Quick Start: AI Agent Teams ​

Beta, early access. A workspace without access gets 403 FEATURE_EARLY_ACCESS and a link to request it. Flag: agent_guard.

This guide puts Bumpkit inside your coding agents. When you are done, each agent:

• starts every session with what your team learned about the code,
• has its work checked before it finishes, and is sent back once to fix proven problems,
• waits for a person before it runs a risky command,
• and, when several agents share a repository, claims what it uses and stays out of the others' way.

Each step has its own flag. Run bumpkit features to see which are on for your workspace.

1. 1

   Install the CLI and sign in ​

   1. Install the CLI, which includes the MCP server, and put it on your PATH. See Install the CLI.

   2. Sign in:

      TerminalCopy

      bumpkit login

   3. The MCP server can use the credentials saved by bumpkit login. For CI or an agent running under a separate account, create a key and pass it through that environment:

      TerminalCopy

      bumpkit keys create agents
      export BUMPKIT_API_KEY=bk_live_...

2. 2

   Install the hooks for your agent ​

   Run the installer from the root of the repository. Each installer merges into your existing config, so running it twice is safe.

   Agent	Install	Flag
   Claude Code	bumpkit hooks install --claude (also the pause hook, see step 4)	agent_guard, and agent_pause for the pause hook
   Cursor	bumpkit hooks install --cursor (also the pause hook, see step 4)	agent_guard, and agent_pause for the pause hook
   OpenAI Codex	bumpkit hooks install --codex (also the pause hook, see step 4)	agent_guard, agent_hooks_extended, and agent_pause for the pause hook
   Gemini CLI	bumpkit hooks install --gemini	agent_guard, agent_hooks_extended
   Windsurf	bumpkit hooks install --windsurf	agent_guard, agent_hooks_extended
   Cline	bumpkit hooks install --cline	agent_guard, agent_hooks_extended
   Any agent or person	bumpkit hooks install --git (a pre-push hook)	agent_guard

   Then finish the setup in the agent:

   • Codex: trust the project, and review the hooks with /hooks.
   • Gemini CLI: trust the project hooks the first time they run.
   • Cline: turn on Enable Hooks in Cline's Feature Settings.

   To try it, ask the agent to make a change, then let it finish. If the change has a proven problem, the agent is told what and where, and goes back to fix it. For the exact files each installer writes, see Agent Hooks.

3. 3

   Connect the MCP server ​

   The MCP server lets the agent call Bumpkit directly: check a release, ask the gate, and read history and memory.

   TerminalCopy

   claude mcp add --transport http bumpkit https://bumpkit.dev/mcp

   Then run /mcp in Claude Code, choose the Bumpkit server and sign in. In the Claude desktop app, add https://bumpkit.dev/mcp as a custom connector instead. To run the server next to your checkout, which tools like release_gate need, use claude mcp add bumpkit -- bumpkit mcp.

   bumpkit mcp uses your saved bumpkit login credentials, or BUMPKIT_API_KEY from the client environment. The Codex, Gemini CLI, Windsurf and Cline hook installers configure bumpkit mcp for those agents. See Connect a client.

   The server lists only the tools whose features are on for your workspace. See MCP Server for every tool.

4. 4

   Pause risky commands ​

   Beta, early access. A workspace without access gets 403 FEATURE_EARLY_ACCESS and a link to request it. Flag: agent_pause.

   The pause hook checks each shell command before the agent runs it. It allows it, asks a person, or denies it. For example, rm -r on the repository root and a force push to main are denied. git reset --hard, terraform apply and a DROP TABLE sent to a database client ask first.

   bumpkit hooks install --claude, bumpkit hooks install --cursor and bumpkit hooks install --codex add the pause hook.

   1. Try a command by hand:

      TerminalCopy

      bumpkit pause git push --force origin main

      It prints allow, or ask: or deny: with the reason. Without a key, or with the flag off, it still checks with the built-in rules.

   2. Tune the rules in bumpkit.yml on your default branch:

      YAMLCopy

      agent:
        deny: [infra_apply]
        ask: ["make deploy*"]
        remote: true
        timeout: 60s

      With remote: true, a teammate can approve or deny an ask in Bumpkit while the agent waits.

   The hook reads the committed policy from the default branch (origin/HEAD), never from the working copy, so an agent can't loosen it by editing the file. For every rule, see Pause risky commands.

5. 5

   Share what your team learned ​

   Beta, early access. A workspace without access gets 403 FEATURE_EARLY_ACCESS and a link to request it. Flag: codebase_memory.

   Memory stores decisions, pitfalls and incidents by path. The hooks load it at the start of each session.

   1. Save a memory:

      TerminalCopy

      bumpkit memory add "Webhook handlers must stay idempotent; GitHub retries on timeout." \
        --kind decision --path services/webhooks/github.go

   2. See what an agent will get for an area:

      TerminalCopy

      bumpkit recall --path services/api

   3. Remove a memory once it is no longer true:

      TerminalCopy

      bumpkit memory forget MEMORY_ID

   Agents save and read memories with the remember, recall and forget MCP tools. To have a person approve memories before agents see them, turn on memory approval.

6. 6

   Run several agents in one repository ​

   Beta, early access. A workspace without access gets 403 FEATURE_EARLY_ACCESS and a link to request it.

   When several agents work in one repository at once, these features stop them colliding. Each has its own flag.

   Give each agent an identity ​

   Set these in each agent's environment, with a different value per agent:

   TerminalCopy

   export BUMPKIT_AGENT_ID=claude-1
   export BUMPKIT_SESSION=claude-1-run-1

   Set both. BUMPKIT_AGENT_ID is the agent's name everywhere; BUMPKIT_AGENT is accepted as an alias. See Identify each agent.

   Claim what you use ​

   Beta, early access. A workspace without access gets 403 FEATURE_EARLY_ACCESS and a link to request it. Flag: swarm_claims.

   Before an agent uses a migration number or creates a shared file, it claims it:

   TerminalCopy

   bumpkit claim migration
   bumpkit claim path services/billing/invoice.go --note "new invoice totals"
   bumpkit claims --mine
   bumpkit unclaim CLAIM_ID

   bumpkit claim migration holds the next free number for you. If another agent already holds a resource, the claim is refused and names the holder. See Claims.

   Declare intent ​

   Beta, early access. A workspace without access gets 403 FEATURE_EARLY_ACCESS and a link to request it. Flag: conflict_prediction.

   Before starting, the agent says what it will change:

   TerminalCopy

   bumpkit intent --paths services/billing --symbols billing.ChargeCustomer

   Bumpkit compares it with open pull requests, claims and other agents' intents. The command exits 1 when it predicts a conflict, so the agent can re-plan before it writes code. See Conflict prediction.

   Merge in the right order ​

   Beta, early access. A workspace without access gets 403 FEATURE_EARLY_ACCESS and a link to request it. Flag: merge_planning. Needs the GitHub App.

   TerminalCopy

   bumpkit merge-plan

   This orders the open pull requests so those with the fewest overlaps merge first, with a reason for each position. Bumpkit re-checks waiting pull requests whenever main moves. See Merge planning.

Next steps ​

• See how each coding agent's releases fare: Agent scorecards.
• Record which agent, model and session made each change: AI change provenance.
• Check the claims in an agent's summary against CI: Verified agent reports.
• Share rules every agent follows, set agent budgets and scopes, and review everything that waits for a person in one place: Swarm lessons, Agent budgets and scopes and the Approval inbox.
• Follow the full multi-agent workflow: Run several agents in one repository.
• See how all the pieces fit: How Bumpkit Works.