anchor-edit

Stateful single-token anchored file editing tools for coding agents, bringing¹:

• ~60% less token usage for file editing (from O(S+R) to O(R))
• Smaller context, better LLM reasoning

Usage

The anchor-edit mcp subcommand exposes the editor as an MCP server for coding agents like Claude Code.

Claude Code as MCP

• CLI: claude mcp add anchor-edit --scope user -- npx -y anchor-edit mcp

• Example .mcp.json:

  {
    "mcpServers": {
      "anchor-edit": {
        "command": "npx",
        "args": ["-y", "anchor-edit", "mcp"]
      }
    }
  }
  

• Verify: claude mcp list

Pi coding agent as extension

anchor-edit also ships a native pi extension.

pi install npm:anchor-edit

Restart pi, or run /reload in an existing pi session. The extension registers:

1. read_anchored
2. edit_anchored
3. write_to_file

Paths may be relative to pi's current working directory or absolute. The extension keeps anchor state in memory for the active pi runtime. After /reload, restart, /new, /resume, or /fork, call read_anchored again before edit_anchored.

For local development from this repository:

pi -e ./src/pi-extension.ts

Implementation

Reimplements the design from Dirac (anchored edits + Myers-diff reconciler) in TypeScript with AI assistance.

anchor-edit is the stateful single-token variant:

1. Pool of ~6,744 single-BPE English words from o200k_base is the anchor namespace.
2. Per-session in-memory Map<file_path, FileState> (lines, anchors, used-pool words) persists across edits.
3. After every edit, a Myers-diff reconciler reassigns anchors only on changed lines, so unchanged lines keep their identity across the session.

For the full mechanism and design rationale, see Dirac's blog post¹ and the dirac-run/dirac source.

Allocation

Mirrors Dirac's AnchorStateManager. Implementation: src/state.ts.

1. Each file has its own monotonically growing usedWords set. Anchors are pulled from the pool in order on first allocation; deleted lines do not return their anchors to the pool.
2. When the pool (6,744 words from o200k_base) is exhausted, the allocator falls back to random two-word combinations (MorelloMagnificent); on collision, three-word, then four-word.
3. Line splitting is /\r?\n/ (CRLF tolerated). Anchored edits rewrite logical lines with \n; write_to_file writes content as supplied.

Tools

1. read_anchored: returns file content with one anchor prefix per line in the format <anchor>§<content>.
2. edit_anchored: replace / insert_before / insert_after / delete by anchor; empty new_content with mode=replace also deletes the range.
3. write_to_file: creates or overwrites a file; returns newly-allocated anchor runs (Myers-added regions). Call read_anchored afterwards if you need the full post-write anchor map.

For multi-line content and new_content, pass real newline characters (LF, U+000A); literal backslash-n is written to the file verbatim.

When not

Edits made via anchor-edit don't render in Claude Code's native diff UI; tool responses are text-only. Reach for Claude Code's built-in Edit tool when visual diffs matter.

Anchor-edit's value is chained surgical edits where line stability matters. Skip it for:

1. Bulk same-string rename in one file: prefer the host's existing replace-all tool, or a CLI command (sed/sd, rg --replace). One tool call.
2. Same-string rename across many files: prefer a CLI command (sd, rg --replace) or a one-shot script. One tool call total.

Development

# Install dependencies
pnpm install

# Build the bin and types
pnpm run build:pool # one-time pool generation
pnpm run build

# Run tests
pnpm test           # bun test, all suites
pnpm run typecheck
pnpm run lint

Credits

Original idea comes from Dirac (Apache-2.0).

Footnotes

1. Hash Anchors + Myers Diff + Single-Token Anchors ↩ ↩²