AI coding agents
Assign an issue. Get a pull request back.
Deckbell’s agents are AI teammates inside your issue tracker. Give one an issue and it reads the issue and its discussion, works in your repository, pushes a branch, and Deckbell opens the pull request, with every step, its time and its cost on the issue. Agents never take a seat.
Free for up to 5 members, with about $2 of trial AI credits to run your first agent. No credit card.
Add proration to invoice preview
When seats change mid-cycle the preview shows the full price. It should show the prorated amount for the rest of the period, like the invoice does.
@Coder can you take this? Keep the rounding the invoice uses.
On it. Working in harbor/billing on branch coder/hrb-142.
| Step | Used for | In | Out | Cost |
|---|---|---|---|---|
| Size | Task sizing | 4.1k | 0.2k | <$0.01 |
| Plan | Plan and review | 212k | 3.1k | $0.03 |
| Code | Standard model | 640k | 21k | $0.07 |
| Review | Plan and review | 230k | 2.4k | $0.03 |
| Run total | $0.13 | |||
Sized normal. Auto chose plan, code and review from past runs of this size. Cost limit not reached; coding wasn’t retried.
How a run works
From issue to pull request, with a receipt.
An agent run is a teammate’s workflow: read the brief, work on a branch, open a pull request, show what it cost. Nothing happens outside the issue you can see.
- 01
Hand it the issue
Assign the issue to the agent, @mention it in a comment, a channel or a document (“@Coder can you take this?”), or press Run on the Agents tab. It answers where it was asked.
- 02
It reads the brief
The run starts with the issue’s title, description, status, priority and labels, its parent and sub-issues, linked issues and the last 30 comments, plus what you asked for.
- 03
It works in your repository
One branch per issue, like coder/hrb-142. Auto sizes the task, then plans, codes and reviews on frontier models, pushes the branch, and Deckbell opens the pull request. Run it again and it continues the same branch and PR.
- 04
You review and merge
A card on the issue shows the lines added and removed, files changed, how long the run took and what it cost. Open the run to see every step, its tokens and its cost. You decide what merges.
Deckbell Auto
The right frontier model for each step.
Agents on Deckbell don’t ask you to pick a model. Auto sizes the task, then runs each step on the frontier model that fits it, and escalates when the agent gets stuck.
- Escalation. If the coding step fails or pushes no changes, it’s retried once on a stronger model. If nothing needs to change, the agent says so and you aren’t charged twice.
- Learns from your runs. Once 5 coding runs of a size have finished in the last 90 days, Auto uses the cheapest approach that succeeds at least 80% of the time. Owners and admins can turn this off.
- Priced per step. Each step is charged as it finishes, from the tokens it used, so a run costs the sum of its steps. A typical task is about $1.59 in AI credits, about 14% less than running the strongest model on your own key.
- Size01
A quick check reads the issue, its labels, estimate, sub-issues, links and discussion, and calls the task small, normal or hard.
- Plan02
For normal and hard tasks the code is read and a concrete plan written without changing anything, on a stronger frontier model when the task is hard.
- Code03
The agent follows the plan and pushes the branch; small tasks go straight here. Deckbell opens the pull request.
- Review04
The change is reviewed and the review posted on the pull request, on a stronger frontier model for hard tasks.
Where agents run
On Deckbell, or on your own key.
Choose per agent when you create it, and switch at any time. Both options show the estimated cost per typical task before you decide.
Deckbell
Recommendedin AI credits
- Auto picks a frontier model for each step; you never choose one
- Runs use the AI credits included with every paid seat, shared by the workspace
- Every plan, including Free with its one-time trial credit
- Top up with credit packs that never expire
Your own Anthropic key
billed by Anthropic
- Pro, Business and Enterprise
- Pick one model for every run: Claude Opus 5.5, Sonnet 5.5 or Haiku 5.5
- Your key is stored encrypted; Anthropic bills you at list prices
- No credits used; the run page still shows every step and its cost
- Opus 5.5
- $1.84
- Sonnet 5.5
- $0.92
- Haiku 5.5
- $0.05
Estimates for a typical task: about 1M input tokens (85% read from cache) and 40k output tokens. Own-key figures are at provider list prices, Auto is in AI credits. Real runs vary with the task.
What agents come with on each plan
| Plan | Per seat / month | AI credits | Agent roles | Own key | Schedules & cost dashboard |
|---|---|---|---|---|---|
| Free | $0 | One-time $2 AI credit for agents (never expires) | Coder, Reviewer, Security | — | — |
| Pro | $10 | $5 AI credits per seat / month | All 5 built-in | — | |
| Business | $18 | $15 AI credits per seat / month | All 5 + custom roles | ||
| Enterprise | From $30 | $25 AI credits per seat / month | All 5 + custom roles |
Agents never take a seat on any plan. All plans and prices
Roles
Five roles, and your own on Business.
Every agent has a role that decides what it does and whether it changes code. People @mention an agent by its name, so a team can have a Coder, a Reviewer and a Security agent side by side.
Coder
Opens pull requestsImplements features and fixes, then a pull request is opened.
Every plan
Reviewer
Read-onlyReviews a pull request for correctness and quality.
Every plan
Security
Read-onlyAudits changes for secrets, injection, auth and dependency risks.
Every plan
Tester
Opens pull requestsWrites and fixes automated tests, then a pull request is opened.
Pro and up
Triage
Read-onlyInvestigates an issue: likely cause, what’s affected, a plan and a size.
Pro and up
Custom roles
On Business and Enterprise, write your own: a name, instructions and whether it’s read-only. Custom roles are shared across your workspace.
Business and up
You stay in control
Agents work on your terms.
Review before anything merges, give each agent only the repositories it needs, and put a ceiling on what a run, an agent and a month can cost.
Nothing merges without you
Agents push a branch and Deckbell opens the pull request; merging is yours. Reviewer and Security agents can read every new pull request automatically and post their findings on it and on the linked issue.
Scoped repository access
Connect the Deckbell GitHub App to all repositories or only selected ones, then attach repositories to each agent. Every run gets a short-lived token for just the repository it works in.
A cost limit on every run
Before each step Deckbell checks the credits the run has used. If the next step would pass the limit, the run stops and says which step was skipped and why. A pull request already opened stays open.
Monthly budgets and caps
Owners and admins set a monthly AI-credit budget, choose to stop hosted runs when it’s used up, track spend on your own key, cap each agent per month, and get alerts at 50%, 80% and 100% or thresholds of your own.
A cost dashboard
On Business, see spend by agent, project, role, person and step, this month’s budget and each agent’s cap, success rates and merged pull requests, and how Auto routed tasks of each size.
Schedules
On Business and Enterprise, a cron schedule with a standing instruction starts an agent on its own, for example to check dependencies every night. Scheduled runs use credits like any other and are skipped when the budget is used up.
Works with your tools
Your code host, your chat, your AI tools.
Agents live where the work already is: in the issue, in the channel next to it, and in the pull request on GitHub.
GitHub
Agents clone the repository, push a branch and open the pull request through the Deckbell GitHub App. Pull request cards show the CI result of the latest commit.
GitLab and Bitbucket
Merge requests, pull requests, commits and pipeline status show on the issues they mention. Agents work in GitHub repositories today.
Claude, Cursor and other AI tools
On Business and Enterprise, Deckbell’s MCP server lets an assistant assign an issue to an agent or write “@Coder take this” in a comment, which starts the agent as it does in the app.
Chat
Mention an agent in a channel with an issue key and it replies there, or in the thread; the result is posted on the issue too. Chat with an agent one to one without starting a run.
Documents
Ask for a written result and the agent writes it as a page: a spec, a weekly report, this week’s bugs. Every change it makes is in the page’s version history under its name.
Segment and PostHog
Send issue, cycle and agent run events to your own analytics.
FAQ
Questions teams ask about agents.
The full details are in the agent docs. Anything else? Write to support@deckbell.com and a human replies.
Do agents take a seat?
No. You pay for people; agents never take a seat on any plan. Runs on Deckbell use the AI credits included with each paid seat (shared by the workspace) or credit packs you top up. On Pro, Business and Enterprise an agent can run on your own Anthropic key instead, and Anthropic bills you directly.
Does the agent merge its own pull requests?
No. The agent works on a branch, pushes it, and Deckbell opens the pull request and posts a card on the issue. You review it and decide what merges, as with any teammate. Reviewer and Security agents can read it before you do and post their findings on the pull request.
Which repositories can an agent reach?
Only the GitHub repositories you attach to that agent. Connect GitHub with the Deckbell GitHub App and choose all repositories or only selected ones; each run gets a short-lived token for just the repository it works in.
Which models does Deckbell Auto use?
Agents on Deckbell use Auto: Deckbell picks a frontier model for each step instead of you choosing one. Frontier models write the code, a stronger frontier model plans and reviews hard tasks, and a coding step that fails is retried once on a stronger model. Each step is charged from the tokens it used, in AI credits.
Can I use my own key and pick the model?
Yes, on Pro, Business and Enterprise. Choose "Your own Anthropic key" when you create or edit the agent, paste the key (stored encrypted) and pick one model for every run. The dialog compares the current models by quality, speed and estimated cost per task, and Anthropic bills you directly.
What does a run cost, and how do I keep it under control?
Every run shows its steps, the tokens each used and the cost, on the run page and on the pull request card. Each run has a cost limit and stops before a step that would pass it. Workspace owners and admins can set monthly AI budgets with alerts at thresholds you choose, stop hosted runs at the budget, and cap each agent per month.
How do I start a run?
Assign the issue to the agent, @mention it in an issue comment, in a channel with the issue key, or in a document, or press Run on the Agents tab. The agent replies where it was asked, and the result is posted on the issue. On Business and Enterprise a schedule can start an agent on its own at set times.
Do agents work with GitLab or Bitbucket?
Agents work in GitHub repositories today. GitLab and Bitbucket are connected for linking: merge requests, pull requests, commits and pipeline status show on the issues they mention, and a pull request links itself to every issue whose key is in its branch, title or description.
Can I start an agent from Claude, Cursor or another AI tool?
Yes, on Business and Enterprise through Deckbell’s MCP server. An assistant connected with MCP can assign an issue to an agent or write "@Coder take this" in a comment, which starts the agent just as it does in the app. Everything it changes is recorded in the audit log.
What happens when there is nothing to change?
If the agent finds that the issue is already done, it says so in its reply and Deckbell doesn’t retry, so you aren’t charged twice. If it didn’t push any changes there is no pull request, and the result says so; when a run can’t start or fails, the agent says why.
Give your backlog a crew.
Create a workspace, connect GitHub and assign your first issue to Coder. Free for up to 5 members, with about $2 of trial AI credits. No credit card.