Coding sources

The first high-value targets for agentic MCP control are code and engineering systems. This page covers what "coding sources" means in Miqro and how they plug into the same gated tool-calling loop as every other capability.

Status: in design, with the connector layer already live for the issue-tracking side (GitHub, Linear, Jira). Full MCP-driven coding control is gated to the Enterprise / miqromanage plan.

What counts as a coding source

A coding source is any system an agent reads or acts on to move engineering work forward:

Source hosts — GitHub (and equivalents): repositories, branches, pull requests, issues, checks.
Issue / project trackers — Linear, Jira: issues, status, assignment, comments. These are already live as typed connectors.
Coding-tool MCP servers — MCP servers that expose a working tree, a build, or a code-intelligence surface (search, references, diagnostics) to an agent.

The connectors give agents a declared, typed slice of these systems today. MCP generalizes it: instead of a hand-written connector per capability, an agent can drive any MCP-exposed coding tool through the one self-hosted client.

Read vs write, in coding terms

The read / write split maps cleanly onto code:

Class	Coding examples	Gate
Read	list repos, read a file, search code, view a PR, fetch CI status, read an issue	runs immediately
Write	open/merge a PR, push a branch, comment, change issue status, close an issue	enqueued; risky writes wait for approval

Reads let an agent understand a codebase or a backlog. Writes — anything that mutates a repo, a PR, or a tracker — go through the action queue. Merging a pull request or pushing to a branch is exactly the kind of high-risk write that pauses for the owner.

How it composes with the rest

Nothing about coding sources is special-cased in the execution path. They are:

discovered through the same catalog / tool-discovery mechanism,
executed through the same self-hosted MCP client,
gated by the same approval + audit machinery described in agentic control,
scoped by the same tenant identity as every other agent action (multi-tenancy).

That means an operator can let an agent triage issues, draft changes, and prepare pull requests autonomously — while keeping the irreversible steps (merge, push, release) behind an explicit human approval.

Today vs. roadmap

Live now: GitHub, Linear, and Jira as typed connectors — reads plus gated writes (issues, PRs where supported). See their connector pages for the exact, generated capability lists.
In design: driving coding-tool MCP servers (working tree, build, code intelligence) through the self-hosted client, so an agent's reach isn't limited to what a hand-written connector exposes.

When the MCP-driven path ships, connection and auth details land here.