Kallip AI

Early stage. Not ready for production use.

An agent harness designed from the start as a multi-agent system. Agents manage their own subagents and context, and coordinate across projects.

Not another coding assistant

Existing coding assistants excel at single-project, single-session work. kallip aims at a different problem: cross-project, multi-agent coordination — without being a "multi-agent manager."

Instead of a top-down orchestration layer, kallip is driven through a headless CLI: the agent itself decides when to spawn, switch between, and coordinate multiple agent instances across projects.

For planned direction, see the roadmap.

Architecture

Crate	Description
kallip-common	Shared types and command parsing
kallip-runtime	Agent runtime: agent context management, policy engine, tool dispatch
kallip-shell	Provider-neutral shell/session tools for LLM applications (used by the runtime)
kallip-daemon	HTTP API server hosting multiple agent instances
kallip-client	Async client library for the daemon HTTP API
kallip	Headless CLI — designed for agents to call, so an agent can manage other agents
kallip-tui	Interactive terminal UI for human users, with approval prompts and markdown rendering
kallip-run	Agent runner for scripting and automation — CI, benchmarks, scripted workflows

Note: just-llm-client is an unrelated upstream crate, not part of this project.

Agentic context management

The most experimental part of the design. Context management in kallip is not hidden behind heuristics — the agent manages its own attention explicitly through tools:

Tool	What it does
context_pin	Mark content as essential — pinned items survive eviction
context_unpin	Remove the pin, allowing the content to be evicted again
context_evict	Evict all turns, replacing them with a summary pinned as context_summary
context_status	Inspect current token usage and pinned items

To make this concrete: context compaction maps directly to context_evict — the agent writes a summary preserving key facts, and the tool atomically pins the summary and evicts all turns. Compaction is not a hidden heuristic but an explicit agent action.

We openly acknowledge that this approach is unproven — it may or may not outperform traditional summarization-only context strategies. But it enables something interesting: skills that emerge naturally.

Emergent skills

When the agent accumulates experience — say, effective patterns for using a particular CLI tool or debugging a class of issues — it can distill that into a file (~/.local/share/kallip/skills/<name>.md). Later, when it encounters a matching situation, it reads the file and pins the content. No dedicated skill system is needed — file read + pin naturally forms skill management.

Asynchronous approval

When a tool call is classified as risky (e.g. rm -rf, sudo, git push --force), the agent does not block waiting for a human. Instead, the call is deferred and the agent manages the lifecycle through tools:

Tool	What it does
approval_list	List approvals, optionally filtered by status
approval_commit	Submit a pending action for approval with justification
approval_redeem	Execute a previously approved action
approval_cancel	Abandon an approval that is no longer needed

The flow:

1. The tool call is deferred — stored in a queue with an approval ID.
2. A deferred result is returned to the LLM immediately, so the agent can continue working on other things.
3. An ApprovalUpdated SSE event is emitted, visible to any client (TUI, CLI, or a supervisor agent).
4. The client approves or denies the request via the daemon's approval API (GET /approvals, POST /approvals/{id}).
5. On the next agent round, the approval notification is injected into context. The agent then calls approval_redeem to execute the stored action.

This design is intentional for multi-agent scenarios: a supervisor agent can monitor approvals from its subagents and make approval decisions programmatically, without a human in the loop. Or it can surface the decision to a human.

Quick start

KALLIP_LLM_PROVIDER=deepseek \
KALLIP_LLM_MODEL=deepseek-v4-flash \
KALLIP_LLM_DEEPSEEK_API_KEY=your-key \
cargo run -p kallip-daemon

# TUI client
cargo run -p kallip-tui