored until you click Confirm and start; an address arriving from anywhere else is ignored, and it re-checks the address on every single request. It never sends your Nuix or Claira sign-in with a request to your model. Your local key, if you set one, is read from that browser and sent only to your model.
What works with a local model
Some features run on your model directly, because your browser already has everything it needs. Others need Claira's servers to prepare the instructions first: Claira's servers build the request, your model answers it in your browser, and Claira's servers make sense of the answer. Either way, the thinking happens on your machine.
Runs on your model directly
| Feature | Notes |
|---|---|
| Single document review | Streamed word by word, as with any other model. |
| Bulk review (live) | The full worker pool, the scan speed slider, the live feed, error annotation, and resume. How many documents run at once starts from what your model measured during setup. |
| Multi-Code (bulk and single) | Available when the Multi-Code check passed for your model. |
| Multi-Code split in the Prompt Lab | Available on your model. |
Runs on your model with Claira preparing the request
Each of these needs something specific from your model. The setup checks tell you in advance whether yours can do it, and Claira says so plainly rather than quietly using a different model.
| Feature | What your model needs | If it cannot |
|---|---|---|
| Prompt Generator | Plain text | Not applicable |
| Insight titles | Plain text | Not applicable |
| Case Context from documents | JSON output mode | Claira asks for JSON in the prompt instead, and reports a clear error if the answer cannot be read |
| Revise with Agent | JSON output mode | Same |
| Insights | A usable context window | Claira reads the documents in batches and merges the findings. If a single document will not fit on its own, Claira says so and asks you to narrow the selection or use a model with a larger window |
| Agent Mode chat | Function calling, for the full experience | The Agent falls back to tool-less chat with a notice that stays on screen: it can answer questions and explain things, but it cannot start scans, propose fields, or edit Case Context |
With a local model, Case Context from documents sends your model each document's extracted text only, never its native file. A document without extracted text is left out, and Claira names it when the context is updated.
Progress is shown while your model works (Your local model is thinking…, Analysing documents 3 of 7, Merging findings, Running tools…) because a local model can take noticeably longer than a cloud one.
Unaffected, because no model is involved
The Prompt Lab, prompt history, pins and comments, prompt templates, Field Migration, Field Create, Connected Fields, reading and editing and approving Case Context, thumbs up and down on a response, usage and billing, and exports all behave exactly as they always do.
Bulk tasks and chat threads still get names automatically, using Claira's standard descriptive naming rather than a model-written title. Insight titles are written by your local model, as listed above.
Not available with a local model
| Feature | What to do instead |
|---|---|
| Background scans | Background processing is not available in Claira Private. Run bulk scans live instead, and keep the window open. |
| Image, audio, video and PDF scans | Your local model receives text only. Scan these documents as Text. |
| Early case assessment | Running an assessment is refused with a message saying so, and nothing is saved. Past assessments stay available to read, download and delete. See Early case assessment. |
The Run as background processing task checkbox and the non-text Scan as options are not offered, so you are never left guessing.
Turning the local model off
- In the Your model section of Settings, click the Shared connection option, or open Local model settings and click Use the shared connection.
- Claira asks Switch to the shared connection? Click Switch to the shared connection.
From then on your work in the case runs on the shared connection again, like everyone else's. If the case has no shared connection yet, AI features tell you that no model is connected until someone sets one up, or until you turn your local model back on.
Because a local model is per user and per case, turning it off in one case does not affect any other case you work on, and never affects a colleague.
Troubleshooting a local model
"No local model server answered on the usual ports." Your server is not running, it does not have its setting yet, or it is on a different port. Confirm the server is up (open its own address in a browser tab) and has the setting listed under Supported local servers, restart it, then click Detect local servers again, or click Enter the endpoint manually.
"Your local model was not confirmed in the bridge window." Claira filled the bridge window's form with your model, but Confirm and start was not clicked there in time, so nothing was turned on. Click Confirm in this browser again, find the bridge window (it may be behind other windows), and click Confirm and start.
"Your local model isn't responding." Nearly always the CORS setting. Check that your server allowlists the Claira address exactly as shown in the Local model panel, including the https:// and with no trailing slash, and restart the server after changing it: most servers read that setting only at startup. If the server runs on another machine, confirm it is listening on the network rather than only on its own loopback address (in LM Studio, that is Serve on Local Network). If this message appears under Confirm and start in the bridge window, nothing was saved: fix the setting, restart the server, and click Confirm and start again.
"Your endpoint answered, but this browser refused to store it." The bridge window keeps the endpoint in that browser's site data, which a private or incognito window (or a site-data block for the Claira address) prevents. Open the bridge in a normal window, or allow site data for Claira, then click Confirm and start again.
"Your browser did not allow Claira to reach your local network." Choose Allow on Chrome's local network prompt, or reset the permission in Chrome's site settings for the Claira address. On a managed device, ask IT to add the Claira address to LocalNetworkAccessAllowedForUrls. See The browser permission.
"The bridge window is closed or not responding." Click Open bridge and leave the window open while you work. If your browser blocked the pop-up, allow pop-ups for Claira. Scans stop cleanly when the bridge closes mid-run: fix it and click Resume Task in Bulk Tasks, and documents that already have results are not scanned again.
Multi-Code is unavailable. Your model did not return well-formed JSON during the checks. Load a more capable instruction-tuned model, raise its output limit if it was small, and click Re-run capability checks. In the meantime, single-field prompts work normally.
The Agent says it cannot start scans or edit Case Context. Your model does not support function calling. It can still answer questions and explain things. Load a model that supports tool use and click Re-run capability checks, or turn your local model off and use the shared connection for that work.
"This document is larger than your local model's context window." Split the document with Unitize, or load a model with a larger window. Local models usually have much smaller windows than hosted ones.
Everything is slow. That is your hardware, not Claira. Lower the scan speed slider so your machine is not thrashing, or load a smaller and faster model. Running a big bulk scan on a laptop while it is on battery is noticeably slower than on mains power.
Results are weaker than you expected. A small local model is genuinely less capable. Test 10 to 25 documents with Single Review before committing to a large bulk run, and tighten your prompt (definitions, allowed values, output format) before blaming the model. See Understanding AI Variability.
Pricing and billing
Claira Private is $899 per case per month, whichever way you connect your model and however many people work on the case. The shared connection and local models are both included.
- No tokens. A Claira Private case has no token balance, no free daily claims, no token packs and no overage, and nothing is estimated or charged per document.
- No document limits. Scan as much as your model can handle.
- Billed like any other plan. Pay by card or on your monthly invoice. Whatever your endpoint provider charges you, or whatever your own hardware costs to run, is between you and them.
See Plans and Billing.
Related
Need help? Contact us at support@claira.to to discuss your model integration needs.
Was this page helpful?
Continue reading
Opt-in controls for sensitive cases
How Claira's opt-in design at the case and user-group level lets administrators keep restricted matters out of AI processing entirely.
Set up Claira Private
A step-by-step guide for a case on Claira Private: activate it, choose how to connect your own model, set that up, check it works, and switch later.
Need more help?
Contact our support team at support@claira.to — we are here to help.