Implements hybrid-custody NEAR key management where the agent holds scoped
function-call keys for routine operations while high-value operations require
explicit user approval through the existing channel approval flow.
Core infrastructure:
- Ed25519 key generation/import via ed25519-dalek (not near-crypto)
- AES-256-GCM encrypted storage via existing SecretsStore
- Hand-rolled borsh-serializable NEAR transaction types
- NEP-413 intent signing and MPC chain signature support
- Configurable policy engine with transaction analysis pipeline
- Daily spend tracking with automatic midnight UTC reset
- Encrypted backup/restore with Argon2id KDF
- CLI subcommands: generate, import, list, info, remove, export, policy, backup, restore
- NEAR ed25519 secret key leak detection (Critical/Block)
- WASM sign-payload host function (keys never enter WASM memory)
- KeyManager wired into AgentDeps for agent-wide access
Security invariants: private keys never reach the LLM or WASM boundary,
signing happens in host Rust code with Zeroize on drop, every transaction
is analyzed before signing, most-restrictive policy rule wins.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
- Fix MCP tool schema deserialization: rename input_schema to match
protocol's camelCase inputSchema, so models receive actual parameter
schemas instead of empty defaults
- Fix conversation history: add tool_calls field to ChatMessage and
include assistant message with tool_calls before tool results, as
required by OpenAI-compatible APIs
- Fix approval loop: pass resume_after_tool flag to run_agentic_loop
so the "force tool use" heuristic doesn't re-trigger after approval
- Fix shutdown: add Submission::Quit, Ctrl+C signal handler, and
graceful shutdown flow
- Fix MCP activate button: auto-attempt auth flow when activation
fails due to missing authentication
- Add inline approval cards in chat via SSE ApprovalNeeded events
- Add markdown rendering in chat (marked.js) with proper streaming
- Add structured fields to log entries (key=value pairs from tracing)
- Collapse log entries to single line with click-to-expand
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Auth mode: when a tool requires an API key, the thread enters a special
mode where the next user message is routed directly to the credential
store, bypassing logs, turns, history, and compaction entirely. This
prevents tokens from leaking into debug output or persistent storage.
Fix MCP auth: auth_mcp now actually uses the token parameter (was
ignored as _token) and falls back to manual token entry when OAuth
and DCR are both unsupported.
Parallel loading: WASM tools, WASM channels, and MCP servers now load
concurrently at startup. Within each loader, individual items also
load in parallel (join_all for WASM, JoinSet for MCP servers).
Co-Authored-By: Claude Opus 4.6 <[email protected]>
write_file now detects workspace files (HEARTBEAT.md, MEMORY.md,
SOUL.md, etc.) and daily/context/ prefixes, returning an error that
tells the LLM to use memory_write with the correct target instead.
This prevents the LLM from writing workspace data to the local
filesystem when it should go to the database.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Introduces a unified extension abstraction over MCP servers and WASM tools
with six agent-callable tools (tool_search, tool_install, tool_auth,
tool_activate, tool_list, tool_remove) so users can add capabilities
conversationally without CLI commands.
Includes built-in registry of 11 MCP servers, online discovery via URL
probing and GitHub search, OAuth 2.1 flows for MCP servers, and manual
token auth for WASM tools.
Co-Authored-By: Claude Opus 4.6 <[email protected]>
- Add dedicated "heartbeat" target in memory_write tool so the LLM
routes HEARTBEAT.md writes to the database instead of the filesystem
- Update tool description to clarify it's database-backed storage
- Broadcast heartbeat notifications to all channels when no explicit
notify target is configured, instead of silently logging them
Co-Authored-By: Claude Opus 4.6 <[email protected]>
Enables connecting to official MCP servers (like Notion) instead of
building custom WASM tools. Uses OAuth 2.1 with PKCE and supports
Dynamic Client Registration for zero-config authentication.
Key features:
- OAuth 2.1 flow with PKCE for secure browser-based auth
- Dynamic Client Registration (DCR) for servers without pre-configured clients
- Automatic token refresh on 401 responses
- Session management with Mcp-Session-Id headers
- SSE streaming response handling
New CLI commands:
- `mcp add <name> <url>` - Add an MCP server
- `mcp remove <name>` - Remove an MCP server
- `mcp list` - List configured servers
- `mcp auth <name>` - Authenticate with a server
- `mcp test <name>` - Test connection
Also removes the Notion WASM tool example since it's superseded by the
Notion MCP server which provides 13 official tools.
Co-Authored-By: Claude Opus 4.5 <[email protected]>
Renamed project from "near-agent" to "ironclaw" throughout the codebase.
Updated documentation to emphasize the core philosophy:
- Your data stays yours (local, encrypted, no telemetry)
- Self-expanding capabilities (build tools on the fly)
- Defense in depth (WASM sandbox, prompt injection defense)
- Always on user's side
Key changes:
- Package name: near-agent -> ironclaw
- Config paths: ~/.near-agent/ -> ~/.ironclaw/
- Database name in docs: near_agent -> ironclaw
- CLI binary: near-agent -> ironclaw
- Log filters: RUST_LOG=near_agent -> RUST_LOG=ironclaw
- All user-facing strings (welcome messages, help text, etc.)
Preserved for compatibility:
- HKDF salt "near-agent-secrets-v1" (changing would break existing secrets)
- WIT interface names (near::agent::*)
- NEAR AI provider config (NEARAI_* env vars)
Co-Authored-By: Claude Opus 4.5 <[email protected]>
The builder would get stuck when the LLM returned JSON specs or planning
text instead of tool calls. The loop would continue for all iterations
without making progress, eventually timing out.
Changes:
- Make initial prompt directive: "Use write_file NOW" instead of passive
"Start by creating the project structure"
- Add consecutive_text_responses counter to detect stuck state
- Fail fast after 2 consecutive text-only responses with clear error
- Send strong nudge on first text response: "STOP. Call write_file..."
- Reset counter once tools have been executed (completion phase)
Co-Authored-By: Claude Opus 4.5 <[email protected]>
- Add escape_telegram_markdown() to handle underscores in tool names
(e.g., build_software was breaking Telegram's Markdown parser)
- Use Telegram-compatible *bold* syntax instead of **bold**
- Clarify workspace memory vs filesystem tool descriptions to prevent
LLM from using read_file on memory_tree paths
- Update build_software to strongly prefer Rust WASM for agent tools
- Rewrite WASM tool template to use Component Model with wit_bindgen
instead of outdated extern "C" approach
Co-Authored-By: Claude Opus 4.5 <[email protected]>
Remove the brittle natural language pattern matching from the router
and add job management tools to the normal tool registry instead.
- Add job tools: create_job, list_jobs, job_status, cancel_job
- Router now only handles explicit /commands
- Natural language goes through agentic loop with all tools
- LLM naturally picks appropriate tools based on user intent
- Share ContextManager between job tools and Agent
Co-Authored-By: Claude Opus 4.5 <[email protected]>
Implements a general-purpose Docker sandbox (inspired by Codex) that provides:
- Container isolation for shell commands with ephemeral containers
- HTTP proxy for network access control with domain allowlist
- Credential injection by proxy (secrets never enter containers)
- Three security policies: ReadOnly, WorkspaceWrite, FullAccess
- Resource limits (memory, CPU, timeout enforcement)
Key components:
- SandboxManager: Main entry point coordinating proxy and containers
- NetworkProxy: HTTP proxy validating requests and injecting credentials
- ContainerRunner: Docker lifecycle management via bollard
- DomainAllowlist: Pattern matching for allowed network destinations
The ShellTool now routes commands through the sandbox when enabled,
with automatic fallback to direct execution if Docker is unavailable.
Configuration via SANDBOX_ENABLED, SANDBOX_POLICY, SANDBOX_TIMEOUT_SECS,
SANDBOX_MEMORY_LIMIT_MB, SANDBOX_IMAGE, SANDBOX_EXTRA_DOMAINS env vars.
Co-Authored-By: Claude Opus 4.5 <[email protected]>
Enable instant message delivery for Telegram via webhooks instead of polling.
Key changes:
- Add tunnel URL configuration for local development (ngrok, cloudflare)
- Auto-register webhook with Telegram API on startup using setWebhook
- Implement webhook secret validation via X-Telegram-Bot-Api-Secret-Token header
- Add credential injection for bot token via URL placeholder substitution
- Fix metadata preservation in respond() to route replies correctly
- Fix serde flatten with Option<T> issue in capabilities schema parsing
The credential injection pattern replaces {TELEGRAM_BOT_TOKEN} placeholders
in URLs with the actual token from the secrets store, keeping credentials
out of WASM module memory until the HTTP request is made.
Co-Authored-By: Claude Opus 4.5 <[email protected]>
- Store event_tx in Arc<Mutex<>> so respond() can send to TUI
- Fix run_event_loop to take owned receiver
- Switch main.rs from SimpleCliChannel to TuiChannel
- Add proper terminal cleanup on TUI error
Co-Authored-By: Claude Opus 4.5 <[email protected]>
Extends the WASM sandbox with HTTP API capabilities, secrets management,
tool aliasing, and leak detection. Key security principle: WASM never
sees credentials, injection happens at host boundary.
New modules:
- secrets: AES-256-GCM encrypted storage with HKDF key derivation
- leak_detector: Aho-Corasick + regex pattern matching for secret exfiltration
- capabilities: Extended capability system (HTTP, ToolInvoke, Secrets)
- allowlist: HTTP endpoint validation with glob patterns
- credential_injector: Host-boundary credential injection
- rate_limiter: Sliding window per-tool rate limiting
- storage: WASM binary storage with BLAKE3 integrity verification
Leak detection happens at two points:
1. Before HTTP request (prevents exfiltration via URL/headers/body)
2. After response (prevents exposure in outputs returned to WASM)
Co-Authored-By: Claude Opus 4.5 <[email protected]>
- Directories indicated by trailing `/` instead of is_directory field
- Files are plain strings, dirs with children are objects
- Remove redundant name/path (just show name in tree)
- Remove preview and updated_at (use memory_read if needed)
- Output is just the tree array, no wrapper object
Example: ["MEMORY.md", {"daily/": ["2024-01-15.md"]}, "projects/"]
Co-Authored-By: Claude Opus 4.5 <[email protected]>
The memory_tree tool provides a hierarchical view of the workspace
with configurable depth (default 1). This is more useful for
exploring nested directory structures.
Co-Authored-By: Claude Opus 4.5 <[email protected]>
- Consolidate all migrations into V1__initial.sql
- Replace DocType enum with flexible path-based file storage
- Add list_workspace_files SQL function for directory listing
- Update memory tools for path-based API (memory_read, memory_write,
memory_search, memory_list)
- Remove unused OpenAI/Anthropic providers (NEAR AI only)
- Simplify config to remove multi-provider support
- Update CLAUDE.md documentation
Co-Authored-By: Claude Opus 4.5 <[email protected]>
