dkod-swarm

Local-first parallel AI coding agents. N agents, same files, one PR.

Design  •  How It Works  •  Status  •  Discord

Status

v0 in flight — milestones 1, 2, 3, 4, and 5 merged. cargo test --workspace is green across 8 PRs of M1, 8 of M2, 3 of M3, 3 of M4, and 3 of M5. Empirical proof of the parallel-vs-serial speedup lives in crates/dkod-mcp/tests/bench_parallel_vs_serial.rs; a human-driven counterpart is documented in bench/MANUAL_E2E.md.

The full design lives in docs/design.md. Milestone 6 (marketplace publish — replaces the cargo run-based .mcp.json with binary distribution) is the remaining ship item.

Try it

# In any git repo:
cargo run -p dkod-cli --bin dkod -- init
cargo run -p dkod-cli --bin dkod -- status
cargo run -p dkod-cli --bin dkod -- abort   # only if a session is active
cargo run -p dkod-cli --bin dkod -- --mcp   # stdio MCP server (Claude Code expected)

dkod init writes .dkod/config.toml. dkod status prints a JSON snapshot of the current session. dkod abort destroys an active dk-branch. dkod --mcp is the stdio entry the Claude Code plugin uses (the plugin manifest landed in M4).

The Problem

Running N AI coding agents in parallel usually breaks one of two ways:

• File-level partition. Each agent owns whole files. Safe, but limits parallelism — real tasks touch overlapping files, so you end up with 2-3 parallel agents when you could have 10.
• Free-for-all writes. Agents edit whatever they want. Merge-at-the-end is a conflict nightmare that often produces broken code.

Your agents are fast. Text-level merges are holding them back.

The Fix

dkod-swarm partitions work at the symbol level — function, struct, method — not the file level. Two agents rewrite different functions in the same auth.rs at the same time. An AST-aware merge composes their symbol-level edits into one coherent file. The partition is computed from the call graph, so coupled symbols stay in the same group — correctness preserved, parallelism unlocked.

No server. No database. No cloud coordination. One shared worktree on your machine. N agents. One PR to your upstream.

Symbol-Level Partition Not "agent A gets file1.rs, agent B gets file2.rs." Instead: agent A gets login + logout, agent B gets validate_credentials. Same file. Parallel. AST-merge composes them at the end. One worktree. N agents. Full parallelism.

Local-First, Zero Network Hops No platform, no API, no cloud. Planner, orchestrator, and AST-merge all run in-process inside a single local binary. The only network I/O is the final gh push. Your code never leaves your machine.

What's Safe in Parallel

Scenario	Result
Two agents edit different functions in the same file	Parallel
Two agents add different fields to the same struct	Parallel (AST-merge composes)
Two agents add the same import	Deduplicated
Caller and callee both change with a new signature	Partitioned together (planner keeps coupled symbols in one group)
Agent A deletes a function Agent B calls	Conflict (partition bug, surfaced pre-merge)

How It Works