English | 简体中文
Fast, context-efficient repository tools for AI agents.
FastCtx is a local Rust tool runtime. It provides file reading, content search, file discovery, batch replacement, and Bash command execution through MCP.
Repository operations run in a persistent process with stable input schemas and output formats. The model can gather the context it needs in fewer steps and spend more attention on understanding code, planning changes, and verifying results.
npm install --global fastctx
fastctx
The fastctx command opens the control terminal. Review the proposed changes, select Apply, then start a new ChatGPT / Codex session.
FastCtx currently provides first-class setup for ChatGPT App and Codex CLI. Any MCP client can also register fastctx serve directly.
What FastCtx solves
Coding agents often assemble shell commands on the fly when they access a repository. They have to handle quotes, escaping, paths, and platform differences, then extract the useful information from terminal output. A simple file read or symbol search can take several tool calls just to confirm that the command is correct and the result is complete.
This work consumes context and reasoning. The model tracks the code problem and the tool mechanics at the same time: whether the PowerShell syntax is correct, whether a path was escaped correctly, whether the encoding produced mojibake, and whether the host truncated a long result. More tool overhead leaves less room for the repository itself.
FastCtx turns common repository operations into structured input and output. The model provides parameters such as a path, pattern, range, and mode. The Rust runtime handles command construction, directory traversal, encoding, pagination, and output boundaries.
The tools cover the main parts of a coding task:
readreads text, images, PDFs, and raw bytes;grepsearches file contents;globfinds files;replaceperforms mechanical batch replacement;run,run_background,job_output,job_kill, andjob_listexecute Bash commands and manage persistent long-running jobs.
This greatly reduces the attention the model spends on tool mechanics, such as checking whether a PowerShell command is correct. It improves context efficiency and helps tasks finish faster with better results.
Installation
Install with npm
Requires Node.js 18 or later:
npm install --global fastctx
fastctx
The first launch opens the full-screen control terminal. The interface supports 17 languages and provides these main actions:
- Adjust the output tier;
- Keep grep/glob on automatic CPU parallelism or set an explicit core limit;
- Enable Bash terminal when needed;
- Set current-user background-job storage, concurrency, and AI list page limits;
- Inspect every currently running job across FastCtx sessions, follow its output, and stop it on the Jobs screen;
- Reset all user preferences to factory defaults through a confirmation screen;
- Review every host configuration change on the Apply screen, apply it, and restart the ChatGPT / Codex session.
Apply copies the current binary to ~/.fastctx/bin/ and points the host configuration at that stable path. The applied setup keeps working after npm cache cleanup or upgrades.
On launch, FastCtx checks its launch channel for updates before the main menu opens. A brief checking screen appears and the wait is strictly bounded: if the check cannot finish — offline, timeout, rate limiting — FastCtx enters silently, and the dedicated Update screen still offers a manual check at any time. When a newer version is installable, the update screen opens directly and asks whether to Update & restart or Continue into the current version. Successful results are cached for 24 hours in machine-private storage outside ~/.fastctx, so most launches skip the network entirely. npm launches query the exact launcher package through a fresh isolated cache with --prefer-online; direct GitHub Release executables read the stable tag from GitHub's releases/latest web redirect.
If GitHub has published a release but npm has not exposed the matching version yet, FastCtx shows a propagation screen instead of trusting stale metadata. Retry always uses another isolated cache; it never clears or mutates the user's normal npm cache. Transient network or rate-limit failures stay quiet and are recorded under Status; malformed publication metadata produces one warning. Status also offers a manual check that bypasses the 24-hour TTL. An accepted npm update installs the exact version with lifecycle scripts disabled. A GitHub Release update downloads this repository's platform archive and aggregate SHA256SUMS, verifies the archive before safely extracting the binary, probes the downloaded version, replaces the executable atomically, and rolls back when restart health fails. A failed npm update restores the exact previous package version; every failed update reopens the previous TUI with a warning. After a successful restart, an owned ~/.fastctx/bin/ Apply copy is synchronized; externally changed copies are left untouched.
cargo install builds and the internal ~/.fastctx/bin/ runtime are not self-updated. Set FASTCTX_DISABLE_UPDATE_CHECK=1 to disable the TUI startup check.
Unapply stops FastCtx process images running from the managed bin directory, removes the configuration managed by FastCtx, and deletes its managed data. Shared settings changed by the user after Apply are preserved.
If the install returns 404
Mirror registries copy new releases from the official registry on a delay. Right after a release, an install through a mirror can fail with 404 Not Found — most often on the platform package, which npm installs as an optional dependency and skips silently, leaving fastctx installed but unable to start.
Install once from the official registry:
npm install --global fastctx --registry=https://registry.npmjs.org/
The flag applies to this command only and leaves the npm configuration unchanged. To use the official registry permanently:
npm config set registry https://registry.npmjs.org/ --location=user
After installation, the Update screen probes the npm registry configured on this machine, the official registry, and registry.npmmirror.com, then installs from the first source that carries both the launcher and the matching platform package. Version numbers always come from the official registry and GitHub, so a mirror can never announce a version the official source has not published.
One-off run
npx fastctx
npx opens the same control terminal without a global installation. Apply still copies the binary to ~/.fastctx/bin/, so the applied setup keeps working after the npx cache is cleaned; only the fastctx command itself requires the global installation.
Non-interactive use
fastctx apply --tier standard --yes
fastctx status
fastctx jobs
fastctx jobs kill j-a1b2c3
fastctx unapply --yes
apply: install FastCtx and write the configuration;status: check the configuration, binary, and MCP handshake;jobs: list running background jobs;jobs kill <job_id>: stop one background job and its full process tree;unapply: remove the content managed by FastCtx;lang <code>: set the control terminal language.
status uses three states: [PASS], [INFO], and [FAIL]. It also reports the detected search CPU ceiling and the configured/effective parallelism. A [FAIL] result returns a non-zero exit code.
Search CPU limit and settings reset
grep/glob uses automatic parallelism by default: the operating system's available parallelism, capped at 16. In Config → Search, choose a preset with ←/→ or press Enter and type auto or any integer in the displayed 1..=maximum range. The setting is read directly by each newly started MCP server, takes effect after restarting the server, and does not require Apply.
The same setting can be written manually in ~/.fastctx/config.toml:
[search]
max_cpu_cores = 4
Omitting the key keeps the previous automatic behavior. Invalid types, empty values, zero, negative numbers, and values above the engine's displayed ceiling prevent serve from starting and produce a diagnostic without rewriting the file. The limit sets the effective search parallelism for one request: its base lane plus shared extra workers is at most N. Concurrent requests retain independent base lanes, so this is not CPU affinity or a strict process-wide/system-wide governor.
Config → Reset → Reset all settings opens with No selected. Confirming restores every user preference, including language, output budgets, Bash/job limits, search CPU limit, and update settings. It preserves the Apply ownership receipt, installed binary, host configuration, and running jobs. Restoring the default 1024 MiB job-history quota may immediately evict the oldest finished records above that quota.
Other distribution channels
cargo install fastctx --locked
GitHub Releases provides a zip archive for Windows x64 and executable-preserving tar.gz archives for Linux x64, macOS x64, and macOS arm64. Every archive includes the binary and license notices; verify it with the release's aggregate SHA256SUMS.
Tools
FastCtx provides nine MCP tools:
| Tool | Purpose |
|---|---|
read |
Read one file in any supported format, or batch 1–32 text files |
grep |
Search contents in a file or repository tree |
glob |
Find files by path pattern |
replace |
Apply mechanical batch replacements to files or a repository tree |
run |
Run a Bash command in the foreground |
run_background |
Start a background Bash job |
job_output |
Query a background job and show its newest unseen output |
job_kill |
Stop the full process tree of a background job |
job_list |
Rediscover running and retained finished jobs |
read, grep, glob, and replace are published by default. The other five tools are enabled with the Bash terminal setting in the control terminal. Once enabled, they share the same mcp__fastctx__* namespace as the file tools.
read
read returns 1-based line numbers for text and supports paging:
{
"file_path": "V:/repo/src/main.rs",
"offset": 120,
"limit": 40
}
120 fn main() {
121 ...
159 }
(Partial: lines 120-159 of 512 shown. Continue with offset=160.)
The continuation parameters in the final status line can be used directly in the next call. In this example, pass offset=160 to read the next section.
When several known text files are relevant, batch them into one call instead of paying one agent round trip per file:
{
"files": [
{"path": "V:/repo/src/main.rs", "offset": 120, "limit": 40},
{"path": "V:/repo/src/config.rs"},
{"path": "V:/repo/docs/legacy.txt", "encoding": "gbk"}
]
}
The files form accepts 1–32 text files, preserves request order, and packs them into one shared read budget. A missing, empty, binary, or undecodable member is reported inside its own segment while the remaining files continue. If the budget fills, the final Partial line contains the exact compact files=[...] array for the next call, including per-file offsets, remaining limits, and encodings. Images, PDFs, and hex view remain single-file calls.
read also supports:
- PNG, JPG, GIF, WebP, and BMP images;
- PDF text layers and rendered page images;
- a paged hex view for any file;
- UTF-8, BOM-based encodings, and common legacy encodings.
Automatic encoding detection accepts results with sufficient evidence. When the encoding is ambiguous, the error lists candidates and retry options. Pass encoding to select one explicitly:
{
"file_path": "V:/repo/docs/legacy.txt",
"encoding": "gbk"
}
Use the hex view for binary files:
{
"file_path": "V:/repo/data/cache.bin",
"view": "hex"
}
grep
grep uses the Rust regex engine from the ripgrep family:
{
"pattern": "fn \\w+_lock",
"path": "V:/repo/src",
"output_mode": "content",
"context": 1
}
V:/repo/src/edit/locks.rs
62-/// Cross-process lock keyed by file identity.
63:pub fn acquire_path_lock(identity: &PathIdentity) -> LockGuard {
64- ...
(Complete: all 1 result shown.)
output_mode has four values:
files_with_matches: return matching files;content: show matches grouped by file;count: return the occurrence count for each file;summary: scan the full target and return global totals.
Searches respect .gitignore and .ignore by default, include hidden files, and exclude .git and binary files. Common filters include glob, type, case_insensitive, multiline, and context. Page through results with head_limit and offset.
Files with uncertain encodings appear in a skip report with their path, reason, and resolution parameters. Use encoding for a single file and fallback_encoding for a directory search.
If a file changes during a directory search, grep reports that file as skipped and continues; a changing single-file target returns an error so partial matches never masquerade as complete results.
glob
glob finds files with a pattern relative to the search root:
{
"pattern": "**/*.toml",
"path": "V:/repo",
"sort": "modified"
}
V:/repo/crates/core/Cargo.toml
V:/repo/Cargo.toml
(Complete: all 2 files shown.)
Main parameters:
filter_mode: "project": apply ignore rules, exclude.git, and keep hidden files visible;filter_mode: "all": list every file;sort: "path": use a stable path order;sort: "modified": order files from newest to oldest;offset/limit: page through the result set.
grep and glob render filename components that are unsafe to place directly in a line as reversible ~fastctx~b...~ or ~fastctx~w...~ escapes. Copy the whole component exactly into a later grep/glob call; do not decode or edit it.
replace
replace handles mechanical, deterministic batch changes such as symbol renames, import rewrites, configuration key migrations, and fixed-pattern deletion. Generated code changes and per-location semantic edits are handled by the host's apply_patch tool.
{
"pattern": "old_name\\(",
"replacement": "new_name(",
"path": "V:/repo/src",
"glob": "**/*.rs",
"dry_run": true
}
...
(Complete: dry run — 12 matches in 3 files; nothing written.)
replace freezes the candidate set and counts every match before the first write. Use dry_run for preview and max_replacements to cap the change scope.
Each file is checked again before commit. Writes use atomic replacement in the same directory and preserve the original encoding, BOM, line endings, trailing newline, Unix mode, and untouched bytes. Concurrent changes move the affected file into the failure report while the remaining files continue.
run
run executes a Bash command in the foreground and returns merged stdout, stderr, and the exit code. It uses Git Bash on Windows and the system Bash on macOS and Linux.
{
"command": "cargo test --quiet 2>&1 | tail -n 40",
"timeout_ms": 180000
}
Commands run in a non-interactive environment. Installation, confirmation, and editor commands need flags such as -y and --no-edit. Non-zero exit codes are returned as execution results.
On Windows, every FastCtx-owned non-interactive child process is created without allocating a console window, including Bash discovery, foreground/background Bash, detached supervisors, and doctor probes. There is no hidden-window parameter to remember. A command that explicitly launches a GUI or a new terminal still has that visible effect.
Output uses bounded memory. When output exceeds the response capacity, the final status line reports the truncated range and gives a path to the complete result: redirect the command output to a file, then page through it with read.
run_background
run_background starts a background Bash job and returns a job id immediately. It is useful for builds, tests, development servers, and other long-running commands.
Each job is owned by a detached supervisor rather than by the MCP server. It keeps running across server exits, ChatGPT / Codex restarts, and session changes until the command exits or job_kill stops it. There is no background timeout parameter.
Output and exit status are stored under ~/.fastctx/jobs/, so another FastCtx server can resume the same job by id. For jobs started by the current format, everything printed is appended to a plain log file there — nothing is rotated away — and its path is returned when the job starts, so read and grep work on it directly.
job_output
job_output queries a background job, including jobs started in earlier sessions, and reports running, exited, or interrupted together with the newest output the caller has not been shown. wait_ms (0–60000, default 30000) is how long the query may take: it returns as soon as the job ends, and otherwise waits the window out, because returning on every intermediate line would spend a whole turn to report that a build is still compiling. Pass wait_ms=0 for an immediate snapshot. Long current-format output is windowed — the newest lines that fit, plus the start of the log on the first call — and a note names the exact lines that were skipped and the log path to read them from. Line numbers in that log are the same seq numbers after_seq takes, so moving between the two tools needs no translation. Records written by the preceding segmented format remain readable, including while an older supervisor is still appending, but they do not advertise direct log coordinates and cannot recover bytes that their original rolling window already evicted.
Keep calling it until the final line says Complete. For current-format jobs, the background log keeps every line, so anything a response leaves out is still one read or grep away; the compatibility limitation above applies only to records created by the preceding format.
job_kill
job_kill stops the selected background job and its full process tree. If the job has already exited, the call returns the existing exit status.
job_list
job_list defaults to status="running". Use status="finished" to inspect retained exited or interrupted records, and status="all" only when both lifecycles are needed. Results are newest first within each lifecycle. offset continues a page; limit overrides the saved page size for one call.
Finished records have no time-to-live. FastCtx retains them until the current user's fastshell.job_storage_limit_mib limit requires eviction of the oldest finished records. The default is 1024 MiB. Running jobs and their records are never evicted; fastshell.max_running_jobs limits concurrent jobs across all FastCtx sessions for that user and defaults to 128. fastshell.job_list_limit is the default page size (20, valid range 1–100). All three settings take effect immediately when saved and do not require Apply; the TUI presets for page size are 10 / 20 / 50 / 100.
The TUI Jobs dashboard scans this same current-user registry but shows only jobs that are currently running, aggregated from every FastCtx server and TUI instance. A finished job disappears with a short notice that its retained output remains available to the agent through job_output. Jobs are grouped by an honest source-session tag with workspace, server PID, and parent-process context. Fixed list columns keep relative age and job ids aligned, while long ASCII or CJK commands end with an ellipsis at one shared edge. The detail panel shows the exact UTC start time to the second and a live HH:MM:SS elapsed time. Horizontal and vertical output navigation remains available; one width-aware footer row keeps the essential keys visible and adds ←/→ output, PgUp/PgDn scroll, and F follow when space permits. ChatGPT / Codex does not expose conversation titles or ids to the MCP server, so FastCtx does not invent one.
Security and privacy
The FastCtx MCP server inherits the local permissions of the host process.
| Capability | Default state | Access scope |
|---|---|---|
read / grep / glob |
Enabled | Local files readable by the host process |
replace |
Enabled | Local file writes with dry-run, CAS, and atomic replacement safeguards |
| Bash tools | Disabled | Bash command execution after the user enables them |
| TUI update check | Enabled for npm and GitHub Release launches | Version metadata from registry.npmjs.org and GitHub's releases/latest web redirect; downloads require explicit confirmation |
| MCP runtime network requests | None | serve and tool calls perform no telemetry or update traffic |
The startup check sends the FastCtx version, normal HTTPS request metadata, and npm's standard registry request; it never sends repository paths, job data, or file contents. Background jobs persist their command, working directory, output, and exit status only in the current user's private ~/.fastctx/jobs/ directory; current-format jobs use a complete plain log. FastCtx does not upload this data. Bash commands can access the network according to the command itself. Prebuilt binaries include the PDF engine.
The MCP server runs outside the host filesystem sandbox. Use an approval mode when every write and command execution should be reviewed:
[mcp_servers.fastctx]
default_tools_approval_mode = "writes"
writes: reviewreplaceand shell execution tools;prompt: review every tool call.
replace is published with the default file tools. The host's read-only mode covers the host's own tools. MCP writes still run with the server process permissions. Set writes or prompt when the workflow depends on a read-only boundary.
Codex configuration
Codex code mode places regular MCP tools inside an execution container. Aggregated results from multiple calls can be truncated in the middle by the host. This setting keeps FastCtx as a direct top-level namespace:
[features.code_mode]
direct_only_tool_namespaces = ["mcp__fastctx"]
Apply maintains this setting automatically and writes a guidance block with explicit markers to ~/.codex/AGENTS.md. The guidance scopes FastCtx to local-file reading, searching, and finding, tells the model to read only what the task needs, and directs several known files into one files=[...] read call.
The control terminal provides three output tiers. Each keeps FastCtx's internal budget at 85% of the host limit so responses close before Codex can truncate their terminal status:
Compact: host limit 10,000; FastCtx budget 8,500;Standard: the default, with host limit 20,000 and FastCtx budget 17,000;High: host limit 30,000; FastCtx budget 25,500.
Higher output tiers allow larger results per call and consume context faster. Choose a tier according to the task.
Manual MCP registration
[mcp_servers.fastctx]
command = "C:/absolute/path/to/fastctx.exe"
args = ["serve"]
startup_timeout_sec = 120
tool_timeout_sec = 300
[features.code_mode]
direct_only_tool_namespaces = ["mcp__fastctx"]
Enable the Bash tools with:
args = ["serve", "--enable-shell"]
When the binary is on PATH, command can be set to fastctx. The compatibility npm package codex-fastctx installs the same fastctx command.
What FastCtx changes
FastCtx uses or manages these paths and settings:
~/.fastctx/bin/fastctx(.exe): the stable self-installed binary;~/.fastctx/config.toml: control terminal settings and the Apply receipt;~/.fastctx/jobs/: persistent background-job records and current-format full output logs, created on demand byrun_background;[mcp_servers.fastctx]in~/.codex/config.toml, includingtool_timeout_sec = 300;- the
mcp__fastctxentry indirect_only_tool_namespaces; - the marker-delimited FastCtx block in
~/.codex/AGENTS.md; - the selected
tool_output_token_limitvalue after user confirmation.
FastCtx edits existing TOML with toml_edit, preserving comments, formatting, and unrelated configuration. Unapply removes entries according to write ownership and preserves later user changes. It stops running background jobs before removing ~/.fastctx/.
Contact
FastCtx is created and maintained by yc-duan. For integration, redistribution, partnership, or anything else, feel free to reach out: [email protected].
Comments