Docs · Start Overview Choose the right entry point and understand the pre-alpha boundary. Docs · Start Getting started Install the pinned toolchain and validate the repository locally. Docs · Start Core contract examples Compile the first Agent, typed Tool, Model stream, and explicit protocol mapping without runtime dependencies. Docs · Start Typed Agent Generate and pin typed schemas, then bind the first-party OpenAI and Anthropic adapters. Docs · Start In-process Agent Run a typed Agent without HTTP while preserving PostgreSQL durability, authorization, recovery, and owned roles. Docs · Start Durable admission Atomically commit authenticated intent, database time, the first event, and the initial checkpoint. Docs · Start Durable runs and results Submit with a durable ingress key, recover the original run, and read a verified public result. Docs · Understand Compile a graph Build a schema-pinned graph, validate its topology, and freeze its canonical digest. Docs · Understand Subgraphs and loops Compose shared-state subgraphs, finite loops, and explicit recovery-safe return paths. Docs · Understand Skill composition Choose capability bundles, shared-state subgraphs, durable child runs, and guarded MCP Skills without creating a second runtime. Docs · Understand Durability model Learn how journal, checkpoint, lease, fence, and recovery evidence fit together. Docs · Understand Durable Graph runtime Register exact executable code, replay noninitial state, and drive deterministic bounded sibling batches under one fence. Docs · Understand Durable Agent Loop Commit lifecycle handoffs and execute one tenant-scoped durable scheduling quantum. Docs · Understand Durable invocations Execute exact model and tool attempts with durable starts, staged ordered Tool coordination, streaming validation, and no-dispatch terminal recovery. Docs · Understand Provider-native Agent Compile and operate the durable multi-turn graph with bounded output repair, ordered parallel read-only Tools, write barriers, policy, accounting, and cancellation evidence. Docs · Integrate Local Rust Tools Generate digest-pinned schemas, register typed local code, and preserve the durable invocation boundary. Docs · Integrate AgentService v1 Authorize and expose exact durable Agent revisions through the versioned embedding boundary. Docs · Integrate Agent HTTP v1 Submit, recover, read and cancel through authenticated bounded JSON HTTP. Docs · Integrate General MCP Tool client Discover and call bounded stateless MCP 2026-07-28 Tools with JSON/SSE, OAuth, custom headers, and mediated MRTR. Docs · Integrate MCP OAuth client Bind one MCP resource to challenge-driven discovery, PKCE, issuer validation, durable credentials, and bounded replay. Docs · Integrate MCP Remote Tool Bind one strict stateless MCP 2026-07-28 Tool with pinned identity, schemas, durable dispatch, and reconciliation. Docs · Integrate MCP Server profile Expose bounded Tools, Resources, Prompts, Completion, and MRTR behind the strict stateless production boundary. Docs · Integrate MCP Skills Publish and activate Final SEP-2640 static Skills with immutable manifests, origin-scoped approval, and verified lazy reads. Docs · Integrate MCP conformance status See the frozen official runner inventory, implemented evidence, and exact boundary of current MCP claims. Docs · Integrate A2A 1.0 Client Discover and pin one agent, call all eleven operations, and reconcile durable unknown sends only through attested remote guarantees. Docs · Integrate Durable artifact storage Materialize terminal A2A parts into immutable PostgreSQL metadata and private integrity-checked object bytes. Docs · Integrate A2A 1.0 Server Expose bounded HTTP+JSON and JSON-RPC/SSE through identity-first policy and a durable service contract. Docs · Integrate A2A conformance status Inspect the frozen official TCK, audited harness patch, exact result, and server-only claim boundary. Docs · Operate Fair scheduling Configure replica-safe weighted tenant selection, exact starvation bounds, rollout, and retention. Docs · Operate PostgreSQL provider Configure, migrate, verify, and test the implemented durability provider. Docs · Operate Host qualification Record bounded capacity, recovery and fault evidence without turning CI into a production SLO claim. Docs · Project Versioning and releases Pin the public alpha exactly and understand compatibility, MSRV, durable-data, and release gates. Docs · Project Implementation status See what is implemented, in progress, and deliberately not claimed. Home Architecture map What exists, what is active, and what remains planned. Home Current cut Implemented slices and explicit gaps. Home Roadmap The path from contracts to a qualified v1. Repository Source code StateKnot/StateKnot on GitHub. Repository RFCs Draft architecture and durable execution contracts. Repository Security policy Private vulnerability reporting and support policy.
↑↓ navigate ↵ open 39 results

Pre-alpha documentation

Publish and activate manifest-bound Agent Skills over MCP.

Use the Final SEP-2640 static Server and Client/Host profiles with immutable manifests, origin-scoped approval, lazy verified reads, and explicit execution gates.

One immutable snapshot drives every response

The server freezes each SKILL.md and supporting file at startup. skills/list, skills/get, and resources/read use those same bytes, so a manifest cannot drift away from the content returned later.

BoundaryGuarantee
DiscoveryBounded pagination, deterministic order, TTL and cache scope.
ManifestEvery file appears once with exact raw-byte SHA-256 and size.
ReadsText or Base64 content from the retained immutable snapshot.
PolicyAuthentication, exact scopes and dynamic authorization before disclosure.
Limits512 files and 16 MiB per Skill, plus aggregate catalog ceilings.
HostOrigin + URI identity, approval before reads, exact digest/size/ frontmatter verification, durable scoped acting windows, exact restart resume, immutable revocation, and manifest-only lazy access.
ExecutionExact owner/version/descriptor binding plus a fresh, acting-window-bound policy permit before every execution and recovery probe, with exact durable correlation and bounded schema-bound input. Each fresh payload-redacted receipt serializes with revocation and validates its active window before provider I/O; later revocation does not recall an authorization already committed.

Activate through the verification-first Host

Rust
let client = McpClient::connect(
    endpoint,
    McpClientIdentity::new("orders-agent", env!("CARGO_PKG_VERSION"))?,
    authorization,
    McpClientOptions::for_skills(),
).await?;

let host = McpSkillHost::new(
    client,
    McpSkillOrigin::new("production/orders-mcp")?,
    SkillActivationScope::new(tenant_id, run_id, thread_id),
    Arc::new(application_skill_policy),
    Arc::new(postgres_store.clone()),
    McpSkillHostOptions::default(),
)?;

let catalog = host.list_skills().await?;
let selected = catalog.find_uri("skill://code-review/SKILL.md")
    .ok_or(AppError::SkillUnavailable)?;
let activation = McpSkillActivationAttempt::generate();
let active = Arc::new(host.activate(selected, activation).await?);

let file = active.read_file("references/checklist.md").await?;
model_context.push_untrusted_skill_file(file.identity(), file.bytes())?;

let guarded = Arc::new(McpSkillBoundTool::new(
    Arc::clone(&active),
    ticket_create_tool,
    Arc::new(postgres_store.clone()),
    McpSkillHostCodeExecution::NotPossible,
)?);
tool_registry.register(guarded)?;
let tools = tool_registry.build();

Listing never prefetches Skill files. Preserve each verified file's origin identity in model-visible context. Register McpSkillBoundTool instead of the raw provider so the ordinary durable Tool runtime enforces an exact policy decision before execution and reconciliation. Frontmatter allowed-tools is review input, never an authorization grant. Policy returns version-pinned activation and per-operation grant evidence. The durable window store and receipt sink fail closed before dispatch when evidence is expired, revoked, crossed, or unavailable.

Reject ambiguous or unsafe packages at startup

The registry rejects duplicate YAML keys at any depth, malformed Agent Skills fields, alias expansion beyond the materialized-value budget, name/URI mismatch, traversal and non-canonical paths, overlapping resource URIs, incomplete packages, and oversized catalogs. Unknown frontmatter fields remain data and are preserved without becoming authority.

Build the static server surface

Rust
let skill = McpServerSkillDefinition::new(
    "skill://code-review/SKILL.md",
    [
        McpServerSkillFile::text(
            "SKILL.md",
            "text/markdown",
            include_str!("skills/code-review/SKILL.md"),
        )?,
        McpServerSkillFile::text(
            "references/checklist.md",
            "text/markdown",
            include_str!("skills/code-review/references/checklist.md"),
        )?,
    ],
)?;

let mut skills = McpServerSkillCatalogBuilder::default();
skills.register(skill)?;

let app = McpServerApplicationBuilder::new(options)
    .with_skills(skills.build()?, skill_authorization)?
    .build()?;

Keep the bytes in reviewed deployment inputs. Do not rebuild Skill registries on the request path.

Reproduce both profile gates

Console
cargo test -p stateknot-integrations mcp_server_skill --locked
cargo test -p stateknot-integrations --test mcp_skills_server --locked
cargo test -p stateknot-integrations mcp_skill_host --locked
cargo test -p stateknot-integrations --test mcp_skills_host --locked
cargo test -p stateknot-store-postgres --test postgres --locked   tool_authorization_receipts_are_exact_immutable_and_page_verifiable

Read the complete implementation and non-claim boundaries in the Server guide and Client/Host guide.