AI agent skills that connect your codebase to Autter. Give coding agents repository wiki, learnings, and architecture context through MCP, or wire Autter Runtime — open-source error tracking, diagnostic logs, usage telemetry, and LLM tracing — into any codebase, regardless of language or framework.
Drop these into Claude Code, Cursor, Codex, or any editor that supports the Agent Skills standard and start using them immediately.
npx @autter/skillsIt finds the coding agents on your machine (Claude Code, Cursor, Codex and
others), asks before installing, and prints one short summary. It needs
Node 18 or newer and works offline from the package, so there is nothing else
to download. Add --global to install for your account instead of the current
project, --agent claude-code cursor to choose agents, or --list to see the
skills first.
Then tell your agent what you want:
- "Use Autter's memory, wiki, and learnings to understand this repo before coding." —
runs
autter-repo-context, which reads the signed-in user's memory, the repository's shared memory for the area being changed (what breaks there, what worked and what did not), its wiki, accepted learnings, production lessons, and indexed architecture through the authenticated Autter MCP, and checks the change against production lessons before committing. - "use the skills to install Autter Runtime in this project." — runs
autter-runtime-setup, which inventories your repo and routes each service to the right style skill automatically — you don't need to pick one yourself. - "Why did that request fail?" (while running the app locally) — runs
autter-runtime-logs, which reads the local request/operation records. - "Audit our error handling for Autter Runtime and propose error codes." —
runs
autter-runtime-errors, which reports coverage gaps and proposes a catalog, and edits only after you approve. /restructure-codebase(optionally followed by an area to focus on) — runsrestructure-codebase, a full structural refactor: it lists every hidden link, needless link, piece of surplus code and special case, redesigns from first principles, fixes the whole list, and repeats until nothing worth fixing is left. It only runs when you invoke it.
Want just one skill? npx @autter/skills --skill otel-node-style
Using the open skills CLI instead?
npx skills add Autter-dev/autter-skills still works, but recent versions of
that tool need a newer Node (22.20 or later) and list every installed folder.
| Skill | Covers |
|---|---|
autter-repo-context |
Reads the signed-in user's memory pages, the repository's shared memory (a map, one page per area of the codebase, and a what-works ledger, rewritten nightly from the team's agent sessions, reviews, production, CI and scans) before editing an area, repository wiki Markdown, accepted process learnings and team rules, production lessons for the files being edited, indexed code/architecture/run commands, and Runtime issues, requests and evidence packs through Autter MCP. Checks the diff with check_change before committing and rates the lessons it used. Uses source commits and freshness metadata to ground coding work in how the user works, how the platform is built, and what already broke in production. |
autter-runtime-setup |
Start here. Inventories services, request entry points, queues and error classes, checks installed and deployed capabilities (through get_runtime_setup when Autter MCP is connected), configures the chosen ingestion path, routes built frontends to source-map upload and Next.js apps to server error capture, verifies stored traces/errors with readable stacks, breadcrumbs and trace links, metrics, logs/operations, request summaries, coded errors and LLM calls where configured, then records supported instrumentation conventions in the repo's existing agent-instruction files. Synthetic tests run only in isolation. |
otel-node-style |
Node.js (Express, Fastify, Koa, NestJS) and Next.js, via @autter/runtime-node / @autter/runtime-next, including per-request summaries, error catalogs and coded errors, server error hooks that keep stacks, caught errors counted from logs and runtimeContext.degraded for deliberate fallbacks, Next.js server errors through onRequestError, withAutter(handler) and error-page reporting, readable stacks for bundled servers, background work, local files, testing and logger-only mode. Has a short edge section; otel-edge-style has the full edge setup. |
otel-edge-style |
Cloudflare Workers, Vercel Edge and Next.js middleware.ts, Deno and Bun, via @autter/runtime-edge, including caught errors, rt.degraded and Worker source maps. |
otel-browser-style |
Browser apps — React, Vue, Svelte, Angular, vanilla SPA, static sites — via @autter/runtime-browser, including caught console.error errors, Next.js error pages, source-map upload in four steps (check, autter-runtime sourcemaps setup, an upload-only key stored straight into CI, sourcemaps verify), breadcrumbs, links from browser errors to server traces, coded API errors and request-id correlation, CSP violations and privacy-conscious recent action context. |
otel-python-style |
Django, FastAPI, Flask, Celery workers, plain WSGI/ASGI, via the standard OpenTelemetry Python SDK plus Runtime's copied autter_tracing.py, which keeps every error under head sampling; coded errors and request summaries through OTel attributes. |
otel-go-rust-style |
Go and Rust backends, via each language's official OTel SDK. |
otel-generic-style |
Everything else — Java, .NET, PHP, Ruby, Elixir, Kotlin, … — via standard OTLP/HTTP + OTel env vars. |
autter-runtime-logs |
Local debugging: reads .autter/runtime/*.jsonl (or autter logs --local) to explain failed or slow requests, one request id, and recent coded errors. |
autter-runtime-errors |
Coverage audit: uncovered routes/jobs, plain throws on user-facing paths, swallowing and log-only catches, invalid codes, sensitive context keys. Proposes a catalog; applies only with approval. |
restructure-codebase |
Whole-codebase refactor so agents read less per change, break less without noticing, and debug faster. Asks whether to also prune comments and tests, logs every problem (hidden links, needless links, surplus, special cases) before changing anything, groups them under first-principles fixes, then fixes and re-checks with subagents until nothing worth fixing is left. Invoked manually only. |
| Skill | Version | Runtime it targets |
|---|---|---|
autter-runtime-setup |
1.11.0 | all of the below |
otel-node-style |
1.7.0 | @autter/runtime-node / @autter/runtime-next 1.9.0+ for caught errors, runtimeContext.degraded, withAutter(handler) and error pages (1.6.0 for error hooks, 1.5.0 for request summaries and codes, 1.4.0 for operations only); @autter/runtime-cli 1.3.0 for server source maps |
otel-edge-style |
1.2.0 | @autter/runtime-edge 1.0.0+ (1.3.0 for request counts, 1.4.0 for caught errors and rt.degraded), @autter/runtime-next/edge 1.5.0+ |
otel-browser-style |
1.5.0 | @autter/runtime-browser 1.6.0+ (1.8.0 for caught console.error errors, 1.4.0 for coded errors and request ids) |
otel-python-style, otel-go-rust-style, otel-generic-style |
1.4.0 | OTel SDKs + ingester 1.5.0+ |
autter-runtime-errors |
1.1.0 | runtime-node/next 1.5.0+ (1.9.0 for the logged-catch fix) |
autter-runtime-logs |
1.0.0 | runtime-node/next 1.5.0+ |
Self-hosted ingesters need 1.5.0+ for request summaries and code grouping, 1.6.0/1.7.0 for source maps and upload keys, 1.10.0 for edge request counts, and 1.11.0 for caught errors, server source maps, root-cause grouping and route percentiles.
Connect your agent to https://api.autter.dev/mcp with OAuth and select the
organization containing your repository. Repository knowledge reads require
mcp:read; they do not require a Runtime ingest key or an SDK install.
check_change also needs mcp:llm, and rate_lesson needs mcp:write.
autter-repo-context uses whoami to match the repository, get_memory for
the user's own notes and the pages other members shared, get_repo_memory and
search_context for the repository's shared memory of the area it is about to
change (update_repo_memory_page, which needs mcp:write, records what it
learned), get_wiki for stored
pages, get_learnings for accepted conventions, get_file_context and
get_lessons for the production lessons in reach of the files it edits,
check_change on the diff before committing, and get_repo_index for
implementation evidence. It checks stored freshness against the current
checkout and reports missing sources explicitly.
With the Autter CLI installed,
autter install also puts repository memory into the session automatically in
Claude Code, Codex, Gemini CLI, Droid, GitHub Copilot, OpenCode, Amp and Pi (the
map at session start, the matching sections with each prompt, an area's page
when a file in it is read). The skill covers every other agent, and tells all of
them to read an area's page before changing code in it; autter memory prints
the same memory in a terminal.
The get_wiki tool requires a backend deployment that includes it. Refresh
the client's MCP tool list after that deployment. If it is unavailable, the
skill continues with exposed learnings/index tools and local source code.
Autter Runtime's ingester uses these key types and HTTP endpoints:
| Credential | Lives in | Can |
|---|---|---|
Server key autter_rt_… |
backend env vars | send OTLP traces/metrics/logs, relay browser events |
Client key autter_rtc_… |
frontend bundles (publishable) | send browser events only, origin-restricted |
Upload key autter_rtu_… |
CI secrets | upload and check source maps only (ingester 1.7.0+) |
| Endpoint | Format |
|---|---|
POST /v1/traces, POST /v1/metrics |
OTLP/HTTP — protobuf or JSON |
POST /v1/logs |
OTLP/HTTP logs — protobuf or JSON; requires the operation-logging ingester release |
POST /v1/browser |
compact JSON (@autter/runtime-browser payload) |
POST /v1/profiles |
symbolized pprof (server key only) |
POST /v1/sourcemaps |
release-keyed source map JSON with an optional debug id (upload or server key) |
POST /v1/sourcemaps/check |
which maps are already stored, so uploads skip them (upload or server key, 1.6.0+) |
GET /v1/compat |
ingester version and supported features (public) |
POST /v1/platform-events |
ECS/Kubernetes OOM and restart JSON (server key only) |
Any language with an OpenTelemetry SDK can send server telemetry — that's
every mainstream language. Node.js, edge runtimes and the browser get dedicated
first-party npm packages (@autter/runtime-node, @autter/runtime-next,
@autter/runtime-edge, @autter/runtime-browser); everything else is a thin style guide over the
standard OTel SDK for that language, which otel-generic-style covers even
when no dedicated skill exists yet.
Memory pressure uses that same OTLP metric endpoint in every backend language. The Node package exports process metrics automatically; other stacks enable a process meter or add the portable RSS/heap gauges described in Runtime's memory contract. The feature also needs the 1.3.3+ ingester and Autter backend/frontend deployment. Applications must be redeployed after instrumentation changes; OOM/restart correlation requires an ECS/Kubernetes event forwarder.
Create one on your Autter dashboard: Settings → Access Tokens → Runtime ingest keys → Create key. Pick the repository, then choose:
- Server — secret, for backends and relays. Keep it in env vars, never commit it.
- Client — publishable, for browser-only apps with no backend. Scoped to the exact origins you register it for.
For source map uploads in CI, create an upload-only key in the repository's Settings → Runtime → Source map upload key, or let the agent create one through Autter MCP.
Set the key in your own environment (shell profile, gitignored .env, or
secret manager) as AUTTER_RUNTIME_KEY — don't paste key values into
the agent chat. The skills are written to only ever reference the env
var by name; the value never needs to reach the agent.
The setup skills cover request histograms, release identity, and trace comparison requirements. Node/Next.js examples use SDK 1.3.2 and opt into bounded slow-request retention. External OTel exporters need explicit-bucket delta histograms. Self-hosted ingesters require 1.3.1 or later.
Enabling detection in the platform does not update customer SDKs. Incident feedback lets a team mark a diagnosis as correct, incorrect, or expected. Fixes open as draft pull requests labelled Verified, Builds or Proposal only, and a person merges them, unless the repository turned on auto-merge for verified, non-breaking fixes (off by default). Production verification uses existing telemetry; selftests belong only in isolated test environments.
Full docs: github.com/Autter-dev/autter-runtime.
The setup skills check installed SDK exports and deployed ingestion before
using runtimeLogger, createRuntimeLogger, or withRuntimeOperation.
Operation logging requires Node/Next.js SDK and ingester 1.4.0+; an existing
latest install or 1.3.x range alone does not establish support. The ingester needs
/v1/logs and migration 0011-runtime-logs; the platform backend, fix worker
and frontend need the corresponding evidence deployment. Release ingester
first, SDKs next, then platform consumers, and redeploy customer applications.
Node/Next.js setup adds stable operation names, nested context, measured steps
and explicit business outcomes to selected checkout/job paths. It preserves
existing loggers, separates diagnostics from issue capture, and checks flushing
and stored evidence. Other backend languages use their OTel logs exporter and
logger bridge. Browser capture keeps its own APIs and endpoint; edge runtimes
use @autter/runtime-edge (see otel-edge-style).
Read Node operation setup or customer logging docs. Logs improve the evidence available to RCA and eligible draft fixes; they do not prove a diagnosis or a working fix. Delivery is best effort.
Runtime's Requests & logs view shows one summary per HTTP request and
per-route counts. It stays empty until the service mounts a request boundary,
so the setup skills add one by default for every Node, Next.js and edge HTTP
service: autterRequests() for Express, Connect and NestJS; autterFastify
for Fastify; withRuntimeRequest() for fetch-style and Next.js route
handlers; and withAutter for edge runtimes. Health, readiness and metrics
paths are ignored, because summaries are never sampled. This needs SDK
1.5.0+ (runtime-edge 1.0.0+) and ingester 1.5.0+, which the skills
check through the public /v1/compat endpoint. Existing error handlers and
response formats are left alone unless the user agrees to
autterErrorResponse(). See
request summaries setup.
Runtime 1.6.0 makes errors easier to diagnose, and the setup skills wire it by default where it applies:
- Source maps. Every frontend that ships built JavaScript gets
@autter/runtime-clias a dev dependency: the Vite/Rollup or webpack plugin,withAutterSourceMapsfor Next.js, or theautter-runtime sourcemaps inject/uploadcommands and theupload-sourcemapsGitHub Action for anything else. Maps stay hidden, carry debug ids, and upload from CI or build secrets, never from the browser bundle. Without them, browser stacks point at minified chunks that RCA and fix agents cannot use. The agent does it in four steps: check (get_runtime_setupover Autter MCP, or/v1/compat, andnpm view @autter/runtime-cli version),autter-runtime sourcemaps setup, an upload-only key (autter_rtu_…) fromcreate_sourcemap_upload_keypiped straight into the CI secret (or one the user creates) and named by prefix only, thenautter-runtime sourcemaps verifyon a production build. With a CLI older than 1.2.0 they wire it by hand; until the CLI is published, they wire nothing and report source maps as pending its release. - Server errors.
autterErrorHandler()(Express/Connect),autterFastifyErrors,autterKoa()andcaptureRequestErrorrecord a thrown error with its stack instead of a stackless 500, without changing existing error handlers. Next.js apps getonRequestErrorfrom@autter/runtime-next/instrumentationandserverExternalPackages: ["@autter/runtime-node"], so routes share the SDK thatregisterAutterstarts. Unhandled promise rejections are recorded by default. - Breadcrumbs and trace links (opt-in, offered with a recommendation):
enableBreadcrumbs()attaches the last 30 steps before a browser error, scrubbed like the rest of the event.enableTracePropagation()links a browser error to the server trace of the same request; cross-origin APIs must allow thetraceparentheader in CORS.
This needs SDK 1.6.0 (runtime-edge 1.1.0, runtime-cli 1.2.0 for
setup and verify) and ingester 1.6.0 (1.7.0 for upload keys and
verify's ingester check), checked through /v1/compat (breadcrumbs,
source_map_debug_ids, fingerprint_override,
browser_trace_propagation, sourcemap_upload_keys, source_map_verify). See
source maps setup and
server errors setup.
Runtime 1.11.0 extends this:
- Caught errors count. An
Errorpassed toconsole.error, a logger orruntimeLogger.errorbecomes a handled occurrence that never fails the request (captureLoggedErrors, on by default; runtime-node/next 1.9.0, runtime-edge 1.4.0, runtime-browser 1.8.0). Acatchthat falls back on purpose says so withruntimeContext.degraded(err, reason), which records awarningthat never pages and ends the requestdegraded. - Next.js.
withAutter(handler)answers a thrown error with its own status (offered, since the response changes), anduseAutterErrorReport,reportNextErrorandAutterGlobalErrorreport what error pages show (wired by default). - Server source maps.
@autter/runtime-cli1.3.0 uploads maps for bundled servers (tsup, esbuild, Vite SSR, webpack Node targets,.next/server, Workers); the plugins cover server output automatically, andsourcemaps verify --serverchecks it.
Runtime 1.5.0 also adds coded errors (RuntimeError, defineRuntimeErrors)
that group into one issue per code across the SDK and connected providers.
Error responses carry the code and requestId, and the browser links its
failures to the server request through x-request-id.
The setup skills check installed exports before using these APIs. Required
versions: runtime-node/next 1.5.0+, @autter/runtime-edge 1.0.0+,
runtime-browser 1.4.0+, and ingester 1.5.0+. A latest tag or a
version string in source does not prove support. Roll out the ingester first,
then SDKs and application redeploys, then platform consumers. Python, Go, Rust
and other stacks send the same attributes (autter.error.code/why/fix/link/expected,
autter.operation.kind=request, autter.request.id) through their OTel SDKs.
Codes are namespaced, stable and never contain ids, PII or secrets.
Application-declared why/fix are shown as declared, and Autter's analysis
treats them as hypotheses. Business failures marked expected are counted but
never open incidents. See error catalogs
and operation logging.
MIT
autter-runtime-setup also supports repository-scoped Sentry, PostHog, Grafana/Loki, Datadog and webhook sources. Ask the agent to connect the chosen source in repository Runtime settings and verify stored logs, RCA and eligible draft fixes. Connector-only setup does not require an application SDK or a Runtime ingest key. See external-source guidance.