agentd v0.1.0-alpha.1

v0.1.0-alpha.1

This is the first public alpha of agentd: a durable, single-host runtime for
personal agents with explicit capabilities, scoped serialization, inspectable
traces, and transactional output delivery.

Highlights

• Tenant-scoped agents, context, memory, artifacts, schedules, MCP catalogs,
  runs, and deliveries in one libSQL database.
• Serialization by (tenant, agent, scope) while unrelated scopes remain
  concurrent.
• Atomic finalization of canonical output, terminal trace, rolling context, and
  an optional delivery reference.
• A native OpenAI-compatible model/tool loop with 15 built-in capabilities and
  tenant-scoped stdio/HTTP MCP discovery.
• A pull delivery outbox with expiring claims, token-checked acknowledgement,
  retry, and no copied output payload.
• Hybrid lexical/semantic memory using a revision- and checksum-pinned
  multilingual E5 Small ONNX model.
• A credential-free deterministic demo that exercises the complete turn,
  result, trace, delivery-check, and artifact lifecycle.
• CI, dependency advisory/license/source checks, current-tree secret scanning,
  a threat model, and an executable reliability evidence matrix.

Try the alpha candidate

Requirements are Rust, Python 3, curl, and jq:

git clone https://github.com/minifish-org/agentd.git
cd agentd
./scripts/demo-e2e.sh

The first run downloads about 448 MiB of checksum-verified embedding assets.
The demo provider is a deterministic protocol fixture; it proves the agentd
lifecycle, not language-model quality or real tool-call interoperability.

Compatibility and support boundary

• This is experimental pre-1.0 software.
• Runtime data is disposable; schema changes require --reset-data.
• HTTP compatibility is not preserved unless release notes explicitly say so.
• The instance bearer token is an operator credential, not end-user auth.
• Tenants are namespaces, not isolation for mutually distrustful users.
• Operator-configured stdio MCP processes are not sandboxed.
• TLS, rate limiting, quotas, backups, and host/network isolation remain
  deployment responsibilities.
• There is no HA, migration framework, deterministic side-effect replay,
  approval workflow, or multi-agent orchestration layer.

Release verification

The tagged source is exercised by CI and security workflows plus a dedicated
release check covering the real embedding path and a complete Docker image
build and inspection. The credential-free demo is also verified from a clean
clone.

See the reliability matrix,
threat model,
and security policy
for the precise evidence and boundaries.