Skip to content

Troubleshooting coding agents

seylo-mcp writes its messages so that a coding agent can act on them, and most agents show them to you as they are. Find the message’s first words in the tables below.

Start here: ask the agent to call list_targets. If that works, the token and the connection are fine, and the problem is with one tool. If the agent’s question is about observability, have it call observability_status next.

When the token is rejected, every cause gets the same message, so a caller holding someone else’s token learns nothing from it:

This deploy token is not valid for Seylo. It may have been revoked, expired, or the namespace it was issued for may no longer be open to coding agents.

The JSON-RPC error carries the exact cause in data.reason:

data.reason HTTP Cause What to do
not-a-deploy-token 401 No Authorization: Bearer header, or the value doesn’t start with seylo-deploy_ Check the client’s header. A user’s sign-in session is not accepted
unknown-token 401 No such token. Often a copy that lost characters Paste the token again. If it is lost, issue a new one
token-revoked 401 An admin revoked it Ask for a new token
token-expired 401 It is past its expiry An admin renews it if it expired within the last 30 days. Otherwise, issue a new one
namespace-not-agent-deployable 403 The namespace was closed to coding agents, or the cluster’s policy no longer permits it An admin opens it again, if that is intended
cluster-gone 403 The token’s cluster was removed from Seylo Issue a token for a cluster that exists

Other connection-level answers:

Message HTTP Cause What to do
This MCP endpoint is POST-only. 405 The client sent a GET, usually to open an SSE stream Use the Streamable HTTP transport, not SSE. The endpoint offers no stream
Origin not allowed. 403 The request carried a browser Origin Call the endpoint from a coding agent or a server, not from a web page
Unsupported MCP protocol version "…". 400 The client asked for a version Seylo doesn’t speak Supported versions are 2026-07-28, 2025-11-25 and 2025-06-18. Update the client
Expected a single JSON-RPC request object. 400 The client sent a JSON-RPC batch, or a body that isn’t an object Send one request per call
Method "…" is not implemented by this server. 200 The client asked for resources, prompts, sampling or roots Nothing to fix. seylo-mcp offers tools only
The request could not be completed. 500 An internal error Retry. Seylo never includes internal detail in this message
Message Cause What to do
No tool named "metrics_query" is available to this token. (any tool name) The token doesn’t carry the capability for that tool Issue a token with the capability ticked. See capabilities
The observability tools don’t appear at all The token was issued before 2026-10-04, when observe didn’t exist. A token never gains capabilities Rotate the token. New tokens include observe by default
A capability added by rotating still isn’t listed The client cached the old tool list, for up to five minutes Reconnect the client, or wait
This token does not carry the "…" capability. As above As above
The free trial ended on …, so changes are blocked. A free trial Seylo granted your organisation has ended. Tools that change something are refused, and reads still work Your plan and usage are under Admin → Billing; write to support@seylo.in
This organisation is suspended, so changes are blocked. The organisation is suspended Contact your Seylo account owner

Every tool call counts against one of two per-token limits, and a call over the limit is refused as a tool error. The message names the family, the limit and how long to wait. The two families are counted separately, so hitting one doesn’t affect the other.

Message begins Cause What to do
Rate limited: this token has made 60 read calls in the last minute, the most allowed for the read family The token made 60 calls to tools that change nothing, such as metrics, logs, diagnose, validate or operation_status, within a minute Wait the number of seconds the message gives (Retry after …s). Combine questions, for example with one sum by (pod) instead of one query per pod. Writes still work meanwhile
Rate limited: this token has made 10 write calls in the last minute, the most allowed for the write family The token made 10 calls to deploy, rollback, put_config, request_secret, dashboard_propose or scrape_target_propose within a minute. A write the tool refused counts too Wait the number of seconds the message gives. Don’t retry a refused write unchanged: read its reason first. Reads still work meanwhile, and the daily cap still applies

deploy and rollback answer with status: "refused" when their checks fail, and nothing is changed. The rest of the answer says which check.

Message begins Cause What to do
This token has used its daily cap of 20 deploys. It resets at 00:00 UTC. (your cap) The daily cap is used up Wait for 00:00 UTC, or have an admin issue a token with a higher cap
This token has used its daily cap and you expected room The cap counts deploys that were applied. Each applied deploy or rollback counts once; a refused or failed attempt counts zero Check the token’s … deploys today badge on the Coding agents card
Namespace "shop-dev" has no ResourceQuota, so there is no ceiling on what a deploy from here can consume. Deploys into a namespace with no quota are refused. So are deploys where Seylo can’t list the namespace’s quotas. The message goes on: An administrator sets one in the Seylo portal under Clusters > Coding agents > "Set a budget" next to this namespace; retry the deploy once it is applied. An admin sets a budget, then the agent retries
Nothing was applied: Deployment shop-dev/checkout is managed by Argo CD. (or Flux, Helm, an operator) Another controller owns an object the deploy would write. The message goes on: A coding-agent token cannot acknowledge a foreign owner - only a person can, in the Seylo portal, on the record. Make the change in that owner’s source, such as the Git repository or Helm values, or ask a person to apply it from Seylo with the owner acknowledged
Nothing was rolled back: Deployment shop-dev/checkout is managed by Argo CD. (or Flux, Helm, an operator) Another controller owns the Deployment the rollback would change. As with a deploy, a coding-agent token cannot acknowledge a foreign owner Roll back in that owner’s source, for example by reverting the commit Argo CD or Flux watches, or ask a person to roll back from Seylo with the owner acknowledged
Cannot deploy: … blocking policy finding(s) A policy rule blocks the generated manifests Fix the findings listed under policy.yours, then call validate again
Cannot deploy: … resource(s) the cluster refused The cluster’s server-side dry run rejected a resource, for example through Pod Security Admission, an admission webhook or a quota Read dryRun.issues. The cluster’s own words are there
image "…" must carry an explicit tag or digest, and must not be ":latest". A mutable image tag Use the immutable tag or digest your build produced
Refusing env "…" (…): it looks like a credential A value in the spec’s env looks like a password, key or token. An env value is stored in plain text in the Deployment and on the operation Leave it out of env. Have the agent call request_secret so a person creates the Secret (or use an ExternalSecret or SecretProviderClass for a vault), then deploy with propose_deploy and manifests that read it through secretKeyRef. Rotate the value: the agent has already seen it
Unsupported field(s): The spec has a field the deploy spec doesn’t support See the deploy spec. Security settings, service accounts, host networking and volumes can’t be set
Namespace "…" on cluster "…" is not open to coding-agent deploys. The namespace was closed An admin opens it again, if intended
There is nothing to roll back to: The workload has fewer than two recorded revisions Nothing to undo
No revision matching "…" in the history of toCommitSha isn’t in that workload’s history Use one of the revisions the message lists
That is the revision currently deployed. Nothing to do. The target is already current Nothing to fix
No operation "…" in this namespace. The operation id is wrong, or the operation belongs to another namespace Use the operationId that deploy or rollback returned

Applied is not running. If deploy succeeds, the manifests were applied, but the rollout may still fail. Have the agent poll operation_status, then call diagnose. If a tasks/cancel was sent, it stopped the polling and didn’t stop the deploy.

Message begins Cause What to do
Refusing to write "…" into a ConfigMap: The value looks like a credential. A ConfigMap is unencrypted and readable by anything with access to the namespace Have the agent call request_secret, then reference the Secret by name
Nothing was written: ConfigMap shop-dev/app-config is managed by Helm. (or Argo CD, Flux, an operator) Another controller owns the ConfigMap. A token can’t acknowledge it Change the ConfigMap in that owner’s source, or ask a person to change it from Seylo
reason must explain what the application does with this secret request_secret was given a reason shorter than 10 characters Explain the use. A person reads the reason
keys must name at least one key the application will read. request_secret was given no key names Name the keys
Message begins Cause What to do
… A deploy token is always limited to its one namespace, so this applies to every token on this cluster. Seylo hasn’t proved that this signal’s store filters by namespace, so it refuses every token A person re-runs the check on Admin → Integrations → Observability stores. observability_status shows the result
metrics_query does not take "namespace". (any tool) The agent tried to set its own namespace or cluster Nothing to fix. Seylo adds the token’s namespace to every query
The query is … characters; at most 4000. The query is too long Split it, or aggregate
That window is … days; at most 30. The window is too long Narrow it
… is more than 11000 points per series. Use a step of at least The step is too small for the window Use the step the message gives, or a shorter window
The … answer is … MiB; Seylo relays at most 4 MiB. The answer is too large Narrow the range, raise the step, lower the limit, or aggregate with sum by
Send either since (e.g. "1h") or start/end in unix seconds, not both. Both kinds of window were sent Send one
logs_stats needs a The query has no | stats pipe Use logs_query to read lines

No data is not the same as zero. A metric with no namespace label, such as node CPU, never appears for a token. query_catalogue and dashboard_templates mark the entries this affects. traces_services returning nothing means no spans arrived from this namespace. The app may not be instrumented, or may send its spans somewhere else.

See what is refused before it reaches a person.

Message Meaning
The container is running and has written nothing to stdout or stderr yet. The read worked. There is nothing to show yet
No previous instance, so there is nothing to read - this container has not restarted. previous: true was asked of a container that never restarted

Problems when opening a namespace, issuing or renewing a token

Section titled “Problems when opening a namespace, issuing or renewing a token”

These appear in Seylo, not in the coding agent.

Message begins What to do
These namespaces cannot be agent-deployable because this cluster's own policy already excludes them: Add the namespace to Allowed namespaces, or remove it from Denied namespaces, first
The in-cluster agent has no RBAC in: Add the namespace to the agent’s watch list, re-run helm upgrade, then open it
Connect a coding agent is disabled, or the card reads No namespace is open to coding agents Open a namespace first
ttlDays must be an integer between 1 and 90. Choose an expiry from 1 to 90 days
This token was revoked, and a revoked token cannot be renewed A revoked token can’t be renewed. Issue a new token for the same namespace
This token expired more than 30 days ago, so it can no longer be renewed. Issue a new token, and choose its capabilities again
This token was revoked or renewed a moment ago, so it was not renewed again. Someone renewed it first. Refresh the list to see its replacement
dailyDeployCap must be an integer between 0 and 1000. Choose a cap from 0 to 1000
A label of 1-80 characters is required Give the token a label

“The key works, but this OpenRouter account’s guardrails or privacy settings block the chosen model”

Section titled ““The key works, but this OpenRouter account’s guardrails or privacy settings block the chosen model””

Your key is fine. OpenRouter refused the model because of your OpenRouter account’s own settings. Allow the model in OpenRouter’s guardrails (openrouter.ai, your workspace’s Guardrails) and data policy (openrouter.ai Settings → Privacy), or choose another model. OpenRouter’s own message is shown below Seylo’s, with your key removed.