MCP tools
The built-in kits on Tara provide MCP tools to agents in Claude Code. This page lists two groups: the six delivery tools of kit-scm-github and kit-scm-gitlab, and the three knowledge tools of kit-knowledge.
How agents see the tools
When you install a kit, or a harness that depends on a kit, Claude Code gets the tools right away. You don't need to install any other MCP server. In Claude Code, tool names get a prefix as follows.
| Kit | Name in Claude Code |
|---|---|
kit-scm-github | mcp__plugin_kit-scm-github_scm__<tool name> |
kit-scm-gitlab | mcp__plugin_kit-scm-gitlab_scm__<tool name> |
kit-knowledge | mcp__plugin_kit-knowledge_tara-knowledge__<tool name> |
Tool names are fixed so that Guardrails rules can match by name. Guardrails records the write tools (comment_issue, create_pull_request) for control, and tools from a kit approved on Tara aren't treated as unknown MCP servers.
SCM kits: kit-scm-github, kit-scm-gitlab
The two kits share the same six tools and the same parameters. They differ only in terminology: GitHub says "pull request" and GitLab says "merge request".
The MCP server gets the repo from the origin remote of the project directory, and PHILLIP_GITHUB_REPO=owner/name overrides it. The token comes from your ph connect connection through a Phillip helper command. The server keeps the token in memory and refreshes it when it's close to expiring. It doesn't read tokens from files or environment variables. If you haven't connected, the tool returns an error for the agent to read, and the server keeps running.
Issue and PR content is external data
Content read from an issue or PR is only data. The agent doesn't treat it as instructions.
get_issue
Reads one issue (task): title, body, Bandum status, assignee, milestone, and labels.
| Parameter | Type | Required | Description |
|---|---|---|---|
number | integer ≥ 1 | Yes | Issue number |
Result: an object { number, title, body, closed, status, assignees, milestone, url, labels }. status is todo, in progress, review, approved, or rejected.
list_my_issues
Issues assigned to the token owner. By default, only open issues.
| Parameter | Type | Required | Description |
|---|---|---|---|
milestone | string | No | Milestone (sprint) name |
include_done | boolean | No | Include closed issues; default false |
Result: { me, issues: [{ number, title, status, closed, milestone, url, labels }] }.
comment_issue
Posts a comment on an issue, usually to report progress or results. Don't put secrets or personal data in the comment.
| Parameter | Type | Required | Description |
|---|---|---|---|
number | integer ≥ 1 | Yes | Issue number |
body | string, 1 to 20000 characters | Yes | Comment text |
Result: { ok: true, number }.
create_pull_request
Opens a PR (an MR on GitLab) from the head branch into the base branch. The head branch must already be pushed. For Bandum to recognize the task, put the code GH-<number> in the branch name or write Closes #<number> in the description.
| Parameter | Type | Required | Description |
|---|---|---|---|
title | string, 1 to 256 characters | Yes | Title |
head | string | Yes | Source branch |
base | string | No | Target branch; defaults to the repo's default branch |
body | string, up to 20000 characters | No | Description |
Result: information about the PR that was just opened, including number, title, headRef, headSha, and url.
get_pull_request
Reads a PR by number, or a task's open PR. Provide only one of the two parameters.
| Parameter | Type | Required | Description |
|---|---|---|---|
number | integer ≥ 1 | One of the two | PR number |
issue | integer ≥ 1 | One of the two | The task's issue number, instead of number |
Result: the title, branches, head commit, and the open, closed, or merged state. If you pass issue and the task has no open PR, the result is { found: false, message }.
list_checks
The CI results of a PR.
| Parameter | Type | Required | Description |
|---|---|---|---|
number | integer ≥ 1 | Yes | PR number |
Result: { number, state, checks: [{ name, state, url }] }. Each check has a state of success, failure, pending, or skipped. The overall state is failure if any check failed, otherwise pending, then success, and none when the PR has no checks. If a kit doesn't support viewing checks yet, the tool returns a corresponding error.
Knowledge kit: kit-knowledge
This kit connects Claude Code to the Knowledge base on Tara through MCP. The server runs over HTTP with these parameters.
| Address | https://tara.phillipcli.cloud/mcp/knowledge |
| Authentication | Authorization: Bearer <Tara token>. The kit calls phillip tara auth-header to get the header |
| Protocol | JSON-RPC; all tools are marked read-only |
You can read only the spaces that belong to your current organization, and a space contains only knowledge that the organization has uploaded. How to upload knowledge is in Knowledge base.
search
Finds relevant document passages in the spaces you can read. The tool uses hybrid search: full-text, business codes such as PRJ-123 or AC-1.2, and vector search if Tara enables it.
| Parameter | Type | Required | Description |
|---|---|---|---|
query | string | Yes | A question or keywords |
space | string | No | Space code; leave empty to search all readable spaces |
top_k | integer 1 to 20 | No | Number of passages returned; default 8 |
Result: text made of numbered passages. Each passage has a title, space, citation source (repo, commit, path, heading), and content, together with structured data { hits, mode }. If there are no passages, the tool returns "No passages found in the spaces you can read." When there is a warning, for example only full-text search was possible, the result adds a "Note:" line.
get_document
Reads a whole document. Get space and path from the search result.
| Parameter | Type | Required | Description |
|---|---|---|---|
space | string | Yes | Space code |
path | string | Yes | The document's source path |
Result: the title, a "Source:" line, and the full content.
list_spaces
Lists the knowledge spaces you can read. It has no parameters.
Each result line contains the space code, name, organization, project (if any), and document count, together with structured data { spaces: [{ space, name, org, project, documents }] }. If you can't read any space yet, the tool returns "You cannot read any space yet."
Errors
When a configuration or permission error occurs, the tool returns a result with isError so the agent can read it and handle it. The server doesn't stop.