# anthropic cheatsheetEvery export of @deckyard/anthropic. Guides are listed in the introduction .
jennifer import "@deckyard/anthropic/" as anthropic;Requests, replies, blocks and tools are json.Value documents in the API's own shape. A refused request throws AnthropicError (kind anthropic) with the HTTP status, Anthropic's error type, the requestId, the wait the server asked for and its rate-limit headers; so does a request the deck refuses to send (status 0). A reply over the cap is kind "limit", a deadline that passed "timeout", a connection the server closed after part of its answer "closed"; a network failure is the bundled http's net.NetError (kind "net"), passed on unchanged; a cancelled task stays "cancelled". A c argument is a Client, a req a Messages request, a reply what chat returned.
# ConstantsConstant Value Meaning VERSION"0.1.0"this deck's version DEFAULT_BASE_URL"https://api.anthropic.com"where the API is served API_VERSION"2026-10-21"the anthropic-version a fresh client sends DEFAULT_MODEL"claude-opus-5-5"the model the docs and examples use DEFAULT_TIMEOUT_MS600000how long one attempt may take (10 minutes) DEFAULT_RETRIES2retries of a 408, 409, 429, 5xx or a failed connection MAX_RETRY_WAIT_MS60000the longest retry-after-ms / retry-after waited for DEFAULT_MAX_BYTES67108864the largest reply read (64 MiB) DEFAULT_MAX_TOOL_INPUT1048576the largest tool input (its JSON) the tool loop hands to a handler (1 MiB) DEFAULT_MAX_TOKENS16000a max_tokens that suits a reply that is not streamed BETA_FALLBACKS"server-side-fallback-2026-07-01"the beta for withFallbacks MAX_BATCH_REQUESTS100000the most requests in one batch
# The clientFunction Returns Does client(apiKey)Clienta client for the Claude API with an API key withBaseUrl(c, baseUrl)Clientanother API root: a gateway, a proxy, a fake withPlainHttp(c)Clientallow plain http to another machine (the key crosses in clear) withTls(c, tls)Clientcertificate options (caCert, skipVerify) withApiVersion(c, version)Clientanother anthropic-version, such as 2023-06-01 withBetas(c, betas)Clientthe anthropic-beta names sent with every request withTimeout(c, ms)Clientthe deadline of one attempt, 1000 to 86400000 ms withRetries(c, retries)Clienthow often a retryable failure is retried, 0 to 10 withMaxBytes(c, most)Clientthe largest reply, 65536 to 1073741824 bytes withMaxToolInput(c, most)Clientthe largest tool input the loop hands to a handler, 1024 to 1073741824 bytes call(c, method, path, body)json.Valueany other endpoint; body JSON null for none models(c)list of json.Valueevery model the key may use, newest first model(c, id)json.Valueone model: context window, output cap, capabilities
# ClientField Meaning apiKey as stringthe API key, sent as x-api-key baseUrl as stringthe API root without a trailing slash version as stringthe anthropic-version sent betas as list of stringthe anthropic-beta names sent tls as http.TlsOptionscertificate checks for https allowPlainHttp as boolwhether plain http to another machine is allowed timeoutMs as inthow long one attempt may take retries as inthow often a retryable failure is retried backoffMs as intthe first wait before a retry without retry-after, doubled each time maxBytes as intthe largest reply read maxToolInput as intthe largest tool input, in bytes of its JSON, the tool loop hands to a handler
# AnthropicErrorField Meaning kind as stringalways "anthropic" message as stringanthropic: <request>: <why> (HTTP <status> <type>, request <id>)status as intthe HTTP status; 0 when no reply was read type as stringAnthropic's error.type, such as rate_limit_error; "" without one requestId as stringthe request-id header or the body's request_id retryAfterMs as intthe wait the server asked for (retry-after-ms, retry-after seconds or HTTP date), at most 1e12; 0 when it named none limits as map of string to stringits rate-limit headers, lower-cased, at most 32: anthropic-ratelimit-*, x-ratelimit-*, retry-after, retry-after-ms
# MessagesFunction Returns Does newRequest(model)json.Valuea new request {"model", "max_tokens": 16000, "messages": []} withMaxTokens(req, n)json.Valueanother max_tokens, 0 to 300000 withSystem(req, text)json.Valuethe system prompt as text withSystemBlocks(req, blocks)json.Valuethe system prompt as text blocks (for cache_control) addUserText(req, text)json.Valuea user turn of text appended addUser(req, blocks)json.Valuea user turn of content blocks appended addAssistant(req, content)json.Valuean assistant turn appended (a string or blocks) addReply(req, reply)json.Valuea reply's whole content appended unchanged, thinking included textBlock(text)json.Valuea text block imageBlock(mediaType, data)json.Valuea base64 image block: jpeg, png, gif or webp pdfBlock(data)json.Valuea base64 PDF document block withCacheControl(v, ttl)json.Valuea cache breakpoint on a block or tool, or automatic caching on a request; ttl "", 5m or 1h withThinking(req, display)json.Valueadaptive thinking; display "", summarized, omitted or updates withEffort(req, effort)json.Valueoutput_config.effort: low to maxwithFormat(req, schema)json.Valuestructured output: output_config.format with a JSON Schema withFallbacks(req)json.Valueserver-side refusal fallbacks (needs BETA_FALLBACKS) chat(c, req)json.Valuesend the request, return the reply countTokens(c, req)intthe input tokens the request would take text(reply)stringthe reply's text blocks joined toolCalls(reply)list of ToolCallthe reply's tool_use blocks stopReason(reply)stringend_turn, max_tokens, tool_use, pause_turn, refusal, compaction, ...refusalCategory(reply)stringa refusal's category, such as cyber; "" otherwise usage(reply)Usagethe reply's token counts
# UsageField Meaning inputTokens as intevery input token: input_tokens + cache reads + cache writes outputTokens as intoutput tokens, thinking included totalTokens as intinputTokens + outputTokens (Anthropic sends no total)cachedTokens as intthe part of inputTokens read from the prompt cache reasoningTokens as intthe part of outputTokens spent thinking cacheCreationInputTokens as intthe part of inputTokens written to the prompt cache
Field Meaning id as stringthe tool_use block's id, which its tool_result answers name as stringthe tool's name input as json.Valuethe input, an object ({} when the block has none)
Function Returns Does tool(name, description, schema)json.Valuea tool definition withStrict(tool)json.Valuethe tool with strict: true addTool(req, tool)json.Valuea tool (or server tool) appended to tools addTools(req, tools)json.Valueseveral tools appended, in order toolResult(toolUseId, content, isError)json.Valuea tool_result block toolText(s)json.Valuetext as a JSON string, for a handler's answer toolError(message)json.Valuean error result for a handler's answer; the model can recover runTools(c, req, handler, maxTurns)ToolRunthe tool loop, until the model stops for something else runToolsWith(c, req, handler, context, maxTurns)ToolRunthe tool loop with a context passed to every handler call
A runTools handler is func(name as string, input as json.Value), a runToolsWith handler func(name as string, input as json.Value, context as json.Value); both return a json.Value. A handler that throws ends the loop with its error. A tool input over maxToolInput is answered with {"error": "the input is over N bytes; call the tool again with a smaller input"} (is_error) without calling the handler.
Field Meaning reply as json.Valuethe last reply request as json.Valuethe conversation, the last reply and its tool results included turns as inthow many requests were sent calls as inthow many tool calls the handler ran usage as Usagethe token counts of every reply, summed stopped as string"" when the model finished, maxTurns when the loop stopped there
# Message batchesFunction Returns Does batchRequest(customId, params)json.Valueone request of a batch createBatch(c, requests)json.Valuecreate a batch batch(c, id)json.Valuea batch and its processing_status batches(c, limit, afterId)json.Valueone page of batches, newest first cancelBatch(c, id)json.Valuecancel a batch in progress deleteBatch(c, id)json.Valuedelete an ended batch waitBatch(c, id, pollMs, timeoutMs)json.Valuepoll until the batch has ended batchResults(c, id)list of json.Valuethe JSONL results, one per request, in any order