Anthropic

Anthropic models support Text, Vision and PDF content. Caching is enabled by default, and Remote MCP is supported.

Streaming timeout

streaming_timeout limits HTTP read inactivity for Anthropic Messages, including Copilot Messages. SSE pings keep a stream alive even when no text or other parsed events arrive. It is not a total response deadline. Parsed-event timing is still recorded for diagnostics.

streaming_timeout=none disables the read and stream-start timeouts; SDK connect/write/pool limits remain unchanged. A numeric value also bounds stream startup as before.

Claude Opus 5.5

Released September 22, 2026. opus and opus55 select claude-opus-5-5; opus5 remains pinned to Opus 5. Copilot uses copilot.claude-opus-5.5 (or copilot.opus / copilot.opus55).

• 1,000,000-token context and 128,000-token maximum output.
• Adaptive thinking is always on and cannot be disabled or assigned a manual token budget. Default effort is medium; auto leaves effort to the provider. Supported effort levels are low, medium, high, xhigh, and max, as confirmed in the Anthropic effort documentation. Fast-mode pricing is not assumed.
• Forced tool use is unsupported. Use auto or none, and native JSON mode for structured output on direct Anthropic. Copilot retains its existing structured-output restrictions (forced-tool fallback is rejected). Unsupported sampling controls are removed, as for Fable 5.1.
• Thinking blocks belong to their originating model and conversation. Do not transplant them into other conversations or models, or edit earlier history. Preserve empty signed thinking blocks and text between tool calls. Thinking display is empty by default; fast-agent does not request summarized display.
• Anthropic pricing per million tokens: input $4, output $20, 5-minute cache writes $5, 1-hour cache writes $8, cache reads $0.20. Cache reads are 5% of input price, not 10%. These are not Copilot charges.

Claude Fable 5.1

Use claude-fable-5-1 (for example, fast-agent go --model "claude-fable-5-1?reasoning=max"). Existing fable and fable5 aliases still select Fable 5.

• 1M-token context; up to 128K output tokens, including thinking.
• Adaptive thinking is always on. auto leaves effort at the API default (high); supported levels are low, medium, high, xhigh, and max. Thinking cannot be disabled or configured with a manual token budget.
• Sampling controls (temperature, top_p, top_k) are removed before sending.
• Tool choice supports auto and none, not forced/required tools. Use native JSON structured output rather than tool_use mode; unsupported forcing fails locally.
• Preserve thinking blocks unchanged, including empty signed blocks, and keep history append-only. Editing earlier turns invalidates thinking; earlier models cannot read Fable 5.1 thinking blocks. Assistant prefill is unsupported.
• Requires 30-day data retention unless Anthropic expressly authorizes otherwise. Input/output pricing is $10/$50 per million tokens; cache reads are $0.25/MTok.

Per-message effort, turn-scoped system messages, and progress-update betas are not enabled automatically by this integration.

See the official overview, migration guide, and thinking constraints.

YAML Configuration:

anthropic:
  api_key: "your_anthropic_key" # Optional if ANTHROPIC_API_KEY or Anthropic SDK credentials are available
  base_url: "https://api.anthropic.com/v1" # Default, only include if required
  cache_mode: "auto" # Options: off, prompt, auto (default: auto)
  cache_ttl: "5m" # Options: 5m, 1h (default: 5m)
  cache_diagnostics: false # First-party Anthropic cache-miss diagnosis (debug only)
  web_search:
    enabled: false
    # max_uses: 3
    # allowed_domains: ["example.com", "*.docs.example.com"]
    # blocked_domains: ["social.example"]  # mutually exclusive with allowed_domains
    # user_location:
    #   type: approximate
    #   city: "London"
    #   country: "UK"
  web_fetch:
    enabled: false
    citations_enabled: false
    # max_uses: 3
    # max_content_tokens: 4096
    # allowed_domains: ["example.com"]
    # blocked_domains: ["tracking.example"]  # mutually exclusive with allowed_domains

Environment Variables:

• ANTHROPIC_API_KEY: Your Anthropic API key
• ANTHROPIC_AUTH_TOKEN: Bearer token auth supported by the Anthropic SDK
• ANTHROPIC_PROFILE / ANTHROPIC_CONFIG_DIR: Select Anthropic SDK profile credentials
• ANTHROPIC_BASE_URL: Override the API endpoint

Authentication precedence:

fast-agent first uses an explicit anthropic.api_key, then ANTHROPIC_API_KEY. If neither is set, it constructs the Anthropic SDK client without an API key so the SDK can use its own credential chain, including ANTHROPIC_AUTH_TOKEN, profile credentials, and workload identity federation environment variables. fast-agent check reports these as Anthropic SDK credentials when the SDK finds an auth source.

Caching Options:

The cache_mode setting controls how prompt caching is applied:

• off: No caching, even if global prompt_caching is enabled
• prompt: Caches tools, system prompt, and template content
• auto: Also advances cache checkpoints through recent conversation turns (default)

The cache_ttl setting controls how long cached content persists:

• 5m: Standard 5-minute cache (default)
• 1h: Extended 1-hour cache (additional cost)

The TTL is a user policy choice. A 5-minute cache has a lower write premium and usually pays off after one future read. A 1-hour cache survives longer gaps but has a higher write premium and generally needs two future reads to pay off. For long-running managed processes, pair the extended cache with a model-local wait such as opus?poll_period=3000; the wait returns early if the process finishes and leaves margin before the one-hour cache expires. Anthropic silently skips cache writes when the marked prefix is below the model's minimum cacheable size; check the cache creation/read usage fields before treating an early request as a cache miss.

Set cache_diagnostics: true only while debugging against the first-party Anthropic API. It enables Anthropic's cache-diagnosis beta and records the provider's cache-miss reason in the fast-agent-provider-diagnostics response channel. It is disabled by default and is not sent by Anthropic-on-Vertex.

Reasoning + Structured Outputs:

Claude reasoning support depends on the model family:

Adaptive models that require an explicit setting use thinking: {"type": "adaptive"} under the hood. Opus 5 has adaptive thinking on by default, so fast-agent omits the field for auto. Use effort levels (low, medium, high, xhigh where supported, max) or auto with anthropic.reasoning:

anthropic:
  reasoning: "high"

Adaptive models default to auto (provider-chosen). Do not configure fixed thinking budgets for these models; use effort levels instead.

task_budget is available for Claude Opus 4.7+ in fast-agent. It gives the model a visible token budget for a full agentic loop, so the model can self-moderate. It is different from max_tokens, which is still the enforced ceiling for one response:

anthropic:
  reasoning: "xhigh"
  task_budget: 128k

Anthropic models using budget-based thinking default to reasoning on with a 1024 token budget. Use anthropic.reasoning to set a budget, map from effort aliases, or disable reasoning entirely:

anthropic:
  reasoning: 16000 # Reasoning budget tokens (minimum: 1024)

• Disable reasoning with reasoning: "0", reasoning: "off", or reasoning: false.
• Budget models also accept low/medium/high/max to map to preset budgets.
• The reasoning budget must be less than max_tokens. If you set a budget that meets/exceeds max_tokens, fast-agent raises max_tokens so the budget fits.

You can also set reasoning per run using the model string:

• sonnet?reasoning=4096
• opus5?reasoning=xhigh&task_budget=128k
• opus47?reasoning=auto&task_budget=64k
• claude-opus-4-6?reasoning=auto

Structured output selection (Anthropic JSON schema vs tool_use):

• Models that support the structured-outputs-2025-11-13 feature default to JSON schema output (structured_output_mode: json). This mode is compatible with reasoning.
• Older models default to the legacy tool_use structured output flow. tool_use is not compatible with reasoning — fast-agent disables reasoning when tool-forced structured output is selected.
• Anthropic on Vertex does not support modern structured outputs in fast-agent; choose structured_output_mode: tool_use / ?structured=tool_use there.

You can override the structured output mode explicitly:

anthropic:
  structured_output_mode: auto # auto (default), json, or tool_use

Deprecated: thinking_enabled and thinking_budget_tokens are ignored. Use reasoning.

Built-in Anthropic web tools (web_search + web_fetch):

fast-agent can enable Anthropic server-side web tools directly (these are not MCP tool calls):

• anthropic.web_search.enabled: true
• anthropic.web_fetch.enabled: true

Optional controls:

• max_uses
• allowed_domains / blocked_domains (mutually exclusive)
• web_search.user_location (approximate city/region/country/timezone)
• web_fetch.max_content_tokens
• web_fetch.citations_enabled

You can override per run in the model string:

• claude-opus-4-6?web_search=on&web_fetch=on
• sonnet?web_search=off

Supported values are on/off (also accepts true/false, 1/0).

Version policy is model-aware:

• Claude Opus 5 supports web_search_20260209 but does not support web_fetch.
• Claude 4.6 models use web_search_20260209 and web_fetch_20260209 (with required beta header code-execution-web-tools-2026-02-09).
• Other supported Anthropic models use legacy versions (web_search_20250305, web_fetch_20250910).

Provider-managed remote MCP:

The direct anthropic provider supports provider-managed remote MCP servers declared with management: provider under mcp.servers or card mcp_connect entries.

• Supported on anthropic
• Not supported on anthropic-vertex
• Server must be a remote http/sse URL
• Use access_token for bearer auth if required

See Configuration Reference for the MCP server schema and Agent Cards for card-scoped runtime targets.

Model Name Aliases: