
tierzero
Agentic production engineering for SWE, SRE and DevOps. Resolve and investigate issues against existing observability, CI/CD, and self-improving knowledge bases.
- version
- 1.3.0
Install
/plugin install tierzero@tierzeroRun it inside Claude Code after adding the marketplace with /plugin marketplace add TierZeroAI/tierzero-claude-code.
Contents
Discovered from sourceSkills 3
TierZero: Fetch Context
Fetch conversation, investigation, or artifact data from TierZero by URL or artifact UUID, so the agent can act on it.
Supported inputs
| Type | Example |
|---|---|
| Chat URL | https://app.tierzero.ai/chat/c/<GlobalID> |
| Investigation URL | https://app.tierzero.ai/investigations/<GlobalID> |
| Artifact UUID | bf904904-afdc-4cf2-94d8-76a4a8bb4f75 |
How to invoke
This plugin registers TierZero as an MCP server (tierzero). Call the MCP tool directly — do not shell out to curl.
Tool: tierzero_fetch_context
Arguments:
url(string, required): the chat URL, investigation URL, or artifact UUID.include_sources(bool, optional, defaultfalse): settrueonly when the user asks for source-by-source breakdowns or you need to cite specific log lines / traces.
After the call
Parse the response and surface what's relevant for the current task:
- Conversations — walk the messages (use
messageTypeandoutput); summarize the conclusion, then note any open questions. - Investigations — present the
outputJsonresult; quote findings, list evidence sources. - Artifacts — show the artifact
valueandtype; if it's a log/metric snapshot, summarize what it contains.
Only include source metadata if the user explicitly asks for it (and call with include_sources: true). Otherwise it's noise.
Authentication
OAuth is automatic.
If an authorization URL surfaces (because the user hasn't authed yet, or the token expired), do NOT print the URL and ask the user to copy it. Claude Code's terminal wraps long URLs across multiple lines, splicing newlines into the copied text and breaking the paste. Instead, run it through the system browser opener:
- macOS:
open '<authorization_url>' - Linux:
xdg-open '<authorization_url>'
Quote the URL — the OAuth params contain & and = which the shell will mangle without quotes. After running, tell the user "Browser opened — log in to TierZero and consent. The MCP server will reconnect automatically once auth completes."
If a tool call fails with an authentication error and you don't have an auth URL in hand, tell the user to run /mcp and re-authorize the tierzero server.
TierZero: Production Telemetry Skill
Query production telemetry data through TierZero's MCP server to inform debugging, incident investigation, feature design, and code review.
When to trigger
TRIGGER when: debugging a production issue, investigating an incident or alert, designing a new feature, refactoring code where you need to know who/what calls the affected paths, reviewing a PR where production behavior would change your verdict, or any task where understanding real production behavior (traffic patterns, error rates, latency, usage) would inform a better decision. Use proactively — don't wait to be asked.
DO NOT TRIGGER when: writing pure unit tests, updating documentation, or tasks with zero production relevance.
How to invoke
This plugin registers TierZero as an MCP server (tierzero). Call the MCP tools directly — do not shell out to curl or wrap them in scripts.
Primary tool: tierzero_ask
- Pass a single
queryargument (plain English) describing what you want to know. - Be specific: name the service, give a time window, describe the symptom.
- The tool runs the investigation server-side and returns the grounded answer with evidence sources.
Writing good queries
A good query for tierzero_ask:
- Names the service or component (e.g.
payment-service, not "the API") - Specifies a time window (e.g. "last 30 minutes", "since 14:00 UTC", "during the deploy at 13:42")
- States the symptom (errors, latency, drop in traffic, unexpected log line)
- Mentions the integration if you know it (Datadog, Sentry, GitHub Actions, etc.)
Examples:
What errors occurred in payment-service in the last hour?Show p95 latency for /api/checkout over the last 24h and flag anomalies.Did any GitHub Actions workflows for the api repo fail today?Has the rate of 5xx on auth-service changed since the deploy at 13:42 UTC?
After the call
- Summarize findings tied back to the user's current task
- Cite the integration the data came from (Datadog logs, Sentry issue, etc.) when relevant
- If the result returns a TierZero investigation or chat URL, surface it so the user can drill in
- If the result is empty or an integration isn't connected, say so and suggest what to connect
Authentication
OAuth is automatic. On first use Claude Code will prompt you to authorize the TierZero MCP server in a browser — log in, consent, done. The token persists in Claude Code.
If an authorization URL surfaces (because the user hasn't authed yet, or the token expired), do NOT print the URL and ask the user to copy it. Claude Code's terminal wraps long URLs across multiple lines, splicing newlines into the copied text and breaking the paste. Instead, run it through the system browser opener:
- macOS:
open '<authorization_url>' - Linux:
xdg-open '<authorization_url>'
Quote the URL — the OAuth params contain & and = which the shell will mangle without quotes. After running, tell the user "Browser opened — log in to TierZero and consent. The MCP server will reconnect automatically once auth completes."
If a tool call fails with an authentication error and you don't have an auth URL in hand, tell the user to run /mcp and re-authorize the tierzero server.
TierZero: Teach a Fact
Save durable knowledge to TierZero's Context Engine so future investigations and answers — for the whole team — can use it.
When to trigger
TRIGGER when: the user explicitly asks to save, remember, or teach something to TierZero; the user says "remember this for next time"; or you've uncovered a durable fact during an investigation that the user confirms is worth retaining.
DO NOT TRIGGER when: the information is ephemeral (a one-off metric value, a transient error count), the user hasn't confirmed they want it saved, or the content is speculative / unverified.
After completing an investigation with tierzero_ask, if you uncovered durable facts or debugging techniques, proactively suggest saving — but always get user confirmation first. Frame it as: "I found that [X]. Want me to save this to TierZero's Context Engine so your team has it for next time?"
How to invoke
This plugin registers TierZero as an MCP server (tierzero). Call the MCP tool directly — do not shell out to curl or wrap it in scripts.
Tool: tierzero_save_snippet
Arguments:
content(string, required): The snippet body — clear and self-contained. Include exact names, identifiers, queries, or commands. Write it so someone with no chat context understands it.scope(string, required): When/where it applies. Describe the service, system, or situation. Normalize from conversation context.memory_type(string, optional, default"FACT"):"FACT"— observable information about the systems (service/infra details, ownership, configuration)."BEST_PRACTICE"— recommended approaches, conventions, or debugging procedures.
Writing good snippets
Examples:
content: "payment-service uses a connection pool of 20 to postgres-primary. Under sustained load (>500 rps) the pool exhausts and requests queue, causing p99 > 2s." /scope: "payment-service / postgres connectivity" /memory_type: "FACT"content: "To debug Kafka consumer lag in order-processor, check consumer_lag_records in Datadog grouped by partition. Single-partition lag usually means a slow handler, not a broker issue." /scope: "order-processor / Kafka debugging" /memory_type: "BEST_PRACTICE"
Avoid saving ephemeral data ("142 errors in the last hour"), vague statements ("the service is broken"), or generic advice ("check Datadog").
After the call
The tool reports the real outcome — relay it honestly:
- "Saved to memory…" — confirm and share the Context Engine link. Note it becomes available after backend processing (a few minutes).
- "Not saved: …" — tell the user why (empty, too long) and suggest a fix. Do not claim it was saved.
- "ERROR: …" — tell the user and suggest retrying.
Authentication
OAuth is automatic.
If an authorization URL surfaces (because the user hasn't authed yet, or the token expired), do NOT print the URL and ask the user to copy it. Claude Code's terminal wraps long URLs across multiple lines, splicing newlines into the copied text and breaking the paste. Instead, run it through the system browser opener:
- macOS:
open '<authorization_url>' - Linux:
xdg-open '<authorization_url>'
Quote the URL — the OAuth params contain & and = which the shell will mangle without quotes. After running, tell the user "Browser opened — log in to TierZero and consent. The MCP server will reconnect automatically once auth completes."
If a tool call fails with an authentication error and you don't have an auth URL in hand, tell the user to run /mcp and re-authorize the tierzero server.
Commands 2
/investigate
Run a TierZero investigation against connected production systems.
The arguments after /investigate are the user's natural-language query.
Steps
- Treat everything the user typed after
/investigateas the investigationquery. - If the query is empty, ask the user what they want to investigate, then stop.
- Call the
tierzero_asktool on thetierzeroMCP server with that query as thequeryargument. - Wait for the response, then summarize the findings tied to the user's current work.
- Surface any TierZero investigation URL returned in the result so the user can drill in.
- If the call fails with an authentication error, run
/mcpand re-authorize thetierzeroserver, then retry.
Tips for a good query
- Name the service or component, not "the API"
- Include a time window ("last hour", "since the 13:42 UTC deploy")
- State the symptom (errors, latency, drop in traffic)
Examples
/investigate what errors hit payment-service in the last hour?/investigate p95 latency for /api/checkout over the last 24h, flag anomalies/investigate did any github actions runs for the api repo fail today?
/teach
Save a snippet to TierZero's organizational memory (Context Engine).
The arguments after /teach are the knowledge to save.
Steps
- Treat everything the user typed after
/teachas the snippetcontent. - If the content is empty, ask the user what they want to save, then stop.
- Infer an appropriate
scopefrom the conversation context — describe when or where this knowledge applies. - Choose a
memory_type: use"FACT"for observable details about systems (default), or"BEST_PRACTICE"for recommended approaches or conventions. - Call
tierzero_save_snippeton thetierzeroMCP server withcontent,scope, andmemory_type. - Relay the outcome honestly — only tell the user it was saved if the tool confirms it.
- If the call fails with an authentication error, run
/mcpand re-authorize thetierzeroserver, then retry.
Examples
/teach payment-service uses a connection pool of 20 to postgres-primary and exhausts under sustained load above 500 rps/teach To debug Kafka consumer lag in order-processor, check consumer_lag_records in Datadog grouped by partition/teach auth-service rate limiter is configured at 100 req/s per IP in the edge-proxy envoy.yaml, not in the service itself
MCP servers 1
Sponsors
Details
- Marketplace
- TierZeroAI/tierzero-claude-code
- Source
- plugins/tierzero
- Author
- TierZero
- Category
- Ai
- License
- MIT
- Keywords
- observability, debugging, incident-response, telemetry