Agent CLI and MCP

Connect Grok Bot or another MCP assistant to your TractionScout project to read evidence and prepare work. TractionScout keeps your data-source credentials, project permissions, approval policy, audit records and budgets on the server. You do not need access to the TractionScout source repository or an npm registry account.

Download the connector package, or let your assistant install the verified release with the prompt below.

Connect Grok Bot

1. Open Security settings. If your account can manage access tokens, create a dedicated named token, such as “Grok Bot”. Otherwise, ask your administrator about API access. Keep the value for the assistant's secure input; never paste it into ordinary chat.
2. Copy the setup prompt below into your own Grok Bot conversation. Tell it which project to use when asked. For another TractionScout deployment, replace the origin in the prompt before using it.
3. Enter the token only in the secure input the Bot presents. It should verify the connection and report the selected project and available tools.

Connect my Grok Bot to TractionScout at https://tractionscout.com in read_only mode.
Ask which of my projects to use; do not choose a different project without me.

Use your own persistent cloud computer. Fetch
https://tractionscout.com/downloads/agent-cli/manifest.json over HTTPS.
Require name @tractionscout/agent-cli, a supported node_engine, and a filename
that is a single .tgz basename. Download that filename only from the same
origin and downloads/agent-cli directory, without redirects. Check the downloaded
byte length against size and its SHA-256 against sha256 before installation.
If a download or check fails, stop and report it; do not fall back to GitHub,
an npm registry, or an unverified package.

Use Node.js >=18.10.0 and npm; stop and report if either is unavailable.
Install the verified local tarball into a new dedicated private directory with
npm install --prefix <private-directory> <absolute-tarball>
--offline --ignore-scripts --no-audit --no-fund. This package has no dependencies.
Do not install globally or change an existing TractionScout connection.
Use <private-directory>/node_modules/.bin/tscout for the installed CLI.

Request my token through secure input and pass it to the CLI through stdin at
runtime. Never put the token in chat, command arguments, generated MCP config,
logs or source files, and never print environment variables or credential files.
Choose a dedicated new private credentials path and set
TRACTIONSCOUT_AGENT_CONFIG_FILE to that path. Map my project's name to its ID
with securely authorized project discovery. If several projects are accessible,
mcp:setup without --project returns their IDs and names without saving credentials.
Ask me to resolve an ambiguous name; never pick an arbitrary project.
Run the installed tscout mcp:setup
--url https://tractionscout.com --project <my-project-id> --mode read_only
--token-stdin --output <new-mcp-config-file>.

Register the resulting mcp_config as a private Command MCP server using its
exact command, args and env. The env contains the credentials filepath, not
the token. Run tscout mcp:doctor with the same credentials path and read_only
mode. Confirm initialization, project access and tools before continuing.
Through the registered Command connection, follow every tools/list nextCursor
until the full tool list is loaded. Then read one small existing context_bundle
(overview, limit 1) without triggering research or refreshing providers. Confirm
the returned project and evidence references. Report any connection failure.
Do not start routines, paid research or staged actions as part of setup.

The release manifest identifies the current download and checksum. If it is unavailable, the deployment has not published a connector package yet; ask its administrator to make the release available.

This private Command connection does not require a marketplace listing. Your Grok account or team may still restrict custom connectors. Grok Bot runs the bridge on its own cloud computer; it does not connect to your laptop's localhost. Personal Bots on the same Grok account share that computer and its credentials.

Use another MCP client or set up locally

Download and verify the package using the release manifest, then install the local tarball into a private prefix as described above. Set TRACTIONSCOUT_CONNECTOR_PREFIX to that installation directory and run:

export TRACTIONSCOUT_AGENT_CONFIG_FILE="$HOME/.tractionscout/my-mcp-credentials.json"
export PATH="$TRACTIONSCOUT_CONNECTOR_PREFIX/node_modules/.bin:$PATH"
tscout mcp:setup --url https://tractionscout.com \
  --project YOUR_PROJECT_ID --mode read_only \
  --token-stdin --output ./tractionscout-mcp.json
tscout mcp:doctor --mode read_only

In a terminal, token entry is hidden. For an assistant, use its secure input and stream the saved secret into stdin without printing it. Setup checks the selected project's access and MCP connection before saving credentials. It writes private credential files and a separate token-free client configuration; --output requires a new file. Import the returned mcp_config into your client. Keep the dedicated credentials path when checking or reconnecting. If you know the project's name but not its ID, setup without --project returns choices when several projects are accessible; a sole accessible project is selected automatically. Confirm the project before using the connection.

Browser OAuth authorization is not available for this connector today. Its token uses your account permissions across TractionScout; selecting a project in the connection does not make the token a project-scoped credential. Revoke the dedicated token in Security settings when access is no longer needed.

Reading and preparing work

The default read_only mode blocks user-authored staging. Some reads, including explicit keyword research, can still incur provider charges and persist evidence. Inspect tool annotations and agree budgets before requesting paid research or recurring work.

To allow supported internal review artifacts, set up a separate connection with --mode review_staging and verify it with the same mode. All MCP modes block external execution. Provider credentials stay in TractionScout's server-side integrations.

Discover before preparing work

Use your selected project ID in the commands below:

export TRACTIONSCOUT_PROJECT_ID=YOUR_PROJECT_ID
tscout auth:status
tscout projects:list
tscout integrations:list "$TRACTIONSCOUT_PROJECT_ID"
tscout capabilities:execution "$TRACTIONSCOUT_PROJECT_ID"

auth:status verifies the configured token with the server; projects:list shows the projects it can access. Inspect the execution capabilities before choosing tools or asking an agent to act.

When the project has suitable evidence and you intend to start an agent run:

tscout agents:orchestrate "$TRACTIONSCOUT_PROJECT_ID" \
  --prompt "Review the connected evidence and stage a useful next growth move."
tscout actions:list "$TRACTIONSCOUT_PROJECT_ID" --status queued

Agent runs use the project's configured AI resources and budget. The orchestration response returns a durable run_id; inspect it with tscout agents:status "$TRACTIONSCOUT_PROJECT_ID" "$RUN_ID". Review the staged work and applicable approvals before execution. Commands return JSON; --compact produces single-line output for automation.

See Connecting your data sources and Getting started for the project setup behind these workflows.