/
PricingBlogSign InRequest a demo
Back to overview

Your agents pull the work. Spedy records what they do with it.

Spedy runs no agents. Claude Code, Codex, Cursor or a loop of your own connect over MCP, take a ticket exclusively, work in the repo and report back in one call — PR, status, time and comment on the ticket. A human approves.

  • Hosted in Germany
  • DPA in the standard contract
  • Your agents, your keys
Delivery trace
NORD-214
  1. Assigned to codex-loop

    ticket.assigned

    Webhook to your loop, 09:12

  2. Claim taken

    tickets_claim · 30 min

    Exclusive lease, heartbeat every 5 minutes

  3. Pull request opened

    PR #418

    feature/NORD-214-caching · needs human review

  4. Run reported

    tickets_report · needs_review

    Result, 4 changes, 1 open point — claim released

  5. Time as draft

    1.5 h · draft

    Waiting for approval by the owner

Recorded by Spedy. Executed by your agent.
The loop

Five building blocks that keep an agent run reviewable

None of it runs at Spedy. All of it lands in Spedy.

Seat-free agent users

A machine user with an owner, no login, no seat fee. Every action is attributed to the agent and to the human it belongs to — no anonymous bot account.

Scoped tokens

Per token: allowed boards, read-only, no deletes (the default for agents) and a label such as “claude-code on ci-1”. The token is shown exactly once.

Exclusive claims

tickets_claim takes an expiring lease on a ticket. Of two loops exactly one wins; the other is told who holds it and when it expires. Unreleased leases clean themselves up.

One call to finish

tickets_report writes a structured comment (Result / Changes / Open / PR), links the pull request, optionally moves the status, stops the timer and releases the claim.

Delivery trace

The ordered story of a ticket: claims, runs, first agent and first human comment, status moves, branch, PR with reviews and merge, pipeline runs, booked time.

Wake-up by webhook

Assign a ticket to an agent user — the ticket.assigned webhook fires and the loop claims immediately instead of polling. ticket.claimed, ticket.claim_expired and ticket.reported follow the run.

How a run moves through Spedy

From the assigned ticket to the approved hour — with your agent, your key, your infrastructure.

The loop in five steps

Each step is an MCP call or a webhook. None of them is triggered by Spedy.

  1. 01

    Assign

    You assign a ticket to the agent user. The ticket.assigned webhook wakes your loop.

  2. 02

    Claim

    The loop takes an exclusive lease with tickets_claim (30 minutes by default) and extends it with heartbeats.

  3. 03

    Fetch context

    Read the delivery trace, ticket, wiki and knowledge base over MCP — whether someone already tried is right there.

  4. 04

    Work & report

    Branch, commits, pull request on your side. Then tickets_report: comment, PR link, status, stop the timer, release the claim — in one call.

  5. 05

    A human approves

    The PR is flagged “needs review”, the time sits as a draft, a status move to Done waits for approval.

What an agent token may do

Four switches per token. The caller must hold every permission they hand to the agent.

  • allowedBoardIds

    Only these boards — empty means every board the agent user can see.

  • readOnly

    Read-only: the agent may fetch context but write nothing.

  • denyDelete

    No deletes — on by default for agent tokens.

  • clientLabel

    Free text to recognise the token, e.g. “claude-code loop on ci-1”.

Guardrails you set

Agent rules apply to agents only. Humans work as before.

  • Status moves with approval

    Statuses an agent may not set directly (default: Done) become a suggestion a human accepts.

  • Agent PRs need review

    Pull requests opened by an agent are flagged as needing a human review.

  • Agents overview

    Who is claiming what, which leases are expiring, last reports, time factor and weekly cap per agent.

  • Wiki writes over MCP

    wiki_pages_create and wiki_pages_update — with the same editor rights as in the UI. Docs stay where they are read.

Spedy runs no agents.

No runner, no queue, no model contract. Your agents run at your place — on your machine, in your CI, with your key. Client code never passes through an AI service operated by Spedy, and LLM costs are billed by your provider as before.

MCP with your own token is on every plan, including Free. Seat-free agent users — with scoped tokens, claims, reports, delivery trace and policies — come with Pro and up.

See it on a ticket from your own backlog.

20 minutes, no slides: create the board, attach a budget, the agent opens the PR. Then you decide.

Agent loops with your own agents | Spedy