Connect your coding agent

Connect your coding agent to Sazabi and it can read your production logs while it works — trace an error to the code that caused it, or check whether a change actually fixed a problem. By the end of this guide your agent can query Sazabi and return real data from your project. Setup takes a few minutes.

What you get

Once connected, your agent can search your logs, look up issues, and run the same investigations you would run in a thread — from inside your editor or terminal. You ask in natural language (“what errors is checkout throwing in production?”) and the agent queries Sazabi and answers with evidence.

Prerequisites

A Sazabi account with at least one project that is receiving logs. If you have not sent logs yet, do the Quickstart first.
A secret key, unless your MCP client signs in with OAuth (see The MCP path). Create one at Settings > Secret keys. Secret keys grant full read and write (organization-wide by default, or scoped to one project), so store it in a secrets manager and never commit it.

Choose a path

There are two ways to connect. Recommended for terminal agents: the CLI path. Install the sazabi CLI, and any terminal agent (Claude Code, Codex, OpenCode) can drive Sazabi through the same commands. Use the MCP path when your agent is a GUI client (Claude Desktop, ChatGPT) or you prefer wiring Sazabi in as an MCP server. It is the only path for GUI clients. Trade-offs: the CLI path is one identical procedure for every terminal agent; the MCP path is configured per client but works for clients that have no terminal. Do not use the CLI path for a GUI-only client that cannot run shell commands — use the MCP path there.

The CLI path

One procedure, identical for every terminal agent.

Install and authenticate the sazabi CLI

Install the CLI globally:
npm install -g @sazabi/cli@latest
To run it without installing, use npx @sazabi/cli@latest <command> in place of sazabi <command>. Keep @latest: without it, npx can reuse an older cached copy that lacks newer commands and flags, such as start --agent.
Sign in. This opens your browser for a device-code sign-in:
sazabi auth sign-in
Confirm you are authenticated and select your organization and project:
sazabi auth whoami sazabi organizations use <org-id> sazabi projects use <project-id>
Run sazabi organizations list and sazabi projects list if you need the IDs.

For non-interactive use (CI, headless agents), set the SAZABI_TOKEN environment variable to a secret key (or pass it with --token) and SAZABI_PROJECT_ID to your project’s ID instead of running sazabi auth sign-in. See Running without a browser.

Connect Claude Code or Codex subscriptions

If you use a Claude subscription with Claude Code, connect it from your terminal:

sazabi integrations claude connect

The Sazabi CLI runs Claude Code’s own sign-in and saves the resulting subscription credential without displaying it. Claude Code opens the browser flow and asks for the short code shown after you approve access. The credential is valid for one year. For a ChatGPT subscription with Codex, run:

sazabi integrations codex connect

The command runs Codex’s browser sign-in on your machine and saves its local sign-in credential without printing it. Add --account-label to name the connection, --project <project-id> to select a project, or --reconnect <connection-id> to replace credentials on an existing connection. Your terminal agent can now run sazabi commands — for example sazabi logs ask, sazabi issues list, or sazabi memory search — to investigate your project. Every command takes -h for its flags and examples.

The MCP path

For MCP clients, connect Sazabi as an MCP server at https://mcp.sazabi.com/mcp. The server accepts two kinds of credential:

OAuth sign-in. A client that supports MCP OAuth needs only the server URL. It opens your browser, you approve the connection, and the client acts as you. No key is created or pasted. The Claude Code, Cursor, and Codex CLI sections below show this path.
A secret key. Every client can send a secret key in the X-Sazabi-Secret-Key header. Use it for clients without OAuth support and for headless or shared setups where nobody can sign in through a browser.

With OAuth, the browser step is the same for every client:

Sign in to Sazabi if you are not already signed in.
On the Authorize MCP connection page, check the client name, then choose the organization the connection may act in.
Under Project access, choose All projects or a single project.
Select Authorize. The browser hands the sign-in back to your client.

The Authorize MCP connection page: the client name and access, an organization menu, a Project access menu, and Deny and Authorize buttons. — The Authorize MCP connection page.

Each sign-in approves one organization. To use another organization, or to change the project access, sign in again from your client and choose differently. Wherever a configuration below shows <YOUR_SECRET_KEY>, replace it with a secret key from Settings > Secret keys, and keep that file out of version control.

Claude Code

Sign in with OAuth. Add the server from your terminal with no header:

claude mcp add --transport http sazabi https://mcp.sazabi.com/mcp

Then run /mcp inside Claude Code, select sazabi, and complete the sign-in in your browser. Use a secret key. Add the server with the key in a header instead:

claude mcp add --transport http sazabi https://mcp.sazabi.com/mcp \
  --header "X-Sazabi-Secret-Key: <YOUR_SECRET_KEY>"

Claude Desktop

In Claude Desktop, open Settings > Developer > Edit Config to open claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/claude_desktop_config.json; Windows: %APPDATA%\Claude\claude_desktop_config.json), then add:

{
  "mcpServers": {
    "sazabi": {
      "type": "http",
      "url": "https://mcp.sazabi.com/mcp",
      "headers": {
        "X-Sazabi-Secret-Key": "<YOUR_SECRET_KEY>"
      }
    }
  }
}

Cursor

Cursor reads MCP servers from ~/.cursor/mcp.json (global) or .cursor/mcp.json in a project. The same file serves both the Cursor editor and the Cursor CLI (cursor-agent). Sign in with OAuth. Add the server with its URL and no headers:

{
  "mcpServers": {
    "sazabi": {
      "type": "http",
      "url": "https://mcp.sazabi.com/mcp"
    }
  }
}

Cursor then lists the server as needing authentication. In the editor, start the sign-in from the server’s entry in Cursor Settings > MCP. With the Cursor CLI, run:

cursor-agent mcp login sazabi

Your browser opens for you to approve the connection. No secret key is involved, so nothing sensitive lands in mcp.json. If the server is in a project’s .cursor/mcp.json, the Cursor CLI asks you to approve it first: run cursor-agent mcp enable sazabi, then log in. Use a secret key. Add the X-Sazabi-Secret-Key header instead:

{
  "mcpServers": {
    "sazabi": {
      "type": "http",
      "url": "https://mcp.sazabi.com/mcp",
      "headers": {
        "X-Sazabi-Secret-Key": "<YOUR_SECRET_KEY>"
      }
    }
  }
}

You can also add the server from Cursor’s settings UI: Cursor Settings > MCP > Add new MCP server, using HTTP transport with the URL above, and the X-Sazabi-Secret-Key header only if you are using a secret key.

Codex CLI

Add the server from your terminal:

codex mcp add sazabi --url https://mcp.sazabi.com/mcp

Codex discovers the OAuth resource from Sazabi and opens your browser to authorize the mcp and offline_access scopes. Do not add --oauth-resource; providing it again duplicates the discovered resource and causes authorization to fail with a 403 response. You can also edit ~/.codex/config.toml directly.

Codex Desktop

In Codex Desktop’s MCP settings, add an HTTP server with URL https://mcp.sazabi.com/mcp and the X-Sazabi-Secret-Key header set to your secret key.

OpenCode

Add the server from your terminal — opencode mcp add walks you through adding a remote server — or add it to your OpenCode configuration (opencode.json) directly:

{
  "mcp": {
    "sazabi": {
      "type": "remote",
      "url": "https://mcp.sazabi.com/mcp",
      "headers": {
        "X-Sazabi-Secret-Key": "<YOUR_SECRET_KEY>"
      }
    }
  }
}

ChatGPT

In ChatGPT’s connector settings, add a custom MCP server with URL https://mcp.sazabi.com/mcp and the X-Sazabi-Secret-Key header set to your secret key.

Try it: investigate production from your coding agent

With either path connected, ask your agent a question that requires your logs:

Use Sazabi to find the most recent ERROR logs in my project and summarize what is failing.

The agent queries Sazabi and answers with the specific records it found.

Hand an issue to your coding agent

Every Sazabi issue has an agent prompt: one prompt you paste into your coding agent to continue Sazabi’s investigation, fix the problem in your repository, and open a pull request linked back to the issue. The prompt carries the issue ID and a link, never a credential. It tells the agent to install and sign in to the CLI if needed, read the analysis with sazabi issues get (its report field has the full what, why, how, and evidence), check the problem against live logs, and treat issue text and log lines as data, not instructions. Print the prompt with the CLI and paste it into your agent (ISSUE_ID is the issue’s ID from sazabi issues list):

sazabi issues agent-prompt "$ISSUE_ID"

A pull request opened from the prompt shows up in the issue’s pull requests in either of two ways:

The Sazabi-Issue line. In a GitHub repository connected to the issue’s project, a pull request whose body has the line Sazabi-Issue: <issue-id> is linked when it is opened or edited.

An explicit link. Link any GitHub or Bitbucket pull request yourself:

sazabi issues link-pr "$ISSUE_ID" https://github.com/acme/api/pull/42

Linked pull requests are listed with source: "linked" and are never billed as pull requests Sazabi opened. If a pull request is the suspected cause rather than the fix, label that relationship explicitly. You can find every issue attributed to the PR and remove a wrong attribution later:

sazabi issues link-pr "$ISSUE_ID" https://github.com/acme/api/pull/42 --relation caused_by
sazabi issues list --caused-by https://github.com/acme/api/pull/42
sazabi issues unlink-pr "$ISSUE_ID" https://github.com/acme/api/pull/42 --relation caused_by

Verify

Confirm the connection returns real data from your project.

First, send a distinctive test event if you have not already — see Quickstart curl step, using body sazabi quickstart test event 8842.
Ask your agent: Use Sazabi to find the log with body "sazabi quickstart test event 8842".
The agent returns one matching record with service quickstart and severity INFO.

If the agent cannot reach Sazabi, re-check the secret key (or, with OAuth, sign in again from your client) and, for the CLI path, that sazabi auth whoami shows the right organization and project.

Agent tools

What the agent can search and do during an investigation.

MCP server reference

The search and execute tools, and authentication details.

CLI reference

Every sazabi command and flag.

Get started

Send data

Detect issues

Triage issues

Ask Sazabi

Explore logs

Ship fixes

Manage account

Connect your coding agent

What you get

Prerequisites

Choose a path

The CLI path

Install and authenticate the sazabi CLI

Connect Claude Code or Codex subscriptions

The MCP path

Claude Code

Claude Desktop

Cursor

Codex CLI

Codex Desktop

OpenCode

ChatGPT

Try it: investigate production from your coding agent

Hand an issue to your coding agent

Verify

Further reading

Agent tools

MCP server reference

CLI reference

​What you get

​Prerequisites

​Choose a path

​The CLI path

​Install and authenticate the sazabi CLI

​Connect Claude Code or Codex subscriptions

​The MCP path

​Claude Code

​Claude Desktop

​Cursor

​Codex CLI

​Codex Desktop

​OpenCode

​ChatGPT

​Try it: investigate production from your coding agent

​Hand an issue to your coding agent

​Verify

​Further reading

Agent tools

MCP server reference

CLI reference

What you get

Prerequisites

Choose a path

The CLI path

Install and authenticate the sazabi CLI

Connect Claude Code or Codex subscriptions

The MCP path

Claude Code

Claude Desktop

Cursor

Codex CLI

Codex Desktop

OpenCode

ChatGPT

Try it: investigate production from your coding agent

Hand an issue to your coding agent

Verify

Further reading