Ardur

A personal AI agent with a security substrate built into the wire format — Rust, multi-channel, multi-provider. Every action an agent takes is scoped by a capability token, recorded in a signed receipt chain, and bounded by a cost gate, so the agent is something you can audit and govern rather than an opaque process.

What is Ardur

Ardur is an agent runtime. You reach it from your terminal or a chat app; it drives a turn through a single fused pipeline that authorizes the request, checks policy and budget, scans for prompt injection, dispatches to a model provider (running any tools the model asks for), then mints a signed receipt, records a bi-temporal memory, and appends to an append-only journal — in that order, short-circuiting on the first failure.

Three primitives thread through every edge of that pipeline:

• Cap-tokens — a scoped, revocable capability (a Biscuit, Ed25519-signed and offline-attenuable) authorizes each action. No ambient authority.
• Receipt-chain — every turn mints a JWS-ES256 receipt whose parent_hash links it to the prior one, so the audit trail is tamper-evident and replayable.
• Cost-tuples — a four-stage cost-admission gate projects, ceiling-checks, reserves, and finalizes the spend of every turn, refunding the difference.

Quickstart

# Build and install the `ardur` CLI from the workspace
cargo install --path crates/cli
#   …or, without installing:
cargo build --release && ./target/release/ardur chat

# Talk to it — Anthropic is the default provider
ARDUR_PROVIDER=anthropic ANTHROPIC_API_KEY=sk-ant-... ardur chat

# Local, no API key — point at an Ollama daemon
ARDUR_PROVIDER=ollama ardur chat

The same environment variables select the provider, memory backend, channels, tools, and skills for both the CLI and the server. The full operator runbook — every provider, the Qdrant/hybrid memory backends, the four chat channels, MCP, skills, OpenTelemetry, and the HTTP surface — lives in RUN.md. For the current implementation inventory, see docs/current-status.md.

Agent Bootstrap

Every LLM or agent session working in this repo must start with the read-only bootstrap:

python3 scripts/agent_bootstrap.py

It reports the current Linear project/issue status, native project progress, Git/worktree isolation state, EXTENDED-drive evidence root, parallel:ready work candidates, Plan Corpus verification backlog, Keychain-backed ARD Linear access status, provider posture, and the no-key smoke-test path. Each implementation must be merged back to dev with required workflows green before that session moves to another item. The conduct rules live in AGENTS.md and docs/agents/session-code-of-conduct.md.

Capabilities

Channels (4). Slack, Matrix, Discord, Telegram — each with bot-token auth, per-channel allowlists, and self-message echo prevention. Inbound messages on any channel run through the same fused turn pipeline; replies post back to the originating channel.

Providers (6). Anthropic (Claude), OpenRouter, generic OpenAI-compatible HTTP endpoints, Ollama (local daemon or hosted cloud), Codex (ChatGPT subscription via the codex CLI), and Claude-CLI (Anthropic subscription via the claude CLI). Selected at boot with ARDUR_PROVIDER. Anthropic, OpenRouter, OpenAI-compatible endpoints, and Ollama stream token-by-token (SSE, SSE, SSE, NDJSON respectively) through the uniform Provider::stream surface.

Tool execution. Tools run inside the fused pipeline — cap-token + Cedar policy + cost gate + injection-defense scan + signed receipt + journal + memory. The hardened built-in tools (shell.run, file.read, file.write, file.list, http.fetch) are implemented and tested with command allowlists, root-confinement, and SSRF defense, but default server boot currently registers only the safe local tools (echo, health_check), optional voice.transcribe, filesystem skills, and remote MCP tools. Default boot wiring for the hardened built-ins is tracked by ARD-457. External tools arrive over MCP via rmcp (Ardur is both an MCP client and an MCP server).

Memory. A bi-temporal store (event-time / valid-from-to / invalidation-time) backed in-process by default, or durably by Qdrant. A hybrid retriever fuses dense recall (fastembed BGE-small) and sparse recall (Tantivy BM25) with reciprocal-rank fusion.

Security substrate. Biscuit cap-tokens (Ed25519, attenuated, revocable) · Cedar policies (Allow / Deny / Indeterminate) · a four-stage cost-admission gate · a JWS-ES256 signed receipt chain (parent-hash linkage) · prompt-injection defense · append-only JSONL session journals.

CLI. The ardur REPL renders streaming Markdown (comrak), syntax-highlighted code (syntect), tool-call boxes, and a live cost line, with three themes (dawn / night / terminal) and slash commands. A full-screen Phase-2 TUI is designed (see docs/cli/ and ADR-Phase2-021) but not yet built.

HTTP surface. ardur-server exposes POST /chat (one full-pipeline turn, reply returned), POST /slack/events (HMAC-verified webhook), GET /healthz, and an optional bearer-gated MCP surface. Two companion binaries ship alongside: ardur-admin, a read-only observability dashboard over the on-disk journals, receipts, and memory; and ardur-eval, a Tau-Bench-style scenario evaluator that POSTs to /chat.

Skills. Filesystem SKILL.md skills (YAML frontmatter + Markdown body, with progressive disclosure via @./file.md references) register as tools from ARDUR_SKILLS_DIRS. Nine example skills ship under examples/skills/.

Architecture

A turn is a single FusedRuntime::submit call that runs ten stages in order, short-circuiting on the first failure: (1) cap-token verify → (2) Cedar policy → (3) cost-gate admit → (4) pre-submit hooks, with (4.5) injection-defense scan → (5) provider dispatch (looping any tool calls back through the pipeline) → (6) mint + sign the receipt → (7) post-receipt hooks → (8) cost-gate finalize/refund → (9) record bi-temporal memory → (10) append to the session journal. A failure in stages 1–5 never reaches the provider's billing; stages 6–10 run after the response and cannot un-happen the turn. FusedRuntime::stream drives the same ten stages but yields a progressive event feed, which is what the CLI consumes for its streaming UX.

Each subsystem is its own workspace package (47 in all). The design corpus and architecture-decision records live under docs/.

Status

Implementation effort wrapping up June 2026. The substrate is feature-complete and production-shaped, with CI green on dev and the no-key baseline passing locally. It is not yet a turnkey production deployment — see RUN.md's Known gaps and docs/current-status.md — so run it in a private channel before exposing it more widely.

Contributing

Contributions are welcome. See CONTRIBUTING.md — note that DCO sign-off and SSH-signed commits are required. Please also read the Code of Conduct and the security policy in SECURITY.md.

License

Licensed under the Apache License, Version 2.0 — chosen for its explicit patent grant and enterprise clarity.