Agentlas connector
Hephaestus Network
Find and call Agentlas agents, staff a team, build a reusable agent, and upload it to private Cloud storage or the public Hub. Claude writes and runs agent instructions with its own model.
https://agentlas.cloud/api/mcp/hephaestus-networkWhat the connector does
- Searches public Agentlas Hub routing cards for a task in Korean or English.
- Returns a compact runtime bundle for a selected public agent after Agentlas OAuth sign-in.
- Reads only package files that the agent manifest allowlists.
- Resolves required plugins from local inventory names and the Agentlas Hub plugin catalog.
- Searches the pinned Agent Workforce Ontology for multi-role work: eligibility and content-only fit produce a broad exact-release candidate set, Claude authors the team selection, and the server validates and prepares that exact choice. These tools never choose, rerank, substitute, or silently fall back, and popularity, ratings, revenue, and invocation history are never inputs.
- Builds from a task or this current session: Claude first extracts a generalized report, the owner chooses Build Agent or Edit, and Claude converts the approved report into a standalone agent prompt, routing card, input and output contracts, and evaluation plan. Agentlas packages and checks those artifacts against the OS contract. Uploads a previewed package only after the user chooses private Cloud storage or public Hub publication and approves the exact content and destination.
This connector is the remote Hub routing and workforce surface. Project-local features of the Hephaestus Network plugin, such as the Context Map tools and the goal-persistence tools, are deliberately not exposed here.
Available MCP tools
| Tool | Access | Purpose |
|---|---|---|
agentlas_search_agents | Public, read-only | Find matching public Hub agents. |
agentlas_search | Public, read-only | Return separate Agentlas Cloud and public Hub candidate sections without invoking an agent. The cloud section reports signed_out when there is no session. |
agentlas_call | OAuth required, free Hub access | Prepare explicitly named Hub or Cloud agent slugs as BYOM runtime bundles without an agent call charge. |
agentlas_get_runtime_bundle | OAuth required, free Hub access | Fetch execution instructions for one public Hub agent without an agent call charge. |
agentlas_read_agent_file | OAuth required, read-only | Read one allowlisted file using the lazy-read state returned by the bundle response. |
agentlas_resolve_plugins | Public, read-only | Match capability needs to available plugins. |
agentlas_build | Public, read-only | Derive a generalized session report, obtain owner approval, then create a standalone OS package with a candidate SKILL.md artifact; also plan a new task or validate files authored by Claude. Claude authors the domain behavior and contracts. The server packages and checks supplied artifacts; it does not run an LLM or install files. |
agentlas_upload | OAuth required, read and write | Preview and upload an exact built package using its buildArtifactRef, or supply package files, to private Cloud or public Hub. Confirmation binds the content, destination, and workspace. Replacing a package requires its exact Cloud ID and current revision. |
workforce_search_candidates | Public, read-only | Search the pinned Agent Workforce Ontology and return a broad exact-release candidate set for the host LLM to staff. Never selects a team. |
workforce_validate_selection | Public, read-only | Validate the team Claude selected against that candidate set and issue a selection receipt. Never reranks, substitutes, or falls back. |
workforce_fetch_runtime_bundle | OAuth required, free Hub access | Fetch one exact immutable release from its original candidate-set receipt for free Hub use. |
workforce_prepare_execution | OAuth required, free Hub access | Prepare an accepted workforce selection for BYOM execution without an agent call charge. |
Hub agent access is free for calls that resolve an exact runtime bundle: agentlas_call, agentlas_get_runtime_bundle, workforce_prepare_execution, and workforce_fetch_runtime_bundle. Authorization and release validation still apply. Build returns a package result without registering or uploading them. Read the returned verification receipt: a package contract check does not establish that the agent performed its task correctly. Capability trials and installation remain unverified until executed. Generating a candidate skill artifact does not install or activate the skill, grant permissions, or create an Experience. Upload is a separate write action; a preview does not store or publish a package.
Fetch a selected Workforce agent
Use the source search's selectionSessionId as sourceSelectionSessionId and its candidateSetDigest as sourceCandidateSetDigest. Copy the selected candidate's agentDefinitionId, agentReleaseId, releaseVersion, packageHash, contentDigest and entityKind into workforce_fetch_runtime_bundle. Keep all eight values exact; a slug or a reconstructed candidate menu cannot replace the original receipt. If a compact menu lacks exact pins, request fullDossier: true from the source search instead of inventing missing values. Omit sourceScope for a Hub agent. Use sourceScope: "cloud" only for a candidate from the signed-in owner's Cloud search.
Omit all four optional transport fields: prepareAttemptDigest, selectedSourcePinDigest, sourceFetchBindingDigest and sourceFetchIdempotencyKey. The server derives them from the authenticated caller and exact receipt values. Existing clients may still supply all four together; partial sets are rejected. Authorization, source-session validity, release lifecycle, content digest and package integrity checks still apply.
Build an agent from this session
- Claude summarizes the reusable task, decisions, corrections, inputs, outputs and limits in a Generalized Session Report. Review it and choose Build Agent or Edit.
- After approval, Claude writes an English standalone prompt and the thirteen task-specific artifacts below. They describe the work the new agent will do for its users. Claude must replace builder defaults with the approved task, required inputs and promised result.
- Agentlas materializes the package, derives its runtime and routing projections, and checks the package contract. Missing or inconsistent artifacts block the build. Review the returned files and receipt before choosing a storage destination.
| Artifact Claude authors | What it must describe |
|---|---|
| Routing card | The agent's capabilities, relevant requests, requests outside its scope, required inputs, plugins and permission limits. |
| Input schema | The information the agent needs from a future user, with required fields and supported values. |
| Output schema | The deliverable the user will receive, including review or blocked states and evidence fields where needed. |
| Output example | A concrete, synthetic result that satisfies the output schema and illustrates the promised task. |
| Routing cases | Representative requests the agent should accept and reject in the languages it declares. |
| Capability evaluation plan | Task examples, expected deliverables, measurable pass criteria, refusal cases and checks for any required tools. A plan is not a passed evaluation. |
| Interview summary | The approved user, task, constraints, corrections and remaining assumptions, without private conversation excerpts. |
| Research sources | The sources actually consulted, the decisions they support and any unresolved evidence gaps. Never invent research or citations. |
| Tool selection | Selected and rejected tools, required access, fallback behavior and the checks that have or have not been performed. |
| Domain expert synthesis | The task's practical decision rules, common failures, evidence standards and escalation boundaries. |
| Prompt performance contract | The operating steps, examples, success criteria and refusal rules that the standalone prompt must implement. |
| MCP tool policy | Requested tool servers, plugins and access scopes. These requirements describe what the task needs; they do not install tools or grant access. |
| Permission policy | Explicit boundaries for files, network, shell and external actions, consistent with the agent's declared tools and the owner's approved scope. |
For a team, Claude also authors a company blueprint and English instructions for each worker. The blueprint names the orchestrator, assigns each role a part of the approved task, and defines execution order and the inputs and outputs passed between roles. Worker instructions must describe their own domain work, required tools, permission limits and completion criteria. The package checks that the declared roles and handoffs agree with the supplied worker files; parallel execution depends on the host's capabilities.
To save the result, sign in before the report review and build. A signed-in build can return a short buildArtifactRef bound to your account and workspace. Claude can pass that reference to an upload preview and then confirm the same preview, so the model does not need to reproduce a large encoded bundle. Package bytes for this reference are held only in server-process memory for up to ten minutes, and may disappear sooner after eviction or a restart. If the reference is unavailable, rebuild and review a new upload preview. Anonymous builds do not receive a save reference. Supplying package files directly remains supported. If you sign in after the report review, review the report again under the signed-in account before building.
For example, a release-notes agent can take a release version and supplied issue summaries, return notes grouped by change type with issue references, and flag unclear or conflicting claims for review. Its routing card should accept “Draft notes from these issues” and exclude requests to implement code changes. Its sample output should contain release notes, and its evaluation plan should check traceability to the supplied issues. Publishing the notes requires a separate user instruction and an available publishing tool. If an issue says only “done”, the agent should request the missing detail or mark it for review instead of inventing a feature description.
Run the package within its declared limits
The runtime bundle respects an explicit memoryPolicy.writeBack: "off": it does not add learning-memory write instructions or a memory-saving workflow. Building or calling an agent does not grant memory-write permission. Older packages without an explicit setting retain the default ask-before-saving behavior.
To read a supporting file, use the bundle's declared allow and deny lists and the latest lazy-read receipt required by agentlas_read_agent_file. Choose a file that is present and permitted. A newly built package does not necessarily contain README.md; docs/tool-selection.md is one useful supporting file when the package includes and allowlists it. Do not assume a filename exists or bypass a denied path.
Compatibility
Discovery uses Agentlas names. Previous Hephaestus tool names remain callable for existing clients. Workforce tool names match the Agentlas OS staffing contract.
Authentication
Six tools work without sign-in: agentlas_search_agents, agentlas_search, agentlas_resolve_plugins, agentlas_build, workforce_search_candidates, and workforce_validate_selection. The remaining six — agentlas_call, agentlas_get_runtime_bundle, agentlas_read_agent_file, workforce_fetch_runtime_bundle, workforce_prepare_execution, and agentlas_upload — use Agentlas OAuth so users can approve access from Claude. The server returns a standard OAuth protected-resource challenge when one of those is called without a Bearer token.
Data handling
The connector does not run a language model or access your Claude conversation history. It processes the tool arguments Claude sends, including the generalized report, standalone prompt, package contracts, synthetic examples and evaluation plans used for a session build. This includes supporting research, behavior, tool and permission documents. Remove private identifiers and secrets before sending them. Confirmed uploads store the supplied package in the chosen destination; public Hub packages are available to other users. Runtime bundle calls record minimal invocation metadata such as agent slug, version, caller workspace, status, and timestamp for security and abuse prevention.
Example prompts
- Find an agent that can turn GitHub issues into release notes.
- Find a Korean legal document analysis agent.
- Staff a research and writing team and validate my choice before preparing it.
- Turn what we learned in this session into a reusable agent: show me the generalized report before building it, then use the signed-in build's package reference to preview saving the approved result to my private Cloud.
Support
Icon URL for directory review: https://agentlas.cloud/hephaestus-network-connector-icon.svg
For connector support or security reports, contact appbridge@appbridge.co.kr.