openapi: 3.1.0
info:
  title: EvolutionaryScale Forge ESM3 API
  description: >
    Hosted inference API for the ESM3 multimodal protein language model from
    EvolutionaryScale. ESM3 reasons jointly across sequence, structure, and
    function tracks. The Forge API exposes generate, batch_generate, encode,
    decode, forward_and_sample, and logits operations across small (1.4B),
    medium (7B), and large (98B) parameter checkpoints. All requests authenticate
    with a Forge API token obtained from forge.evolutionaryscale.ai. This OpenAPI
    is reconstructed from the open-source `esm` Python SDK (Biohub/esm) which is
    the canonical client for the Forge service.
  version: 2024-08-01
  contact:
    name: EvolutionaryScale Forge
    url: https://forge.evolutionaryscale.ai
  license:
    name: Cambrian Inference Clickthrough License Agreement
    url: https://www.evolutionaryscale.ai/policies/cambrian-inference-clickthrough-license-agreement
servers:
  - url: https://forge.evolutionaryscale.ai
    description: EvolutionaryScale Forge Production
security:
  - BearerAuth: []
tags:
  - name: Generation
    description: Generate proteins across sequence, structure, and function tracks.
  - name: Encoding
    description: Tokenize ESMProtein objects into ESMProteinTensor inputs (and back).
  - name: Sampling
    description: Low-level forward passes with logits and sampling control.
paths:
  /api/v1/generate:
    post:
      summary: Generate Protein With ESM3
      description: >
        Iteratively samples tokens on the requested track of an `ESMProtein`
        input, filling in masked positions until the track is complete. Wraps
        encode, iterative sampling, and decode in a single call. Returns a
        fully populated `ESMProtein`.
      operationId: generate
      tags:
        - Generation
      requestBody:
        required: true
        content:
          application/json:
            schema:
              $ref: '#/components/schemas/GenerateRequest'
            examples:
              SequenceCompletion:
                $ref: '#/components/examples/SequenceCompletionRequest'
      responses:
        '200':
          description: Generated protein with completed track.
          content:
            application/json:
              schema:
                $ref: '#/components/schemas/ESMProtein'
        '4XX':
          $ref: '#/components/responses/ErrorResponse'
  /api/v1/batch_generate:
    post:
      summary: Batch Generate Proteins With ESM3
      description: >
        Generate multiple proteins concurrently using a list of `ESMProtein`
        inputs and a parallel list of `GenerationConfig` objects. Returns a
        list of completed `ESMProtein` results in the same order.
      operationId: batchGenerate
      tags:
        - Generation
      requestBody:
        required: true
        content:
          application/json:
            schema:
              $ref: '#/components/schemas/BatchGenerateRequest'
      responses:
        '200':
          description: List of generated proteins.
          content:
            application/json:
              schema:
                type: array
                items:
                  $ref: '#/components/schemas/ESMProtein'
        '4XX':
          $ref: '#/components/responses/ErrorResponse'
  /api/v1/encode:
    post:
      summary: Encode Protein To Tensor
      description: >
        Convert an `ESMProtein` into an `ESMProteinTensor` representation by
        running the structure tokenizer (and other per-track tokenizers) over
        the input. The tensor representation is required for low-level
        forward_and_sample and logits operations.
      operationId: encode
      tags:
        - Encoding
      requestBody:
        required: true
        content:
          application/json:
            schema:
              $ref: '#/components/schemas/EncodeRequest'
      responses:
        '200':
          description: Tokenized protein tensor.
          content:
            application/json:
              schema:
                $ref: '#/components/schemas/ESMProteinTensor'
        '4XX':
          $ref: '#/components/responses/ErrorResponse'
  /api/v1/decode:
    post:
      summary: Decode Protein Tensor
      description: >
        Inverse of encode. Runs the structure token decoder to convert an
        `ESMProteinTensor` back into an `ESMProtein` with sequence and atom
        coordinates.
      operationId: decode
      tags:
        - Encoding
      requestBody:
        required: true
        content:
          application/json:
            schema:
              $ref: '#/components/schemas/DecodeRequest'
      responses:
        '200':
          description: Decoded protein.
          content:
            application/json:
              schema:
                $ref: '#/components/schemas/ESMProtein'
        '4XX':
          $ref: '#/components/responses/ErrorResponse'
  /api/v1/forward_and_sample:
    post:
      summary: Forward And Sample
      description: >
        Executes a single forward pass of ESM3 against a tokenized protein and
        samples tokens according to a `SamplingConfig`. Intended for power
        users who need fine-grained control over each iteration (for example,
        chain-of-thought protein design or custom sampling schedules).
      operationId: forwardAndSample
      tags:
        - Sampling
      requestBody:
        required: true
        content:
          application/json:
            schema:
              $ref: '#/components/schemas/ForwardAndSampleRequest'
      responses:
        '200':
          description: Sampled output with optional logits and entropy.
          content:
            application/json:
              schema:
                $ref: '#/components/schemas/ForwardAndSampleOutput'
        '4XX':
          $ref: '#/components/responses/ErrorResponse'
  /api/v1/logits:
    post:
      summary: Get Logits And Embeddings
      description: >
        Returns raw per-position logits, embeddings, and optional hidden states
        for a tokenized protein. The SDK discourages routine use because logits
        payloads are large; prefer encode + forward_and_sample for typical
        workflows.
      operationId: logits
      tags:
        - Sampling
      requestBody:
        required: true
        content:
          application/json:
            schema:
              $ref: '#/components/schemas/LogitsRequest'
      responses:
        '200':
          description: Logits and embeddings response.
          content:
            application/json:
              schema:
                $ref: '#/components/schemas/LogitsOutput'
        '4XX':
          $ref: '#/components/responses/ErrorResponse'
components:
  securitySchemes:
    BearerAuth:
      type: http
      scheme: bearer
      description: Forge API token issued via forge.evolutionaryscale.ai.
  schemas:
    ESMProtein:
      type: object
      description: Raw protein representation with optional sequence, structure, function, secondary structure, and SASA tracks.
      properties:
        sequence:
          type: string
          description: Amino acid sequence using one-letter codes (A, R, N, ..., V). Underscore `_` marks masked positions.
          example: "MKTAYIAKQRQISFVK_____SSERVKKLLVGDIVT"
        coordinates:
          type: array
          description: Per-residue atom37 coordinates (list of [37,3] arrays).
          items:
            type: array
            items:
              type: array
              items:
                type: number
        secondary_structure:
          type: string
          description: DSSP-style per-residue secondary structure annotation.
        sasa:
          type: array
          description: Per-residue solvent accessible surface area.
          items:
            type: number
        function_annotations:
          type: array
          description: Optional per-residue function annotations.
          items:
            type: object
        ptm:
          type: number
          description: Predicted TM-score (if available).
        plddt:
          type: array
          description: Per-residue pLDDT confidence (if available).
          items:
            type: number
    ESMProteinTensor:
      type: object
      description: Tokenized representation of a protein across all ESM3 tracks.
      properties:
        sequence:
          type: array
          items:
            type: integer
        structure:
          type: array
          items:
            type: integer
        secondary_structure:
          type: array
          items:
            type: integer
        sasa:
          type: array
          items:
            type: integer
        function:
          type: array
          items:
            type: integer
        residue_annotations:
          type: array
          items:
            type: integer
    GenerationConfig:
      type: object
      description: Configuration for an ESM3 generate call.
      required:
        - track
      properties:
        track:
          type: string
          description: Which protein track to generate.
          enum:
            - sequence
            - structure
            - secondary_structure
            - sasa
            - function
        num_steps:
          type: integer
          description: Number of iterative sampling steps.
          default: 8
          minimum: 1
        temperature:
          type: number
          description: Sampling temperature.
          default: 1.0
          minimum: 0.0
        top_p:
          type: number
          description: Nucleus sampling cutoff.
          minimum: 0.0
          maximum: 1.0
        schedule:
          type: string
          description: Decoding schedule (e.g. cosine, linear).
          enum:
            - cosine
            - linear
        invalid_ids:
          type: array
          description: Token IDs that the sampler must never emit.
          items:
            type: integer
    SamplingConfig:
      type: object
      description: Per-track sampling configuration for forward_and_sample.
      properties:
        sequence:
          $ref: '#/components/schemas/SamplingTrackConfig'
        structure:
          $ref: '#/components/schemas/SamplingTrackConfig'
        secondary_structure:
          $ref: '#/components/schemas/SamplingTrackConfig'
        sasa:
          $ref: '#/components/schemas/SamplingTrackConfig'
        function:
          $ref: '#/components/schemas/SamplingTrackConfig'
    SamplingTrackConfig:
      type: object
      properties:
        temperature:
          type: number
        top_p:
          type: number
        invalid_ids:
          type: array
          items:
            type: integer
        only_sample_masked_tokens:
          type: boolean
          default: true
    GenerateRequest:
      type: object
      required:
        - model
        - protein
        - config
      properties:
        model:
          type: string
          description: ESM3 model checkpoint identifier.
          example: esm3-medium-2024-08
          enum:
            - esm3-large-2024-03
            - esm3-medium-2024-08
            - esm3-small-2024-08
            - esm3-open
        protein:
          $ref: '#/components/schemas/ESMProtein'
        config:
          $ref: '#/components/schemas/GenerationConfig'
    BatchGenerateRequest:
      type: object
      required:
        - model
        - proteins
        - configs
      properties:
        model:
          type: string
          example: esm3-medium-2024-08
        proteins:
          type: array
          items:
            $ref: '#/components/schemas/ESMProtein'
        configs:
          type: array
          items:
            $ref: '#/components/schemas/GenerationConfig'
    EncodeRequest:
      type: object
      required:
        - model
        - protein
      properties:
        model:
          type: string
          example: esm3-medium-2024-08
        protein:
          $ref: '#/components/schemas/ESMProtein'
    DecodeRequest:
      type: object
      required:
        - model
        - protein_tensor
      properties:
        model:
          type: string
          example: esm3-medium-2024-08
        protein_tensor:
          $ref: '#/components/schemas/ESMProteinTensor'
    ForwardAndSampleRequest:
      type: object
      required:
        - model
        - protein_tensor
        - sampling_config
      properties:
        model:
          type: string
          example: esm3-medium-2024-08
        protein_tensor:
          $ref: '#/components/schemas/ESMProteinTensor'
        sampling_config:
          $ref: '#/components/schemas/SamplingConfig'
    LogitsRequest:
      type: object
      required:
        - model
        - protein_tensor
      properties:
        model:
          type: string
          example: esm3-medium-2024-08
        protein_tensor:
          $ref: '#/components/schemas/ESMProteinTensor'
        config:
          type: object
          properties:
            sequence:
              type: boolean
              default: false
            structure:
              type: boolean
              default: false
            secondary_structure:
              type: boolean
              default: false
            sasa:
              type: boolean
              default: false
            function:
              type: boolean
              default: false
            return_embeddings:
              type: boolean
              default: false
            return_hidden_states:
              type: boolean
              default: false
            ith_hidden_layer:
              type: integer
              default: -1
    LogitsOutput:
      type: object
      description: Logits, embeddings, and optional hidden states.
      properties:
        logits:
          type: object
          properties:
            sequence:
              type: array
              items:
                type: array
                items:
                  type: number
            structure:
              type: array
              items:
                type: array
                items:
                  type: number
        embeddings:
          type: array
          description: Per-residue embedding vectors.
          items:
            type: array
            items:
              type: number
        hidden_states:
          type: array
          items:
            type: array
            items:
              type: array
              items:
                type: number
    ForwardAndSampleOutput:
      type: object
      properties:
        protein_tensor:
          $ref: '#/components/schemas/ESMProteinTensor'
        per_residue_entropy:
          type: array
          items:
            type: number
        topk_logprobs:
          type: object
        embeddings:
          type: array
          items:
            type: array
            items:
              type: number
    Error:
      type: object
      properties:
        error:
          type: object
          properties:
            type:
              type: string
            message:
              type: string
            code:
              type: string
  responses:
    ErrorResponse:
      description: Error returned by the Forge API.
      content:
        application/json:
          schema:
            $ref: '#/components/schemas/Error'
  examples:
    SequenceCompletionRequest:
      summary: Fill masked positions in a partial sequence
      value:
        model: esm3-medium-2024-08
        protein:
          sequence: "MKTAYIAKQRQISFVK_____SSERVKKLLVGDIVT"
        config:
          track: sequence
          num_steps: 8
          temperature: 1.0