Skip to content

The hosted coding agent

The rest of this section is about connecting your own coding agent to Seylo through seylo-mcp. The hosted coding agent is different: Seylo runs the agent for you, in a sandbox, and you work with it from Build → Coding agent in the portal.

  1. Connect GitHub. Repositories come through the Seylo GitHub App and your own GitHub account; the page lists the repositories both can reach. If there are none in common, Admin → Integrations → GitHub says which side to ask.
  2. Describe a task, in plain words, as you would brief a colleague: the repository, optionally a branch (the default branch otherwise), and optionally the cluster and namespace you may later deploy to. The branch must already exist on GitHub; see when the branch does not exist.
  3. Watch it work. The agent works on a copy of the repository in an isolated sandbox. Its steps appear as it goes, and you steer it with follow-up messages.
  4. Review the diff. The diff is taken after each of the agent’s turns, so you can read it while the sandbox sleeps.
  5. Test and scan. The agent runs the repository’s tests and a security scan in the sandbox, and fixes what they find (see below).
  6. Open a pull request — one commit on a new branch, opened as you, with the agent named as co-author. The agent itself cannot push. The pull request’s description carries the task, the agent’s summary, the test result and the scan result.
  7. Deploy — once your CI has built an image from that branch, deploy the session’s manifests to a connected cluster, into an existing namespace or a new one, or propose moving an existing Deployment to that image tag. It is filed in Approvals; nothing changes on the cluster until someone approves it, unless an admin gave you a deploy allowance that covers it.

A session’s sandbox can be put to sleep, woken and stopped from the session. The page shows how many sandbox minutes your organisation has used this month.

Build → Coding agent is always your workspace, with three panes:

  • Sessions on the left, grouped by This week, This month and Older, each with its status: running, waiting for you, hibernated, done or failed. Collapse it to a rail when you need the room.

  • The conversation in the middle. The agent’s steps are condensed into lines such as “Explored 6 files, Edited 2 files, Ran tests (12 passed)”; open one to see what it did. The model menu sits in the message box and shows each model’s cost as a multiple (×).

  • A panel on the right with Changes (the diff), Files (the repository and a read-only viewer), Terminal (every command the agent ran, with its output), Checks and Preview. Drag to resize it, or expand it.

    To give the conversation the full width, press Hide the details panel at the right end of the session’s header. Show the details panel in the same place brings it back, and so does a status-bar item that opens one of its tabs, such as the last test result. Your browser remembers the choice, so the panel stays hidden the next time you open a session in that browser; another browser or device keeps its own. With no session open, the panel is always shown.

A status bar along the bottom shows the repository and branch, the model, the sandbox state, the last test and scan results, the sandbox minutes used this month, and, while you are on Seylo’s model, this month’s use of Seylo’s model against the amount your seat includes, such as $0.05/$5.00. Hover it for the full sentence. It is marked as a warning from 90% of the included amount, and as spent once it is all used (see choose the model). Use on your organisation’s own key is billed by your provider and is not counted there. If Seylo cannot count this month’s use, the item says so instead of showing a figure.

Click an item to go where it is changed. A value Seylo cannot read says “unknown” and why. On a phone you see one pane at a time; use the Sessions / Chat / Details switcher at the top. The status bar then keeps only the repository, the model and the sandbox.

If something is missing, a setup card in the middle lists it in order. Each step disappears when it is done:

  1. Sandbox not configured on the server. Administrators see the setting names to fix (never their values); everyone else is asked to contact an administrator.
  2. No coding model. Shown only when no model can be used. Seylo’s model is the default, so an organisation usually never sees this step; when it does, an administrator adds your organisation’s own key in the form on the card (see choose the model).
  3. GitHub. Connect your GitHub account, install the Seylo GitHub App on your organisation (administrators), or add repositories.
  4. Seat. The coding agent needs a paid seat for this month, bought for you by an organisation admin (see your seat).

When nothing is missing, the task box appears.

Seylo looks the branch up on GitHub before it starts a sandbox. If the branch does not exist, the session is refused with the error code repo-branch-missing, no sandbox is started, and the page says which of these it is:

  • You typed a branch that does not exist. Leave Branch empty to start from the repository’s default branch (the form names it), or type a branch that exists.
  • The repository has no commits yet. GitHub names a default branch, such as main, even for a repository that has no branch at all. Add a first commit on GitHub (for example, a README), or push one, then start the session again.

If GitHub cannot be asked at that moment, the session is not refused: it starts, and if the sandbox then cannot fetch the branch, the session shows what git reported.

The list holds every repository that both the Seylo GitHub App and your own GitHub account can reach. Search it by owner or name; each row shows the default branch. A repository is missing when the App was not given it, or when your account cannot reach it. Add repositories opens the App’s installation settings on GitHub, where an owner of the organisation grants it more. Each session works on one repository.

After you install the Seylo App on GitHub, Seylo finds the installation by itself, even if GitHub’s redirect does not say which one. If it is not shown yet, open Admin → Integrations → GitHub and press Refresh. If you are not an owner of the GitHub organisation, GitHub sends an owner a request, and Seylo shows Waiting for an owner’s approval beside the organisation until they approve it (an owner approves it on GitHub under Settings → GitHub Apps → Pending requests); Refresh checks again. An installation can be attached to only one Seylo organisation, and only an administrator can attach it.

The sandbox holds only a read-only token for the session’s one repository, so the agent cannot push. When you press Open pull request, Seylo writes one commit on a new seylo/ branch using your own GitHub token and opens the pull request as you, with the agent named as co-author. It opens only once the last security scan covers the current change and nothing in it blocks.

  • Tests. Seylo finds the repository’s test command — a package.json test script, pytest, go test ./..., cargo test or a Makefile test target — or uses one you type. Shell operators such as &&, | and ; are refused. The pass and fail counts come from the test runner’s own summary line; when Seylo does not recognise the format, it shows no counts rather than a guess.
  • Security scan. Open-source scanners run inside the sandbox and look for known vulnerabilities in dependencies, misconfigured Dockerfiles and Kubernetes manifests, insecure code patterns (with Seylo’s own starter rules), and credentials committed in the change. Nothing is sent to a scanning service.
  • The agent runs both itself and is asked to fix what they find, then scan again. You can also run either from the Checks tab between turns.
  • What counts against the change. A finding counts when it is in a file the session changed, or in a lockfile it changed. Findings that were already in the repository are listed separately, never hidden, and do not block.
  • The pull request is gated. It opens only when the last scan covered the current diff. A critical finding blocks it; so does a detected credential. A high finding blocks it unless someone accepts it in writing on the Checks tab, with a reason; the acceptance and its reason go into the pull request.
  • A scanner that did not run is named as not run — on the page, to the agent and in the pull request. It is never reported as clean.
  • No editor. You describe, watch, review a read-only diff and decide. A follow-up message is how a change is steered.
  • No secrets. A task that contains a password or token is refused, because everything in a session is sent to the model. Everything the session shows is redacted first.
  • No credentials in the sandbox. It holds no registry or cluster credential, so it does not build or push images and cannot reach your cluster. Your CI builds the image; the deploy is a proposal.
  • No deploy without a person. A deploy is an approval request, checked like any other change Seylo makes. Under a deploy allowance an admin gave you, your own press of Deploy is the decision; the agent never deploys on its own.

releaser and above: a session ends in a pull request and a deploy proposal, both releaser actions. With only read access to a repository, the agent can work on it, but you cannot open a pull request from it.

Seylo’s model is the default. Every coding agent seat includes $5 of Seylo’s model (GLM-5.3) use per month, so someone with a seat can start a session without setting up a model. The status bar shows how much of it is used this month. The model menu in the message box lists the models you can use and shows each one’s cost as a multiple (×).

On Seylo’s model, one call is limited to 200,000 tokens of context and 32,768 tokens of output.

When the included use is spent, switch to your organisation’s own model key from the same menu, with Bring your own key…. Top-ups of Seylo’s model are not available yet. Tokens on your own key are billed by your provider, never by Seylo.

An admin sets the organisation’s default model and its own key on Build → Coding agent → Model settings. Nightshift → Settings links there too, under The coding agent’s model.

The coding agent’s own key is not the key Nightshift uses. The two are set separately and neither ever falls back to the other, because the coding agent sends your source code to its model, while Nightshift sends only redacted evidence to its own.

An admin adds it from Bring your own key… in the model menu, or on Model settings. Anyone else sees the item disabled, saying an admin adds it.

  1. Choose your provider and paste its API key.
  2. Press Verify key. Seylo asks the provider which models that key can use and lists them.
  3. Pick a model and Save. Saving tests it: Seylo sends one short request through the same client a session uses and shows the answer, or why it failed.

Seylo also shows the model’s context window and output limit, and where they came from: the provider’s own model list, or Seylo’s catalogue copied from the provider’s model pages (with the date checked). If neither knows the model, open Advanced and type both numbers from your provider’s model page; Seylo does not guess them.

Where to get a key:

Provider Where
Anthropic console.anthropic.com → Settings → API keys
OpenAI platform.openai.com → API keys. The models listed are the ones your organisation can use
Google Gemini aistudio.google.com → Get API key
Azure OpenAI / Azure AI Foundry Azure portal → your resource → Keys and Endpoint: copy the Endpoint and Key 1. Under Deployments, copy the deployment name (the name you gave it, not the model’s). Verify checks the deployment and shows which model it runs
Other OpenAI-compatible The https address your API answers on, ending before /chat/completions, plus its key

The key is write-only. It is sealed when you save it and never shown again, by the page or the API, and it is never sent into the sandbox. The page shows only that a key is set and its length. Leave the field empty to keep the stored key; paste a new one to replace it. See how secrets are handled.

A key with whitespace in it, or containing $ or starting with !, is refused: no provider issues one, so it is almost always a paste error.

Remove deletes the organisation’s key. To carry on, pick another model from the menu.

The coding agent is $20 per seat per month, paid before use.

  • An organisation admin buys seats for named people. A seat starts working only once the payment is confirmed.
  • A seat covers the calendar month (UTC) it is bought for and is not prorated: bought on the 20th, it is still $20 and ends at the month’s end.
  • Renewal for the next month opens on the 25th. A seat that is not renewed stops at 00:00 UTC on the 1st, until it is paid.
  • Each seat includes sandbox minutes and $5 of Seylo’s model use per month (see choose the model). Sandbox minutes beyond a seat’s included minutes are not priced yet.
  • A free trial Seylo grants an organisation does not cover coding agent seats.

Each seat payment gets its own invoice, already marked paid. See plans and billing.