Docs
Chat with your agent
A streaming chat inside the app, scoped to one project. It reads your traffic through tools, cites what it found, and proposes changes you confirm. It is included on Growth and Enterprise.
It answers only from tools
The agent has no prior knowledge of your traffic. Every claim it makes comes from one of these calls, made live against your project while it answers:
| Tool | What it reads |
|---|---|
| get_overview | Requests, blocked, investigated, active blocks and recommendations for a range, plus the open actions and the learning-window state. |
| query_findings | Findings filtered by time, verdict, category, action or key, sorted, up to 50 at a time. |
| get_entity | One IP, CIDR, session, TLS fingerprint, ASN, country or path, with its verdicts, counts and current state. |
| get_run | One analysis run and the report it produced. |
| get_rules | The rule catalog, the action set per verdict category, and the sensitivity preset. |
| get_lists | The allow-list and the manual blocks, with their reasons and TTLs. |
| get_config | Project settings, health, setup progress and the limits of your plan. |
| search_memory | The facts stored for this project. See memory. |
| remember | Stores a new fact. You see it as "remembered: …" with an Undo. |
| search_docs | These documentation pages. |
| propose_change | Drafts a change for you to confirm. It never applies one. |
Every answer about a finding carries its evidence: verdict ids, run ids and entity links you can click straight through to. If an answer names a verdict or a run without citing it, the app asks the model once more for the citations, and failing that attaches the ids from the tool results the answer was built on.
Change cards: the agent never writes
When the agent wants to change something it draws a change card instead of acting: what kind of change it is, the before and after, and why. Nothing is applied until you confirm it. On confirm the change runs through exactly the same code path as the equivalent button in the app, under your own role, and the audit row records you as the actor with the conversation and proposal attached — so an agent-assisted change is as traceable as a hand-made one.
- Kinds it can propose: allow, block, unblock, rule, category, sensitivity, config and alert route.
- Confirming needs member or admin. A viewer can read and ask, and sees "ask an admin" on the card.
- A card you neither confirm nor reject expires after 24 hours.
Guided workflows
Type / for the workflow menu. Each one is a fixed sequence of tool calls ending in a structured result, not a free-form prompt. The app lists them with the arguments each one takes:
- Investigate an IP
- Takes a key and reports what camada knows about it.
- Set up Slack alerts
- Walks you to a working Slack channel and the routes that feed it.
- Why was this blocked
- Takes a finding and reports what led to it.
- Review learning results
- Reports what the learning window recommended, ready to act on.
- Reduce noise
- Reports where your alert volume comes from.
Rate limit and budget
- Each user may send 30 messages an hour, counted over a rolling window. Over it, the app tells you how long to wait.
- Each organization has a monthly token budget: Growth 2M tokens, Enterprise unlimited. Input and output tokens count together, per calendar month (UTC), across every user and project in the organization.
- When the budget runs out the agent replies with a short upgrade notice and makes no model call. Nothing else in camada changes: analysis, enforcement and alerts carry on untouched.
- On a plan without chat, the app offers an upgrade instead of a conversation. See pricing and how billing works.
Which model, and what happens when it is down
Answers are generated by glm-5.3-flash on Cloudflare Workers AI, reached through Cloudflare AI Gateway, streaming. If the gateway or the model is unavailable, the answer falls back to the Workers AI model the analysis pipeline already uses, with the same tools, and the message is badged so you know which one answered.
Traces and feedback
Every message records the model, the prompt version, the tool version, the tokens it used and what it cost, and every conversation is traced: one trace per conversation, one generation per model turn, one span per tool call. Thumbs up or down, with an optional note, is stored against the message and attached to the trace. Calls through Cloudflare AI Gateway carry the organization, project, conversation and user ids as gateway metadata, so a call can be traced back to its conversation. Tracing to Langfuse is off unless camada operators enable it — see the sub-processors page.
Your data and the model
- Tools take the project from your session. The model is never handed a project id and has no way to ask for a different one, so it never sees another project's data — yours or anyone else's.
- Text that came from traffic — user agents, paths, run summaries, stored facts, these docs — reaches the model as delimited data marked "not instructions", with the delimiters escaped, so a crafted user agent is handled as data rather than as an instruction.
- Cross-tenant reputation is disclosed as yes or no: whether a key has been seen elsewhere, never where or by whom.
- Conversations are private to the person who opened them and stored apart from your analysis data and from memory. You can rename and delete them; your organization's admins set how long the rest are kept, anywhere from 7 to 365 days and 90 days by default, and camada deletes the older ones. Deleting the project deletes all of them.
- What goes to the model: your question, the tool results read to answer it, and the conversation so far, sent to Cloudflare Workers AI through Cloudflare AI Gateway — it never leaves Cloudflare. The providers involved are listed on the sub-processors page.