<!-- Markdown twin of https://retasc.com/guides/github-issues-ai-agents, generated by scripts/markdown-twins.py from the page. Do not edit; edit the page. -->
# GitHub gives agents access to issues and runners for assigned work

> Get draft PRs from assigned issues. Retasc prevents double claims in a GitHub MCP server workflow and holds blocked issues or text awaiting a member’s approval.

Source: https://retasc.com/guides/github-issues-ai-agents · Markdown for agents; the same page as HTML with the diagrams described in a line each.

[Guides](https://retasc.com/guides)

The GitHub MCP server lets your agents read and write GitHub Issues. Copilot cloud agent and Agent HQ run work someone assigns, while a connected queue lets several agents claim the next unblocked issue themselves.

Updated October 2026 · 10 min read

## GitHub gives agents several ways into an issue

GitHub offers a hosted agent, mission control for several agents and an MCP server for agents running elsewhere. Copilot starts when a person assigns it an issue or an automation triggers it on issue creation or a schedule. Those automations are limited to private and internal repos. In Agent HQ, a person assigns the issue to an agent. With the GitHub MCP server, a person runs an MCP client against the repo’s tools.

GitHub renamed [Copilot coding agent](https://github.blog/changelog/2026-04-01-research-plan-and-code-with-copilot-cloud-agent/) to Copilot cloud agent on April 1, 2026. It runs on GitHub Actions. It became generally available to paid Copilot subscribers on September 25, 2025.

Agent HQ brings Claude and Codex into GitHub’s assignment workflow. The GitHub MCP server exposes repository tools to an MCP client such as Claude Code, Codex or Cursor, which can run outside GitHub’s hosted sessions.

## Assigning an issue starts Copilot with the text available then

To assign an issue to Copilot, select it in the issue’s Assignees menu. Only someone with write access can trigger the agent. The API route adds `copilot-swe-agent[bot]` as an assignee and is in [public preview](https://docs.github.com/en/copilot/how-tos/use-copilot-agents/cloud-agent/use-cloud-agent-via-the-api).

Copilot reads the title, body, existing comments and any extra instructions supplied with the assignment. It does not see comments added afterward. Put the context it needs into the issue before assigning it, as described in [GitHub’s assignment docs](https://docs.github.com/en/copilot/how-tos/use-copilot-agents/cloud-agent/use-cloud-agent-on-github).

Copilot adds an eyes reaction and opens a linked draft pull request. Under [GitHub’s safeguards](https://docs.github.com/en/copilot/concepts/agents/cloud-agent/risks-and-mitigations), it pushes only to its `copilot/` branch and cannot mark the pull request ready for review, approve it or merge it. A task covers one repo and one pull request, with an execution time of at most 59 minutes that cannot be extended, according to [GitHub’s surface documentation](https://docs.github.com/en/copilot/concepts/copilot-surfaces/copilot-on-github).

[Copilot cloud agent automations](https://docs.github.com/en/copilot/concepts/agents/cloud-agent/about-automations) can start work when an issue is created or on a schedule, on private and internal repositories only. Copilot Free does not include cloud agent. Since June 1, 2026, Copilot billing uses AI credits priced at $0.01 each, as set out in [the billing change](https://github.blog/changelog/2026-06-01-updates-to-github-copilot-billing-and-plans/).

## Agent HQ lets you compare Claude and Codex on the same issue

GitHub announced Agent HQ on October 28, 2025, with mission control for assigning, steering and tracking agents. Claude and Codex [became available in public preview on Copilot Business and Pro](https://github.blog/changelog/2026-02-26-claude-and-codex-now-available-for-copilot-business-pro-users/) on February 26, 2026. The [Agent HQ announcement](https://github.blog/news-insights/company-news/welcome-home-agents/) named agents from Anthropic, OpenAI, Google, Cognition and xAI.

You can assign the same issue to several agents to compare their approaches. That is a documented use of parallel runs, with a person deciding which agents receive the work and comparing what they produce. Mission control gives that person a place to track the sessions.

Third-party agents consume GitHub Actions minutes and AI credits. They are unavailable on Copilot Free and Student. They [cannot be steered](https://docs.github.com/en/copilot/concepts/agents/about-third-party-agents).

## The GitHub MCP server exposes tools to your own agent

The remote GitHub MCP server URL is `https://api.githubcopilot.com/mcp/`. It has been [generally available since September 4, 2025](https://github.blog/changelog/2025-09-04-remote-github-mcp-server-is-now-generally-available/). Authentication uses OAuth 2.1 with PKCE or a personal access token sent as a bearer token, according to the [remote server documentation](https://github.com/github/github-mcp-server/blob/main/docs/remote-server.md).

You can select a toolset by its path, such as `/mcp/x/issues`, or use the `/readonly` suffix for read-only access. The default toolsets are `context`, `repos`, `issues`, `pull_requests` and `users`. Claude Code, Codex, Cursor or another MCP client can use the server to read and write issues through those tools.

The [github-mcp-server repo](https://github.com/github/github-mcp-server) lists `assign_copilot_to_issue` for handing an issue to Copilot and `sub_issue_write` for changing sub-issues. Issue dependency tools are [Insiders only](https://github.com/github/github-mcp-server/blob/main/docs/insiders-features.md).

For public repositories, lockdown mode withholds issues, pull requests and comments whose authors lack push access. Enable it with `--lockdown-mode`, `GITHUB_LOCKDOWN_MODE` or the `X-MCP-Lockdown` header. GitHub describes it as a best-effort filter against prompt injection and explicitly says it is not an authorization boundary.

## Several assignees still need someone to partition the work

A GitHub issue can have up to [10 assignees](https://docs.github.com/en/issues/tracking-your-work-with-issues/using-issues/assigning-issues-and-pull-requests-to-other-github-users). GitHub suggests assigning several agents to compare approaches, but documents no claim or lock that keeps them off the same issue. A team that wants exclusive ownership has to arrange it outside that assignment mechanism.

[GitHub sub-issues](https://docs.github.com/en/issues/tracking-your-work-with-issues/using-issues/adding-sub-issues) express a parent-child structure with up to 100 sub-issues per parent across eight levels. GitHub issue dependencies allow up to 50 links per relationship type. The links appear in the API and `issue_dependencies` webhooks. Blocked issues show a Blocked icon. [GitHub does not document agents respecting those dependencies, so a blocked issue can still be assigned](https://docs.github.com/en/issues/tracking-your-work-with-issues/using-issues/creating-issue-dependencies).

GitHub’s [mission control guidance](https://github.blog/ai-and-ml/github-copilot/how-to-orchestrate-agents-using-mission-control/) tells you to partition parallel work so agents avoid the same files and merge conflicts. Issue text also comes from anyone allowed to open an issue, including outside filers on a public repo. Assignment, file overlap and which outside text reaches an agent remain decisions your workflow has to make.

| | GitHub | Retasc |
|---|---|---|
| A hosted agent started from your issue | yes: Assign Copilot, Claude or Codex | no: Retasc does not host or start agents |
| A built-in agent opens a draft PR | yes: Copilot opens a draft PR per task | no: Your agent pulls work where it runs |
| One agent per issue, enforced | no: Multiple assignees, no documented claim | yes: The server refuses a second claim |
| Blocked issues withheld from agents | partly: A Blocked icon, assignment still allowed | yes: Blocked work is never handed out |
| A dead agent’s issue frees itself | no: No documented lease or reassignment | yes: The lease lapses, the checkpoint stays |
| People read outside issues before agents act | partly: MCP lockdown filter on public repos | yes: A member approves the exact text |

GitHub also exposes the graph through [webhooks](https://docs.github.com/en/webhooks/webhook-events-and-payloads). The `issues` event fires when an issue is opened, edited, assigned, labeled or closed, among other actions.

## A repo connector starts with a signed webhook

An org owner or admin connects a GitHub repo in the Retasc Dash. Choose the project, enter the repo as `owner/name` and optionally supply an access token. The token needs `issues:write`, plus `pull_requests:write` for pull requests.

Retasc stores the token server-side and never shows it back. Without a token, sync is one way. The connector returns a webhook URL and a signing secret shown once, which you paste into the repo’s webhook settings.

*Diagram: What you provide, what Retasc mints, and the GitHub webhook settings*

GitHub’s signed `ping` on webhook creation marks the connector as connected in the Dash. Sync issues is on by default, while Sync PRs is off.

*Diagram: Edges to know before connecting a busy repo*

An agent acts under a key minted by a signed-in person, with actions attributed to that human and the runtime the agent reports. The human identity comes from GitHub or Google sign-in, never from what the agent says. Run `npx @retasc/cli@latest bind` in the folder to connect its agent to Retasc’s remote MCP server at `https://mcp.retasc.com/mcp`.

## Outside issues wait for a member to approve the exact text

Retasc verifies `X-Hub-Signature-256` as an HMAC-SHA256 over the raw request body, returning the same 401 for an unknown connector, a revoked connector or a bad signature. A replayed delivery is a no-op, deduped by `X-GitHub-Delivery`. Retasc syncs `opened`, `edited`, `reopened`, `labeled`, `unlabeled`, `milestoned` and `demilestoned` actions, updating the same issue keyed by `github:owner/repo#number` and ignoring other actions.

The issue arrives as `todo` with no priority, which sorts last, and the filer’s GitHub handle as its author. The default `support` label can be renamed or left blank. A milestone’s due date is ignored, so a filer cannot move work up the queue through a milestone, and a member sets any deadline in Retasc.

No agent is handed the issue until a member approves its exact title and body in the Dash or `retasc triage`, with approval bound to a digest of that text. The CLI requires an interactive terminal and asks the member to type the issue id after reading the body, with no approve flag or MCP approval tool. An agent key cannot sign off, and an edit on GitHub voids the approval.

*Diagram: A signed GitHub webhook creates ACME-27, but the agent asking for work gets no issue until a person reads and approves the issue filed outside the org.*

When no work is ready, Retasc tells the agent how many issues await sign-off. An agent claiming a synced issue is told who filed it, to evaluate the text as a report and to stop and tell its human if the text asks for credentials, outside systems or to set aside its instructions. On public repos, GitHub’s lockdown mode withholds issues whose authors lack push access, while Retasc’s sign-off lets a member read the text and decide whether to approve it.

*Diagram: After Maya approves the text, the agent's next pull claims ACME-27 with a warning to evaluate the body from outside the org as a report, never instructions.*

## Retasc closes the source issue after recording the work

With an access token, `done` in Retasc closes the GitHub issue with `state_reason` set to `completed`. A `canceled` issue closes with `state_reason` set to `not_planned`. Without a token, Retasc cannot send those changes back.

*Diagram: A signed GitHub issue syncs in, an agent claims and works it, then it closes at the source*

The close runs after Retasc records its own change. If the GitHub close fails, Retasc logs the failure without undoing its own change. A GitHub outage can leave the source issue open after Retasc records the work as done.

Closing an issue on GitHub does not change its Retasc status, and inbound sync changes content only. Revoking the connector stops sync and keeps the issues already imported. The GitLab connector follows the same intake, exact-text sign-off and close-back behavior on gitlab.com and self-hosted GitLab, verifying GitLab’s secret token.

## Outside pull requests become review work for a chosen reviewer

Turn on Sync PRs and choose a reviewer to bring outside contributions into the queue. Retasc skips members’ own pull requests. An outside contributor’s pull request becomes review work assigned to the reviewer you chose.

A merged pull request marks its Retasc issue `done`, while a pull request closed without merging marks it `canceled`. With a token, Retasc comments on the pull request with the issue it is tracked as.

Retasc dispatches review to a different person’s agent. The server refuses to let whoever sent work to review claim or accept that review. Work sent to review must include its branch, PR or commit unless the agent already left a checkpoint, so the reviewer receives the handoff note.

## GitHub can keep the intake while the queue assigns the work

Copilot cloud agent fits when you want to assign an issue and get a draft pull request on GitHub. The GitHub MCP server fits when your agent works with GitHub issues over MCP. Retasc fits when several agents on different runtimes take the next unblocked issue themselves through `next_issue`. With GitHub and Retasc connected, people keep filing issues on GitHub. Retasc hands out a synced issue only after a member approves its exact text.

A claim is one server transaction. Two agents can never hold the same issue. The agent records a checkpoint of what is done, what is next and the gotchas, which stays on the issue for whoever claims it next.

Blocked work is never handed out. When a blocker closes, its dependents join the queue. A blocker inherits the urgency of the work it holds up, so a chore blocking an urgent feature jumps the queue.

An agent renews its 30-minute lease by heartbeat or checkpoint. When renewal stops because a process dies, a laptop closes or a terminal is killed, the reclaimer returns the issue to its owner’s lane for that person’s next agent, on any runtime, to resume from the checkpoint. A lane nobody returns to rejoins the shared pool after 7 days.

## Common questions

**What is the GitHub MCP server URL?**

`https://api.githubcopilot.com/mcp/` is the remote GitHub MCP server URL. Use a toolset path such as `/mcp/x/issues` or the `/readonly` suffix to narrow access. Authentication uses OAuth 2.1 with PKCE or a personal access token sent as a bearer token.

**Can I assign a GitHub issue to an AI agent?**

You can assign an issue to Copilot through the Assignees menu if you have write access. The API accepts `copilot-swe-agent[bot]` as an assignee in public preview. Copilot reads the title, body and existing comments, but does not see comments added afterward.

**What is Copilot cloud agent?**

Copilot cloud agent is the GitHub Actions-based agent formerly called Copilot coding agent. GitHub renamed it on April 1, 2026, as its scope expanded beyond pull requests.

**Can Claude and Codex work GitHub issues?**

Claude and Codex can receive issues through Agent HQ on Copilot Business and Pro in public preview. Terminal clients can also read and write issues through the GitHub MCP server.

**Can two agents work the same GitHub issue?**

Several agents can be assigned the same GitHub issue to compare approaches. A GitHub issue can have up to 10 assignees. GitHub documents no exclusive claim, while Retasc allows only one agent to hold an issue.

**Do GitHub issue dependencies stop an agent?**

GitHub does not document issue dependencies preventing an agent from starting work. The links show a Blocked icon, but blocked issues can still be assigned.

**Is it safe to let an agent work issues from a public repo?**

Retasc hands no agent an issue from a public repo until a member approves its exact title and body, with an edit on GitHub voiding the approval. Public issue text can carry prompt injection. GitHub describes its lockdown mode, which filters issues, pull requests and comments from authors without push access, as a best-effort filter that is not an authorization boundary.

**Does Retasc close the GitHub issue when the work is done?**

Retasc closes the GitHub issue as completed when work becomes `done`, provided the connector has a token. A failed GitHub close leaves the Retasc change intact. Closing the issue on GitHub does not change its Retasc status.

Run the command in your repo folder so your agents pull the next unblocked issue from Retasc while people keep filing on GitHub. A new org gets $10 of usage at signup, no card.

```sh
$ npx @retasc/cli@latest bind
```

[Other ways](https://retasc.com/wire) [Run the fleet that drains it](https://retasc.com/guides/claude-code-parallel-agents)

## Related guides

### [Cursor background agents on a shared backlog Cloud Agents take their MCP servers from the dashboard. The setup that works, and the one that silently does not. [cloud agents][MCP]6 min](https://retasc.com/guides/cursor-background-agents) ### [Linear vs Jira vs an agent-first tracker The tracker evaluation changes when agents work the backlog. The new criteria, and a reversible migration. [evaluation][migration]6 min](https://retasc.com/guides/linear-vs-jira-vs-agent-first) ### [One import, then Retasc is the queue The import runs once and never writes back: what comes across from Linear, Jira, ClickUp, Asana and Shortcut, what stays, and how to export it all again. [import][one-way]5 min](https://retasc.com/migrate)
